Impellam CFO Links Finance AI Failures to Bad Workflows, Not Models

You Can't Automate What You Can't See With Impellam Group — Ai4

Ai4

Enterprise finance AI underperforms because organizations deploy algorithms over broken processes and untrusted data without assigning single-threaded human owners to govern outputs.

AI-assisted summary of You Can't Automate What You Can't See With Impellam Group — original source. Check the original for full context.

Source published: 2026-08-18

Detailed notes

## Background Published on YouTube by Ai4 on August 18, 2026, this presentation features Thomas DeFabrizio, CFO at Impellam Group. According to the description, it serves as a practitioner field report addressing why 63% of finance teams deploy AI while only about one in five can prove ROI, introducing a five-step framework (see it, clean it, frame it, build it, refine it) to make finance automations repeatable and reliable.

## TL;DR Thomas DeFabrizio argues that AI underperformance in corporate finance stems not from deficient models, but from broken workflows, untrusted data, and lack of single-threaded ownership. Drawing parallels to the 2008 financial crisis, he emphasizes that raw data access is meaningless without structural visibility and strict human governance. DeFabrizio demonstrates proven use cases—such as account reconciliations and rapid reporting deck overhauls—and outlines a five-step operational methodology to implement AI safely without delegating ultimate authority. He concludes by warning about the erosion of junior finance apprenticeship, urging modern CFOs to transform from historical record-keepers into active enterprise system architects.

## Atmosphere & style The presentation is brisk, candid, and refreshingly pragmatic, cutting through conference floor hype with the grounded skepticism of an operating CFO. DeFabrizio blends sharp humor about standing between the audience and lunch with personal war stories from the 2008 financial crisis and weekend-long coding sprints. The talk feels like a direct, roll-up-your-sleeves executive briefing that values unvarnished operational discipline and accountability far above vendor sales pitches.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=pG2FNIJViz8&t=4s) The CFO's primary mandate in enterprise AI is translating raw technological capability into durable economic value, operational risk mitigation, and verified bottom-line results. - [01:22](https://www.youtube.com/watch?v=pG2FNIJViz8&t=82s) While 63% of finance departments have deployed AI, only 21% can prove measurable ROI, highlighting a massive gap between crossing the adoption line and crossing the value line. - [01:22](https://www.youtube.com/watch?v=pG2FNIJViz8&t=82s) Access to massive volumes of real-time data is not equivalent to structural visibility, a core operational lesson forged during the 2008 financial crisis. - [03:10](https://www.youtube.com/watch?v=pG2FNIJViz8&t=190s) AI underperformance is rarely caused by model limitations; breakdowns in governance, untrusted data, and lack of single-threaded ownership drive failure in 92% of underperforming finance initiatives. - [03:10](https://www.youtube.com/watch?v=pG2FNIJViz8&t=190s) Fewer than 4% of mid-market CFOs cite technology enablement as their primary obstacle, identifying strategic prioritization, talent scarcity, and organizational capacity as the real blockers. - [05:40](https://www.youtube.com/watch?v=pG2FNIJViz8&t=340s) Layering generative AI over broken manual workflows or legacy tech debt merely accelerates the propagation of errors under the guise of polished, authoritative deliverables. - [07:42](https://www.youtube.com/watch?v=pG2FNIJViz8&t=462s) High-ROI finance AI use cases succeed by focusing on recurring cadence and clean data, allowing AI to process high-volume matching while humans retain sole sign-off authority. - [09:59](https://www.youtube.com/watch?v=pG2FNIJViz8&t=599s) Delivering complex AI projects rapidly requires deep business context, component-by-component prompt iteration, and specialized tools rather than monolithic, one-click prompts. - [12:17](https://www.youtube.com/watch?v=pG2FNIJViz8&t=737s) Modern AI stacks allow non-technical leaders to build functional production applications, but strategic logic, scoring weights, and domain judgment cannot be outsourced to the model. - [13:42](https://www.youtube.com/watch?v=pG2FNIJViz8&t=822s) A disciplined five-step operating sequence—See It, Clean It, Frame It, Build It, Refine It—enforces that autonomy is not authority before software is ever built or procured. - [15:11](https://www.youtube.com/watch?v=pG2FNIJViz8&t=911s) The automation of entry-level tasks creates an apprenticeship dilemma, requiring finance organizations to actively invest in reskilling to replace the tacit business knowledge historically gained from manual work. - [16:54](https://www.youtube.com/watch?v=pG2FNIJViz8&t=1014s) The Office of the CFO must transition from administrative record-keepers reporting the past into enterprise system architects who design the operational engines of the future.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=pG2FNIJViz8&t=4s) The CFO's Mandate: Turning Possibility into Economic Value

Taking the stage as the literal obstacle standing between the audience and lunch, Thomas DeFabrizio acknowledges the high ROI required of his twenty-minute window before cutting straight to the fundamental tension in enterprise artificial intelligence. Demos across the conference floor display dazzling visions of what is technologically possible, but the Chief Financial Officer's mandate is fundamentally different: asking what must be true inside the business for that raw potential to translate into durable enterprise value. That pragmatic skepticism often earns finance leaders the reputation of being the "bad guys and gals" who slow innovation down. However, DeFabrizio clarifies that this caution is not rooted in obstructionism, but in ultimate accountability for enterprise economics, operational risk, and measurable bottom-line results.

The central thesis of the presentation is simple yet uncompromising: "You can't automate what you can't see." While the broader market fixates on model architectures and cutting-edge software showcases, real-world finance automation succeeds or fails on the discipline of the operating machine beneath it. DeFabrizio emphasizes that he is not on stage to pitch software; he builds AI tools himself, implements them within his organization, and carries direct CFO accountability for governing their behavior.

[01:22](https://www.youtube.com/watch?v=pG2FNIJViz8&t=82s) The Value Line and Lessons from the 2008 Financial Crisis

Finance functions are confronting a stark performance paradox highlighted by recent industry research: while a Deloitte study shows that 63% of finance departments have deployed AI in some capacity, a mere 21% can point to clear, measurable return on investment. That massive disparity demonstrates that while corporate finance has successfully crossed the adoption line, it remains stalled far short of the value line. The minority of teams achieving actual financial returns are not simply running an endless string of speculative experiments to generate happy-hour talking points; they are disciplined operators who target well-defined business problems, insist on verified data, assign single-threaded owners, build strict controls, and systematically quantify operational changes.

This perspective was forged during the early days of DeFabrizio's career at the Bank of New York in the run-up to the 2008 global financial crisis. Sitting behind Bloomberg terminals, surrounded by a deluge of real-time transactions, asset pricing feeds, counterparty risk sheets, and complex valuation models, he watched Tier-1 financial institutions collapse in real time. The data existed in staggering volumes, yet it failed to prevent disaster because raw information is not synonymous with visibility. The systemic failure of 2008 proved that having access to metrics means nothing without structural clarity, establishing the philosophical bedrock for how modern finance must approach algorithmic tools.

[03:10](https://www.youtube.com/watch?v=pG2FNIJViz8&t=190s) Governance, Token Burn, and the Real Bottlenecks of AI Underperformance

When diagnosing why enterprise finance AI initiatives miss expectations, corporate leaders often look in the wrong direction. In a Grant Thornton survey of nearly 1,000 senior financial leaders, 92% acknowledged experiencing AI underperformance. When asked for the primary cause, the leading response was not a lack of sophisticated models, but breakdowns in governance and compliance. The model is almost never the bottleneck; failure occurs in the structural vacuum surrounding it—unclear manual workflows, untrusted underlying databases, and zero individual accountability for the algorithm's output. Bolting advanced frontier intelligence onto weak operational hygiene merely amplifies organizational confusion.

This diagnostic gap becomes even clearer when polling mid-market CFOs directly. In a survey of 230 finance chiefs conducted by a peer board on which DeFabrizio serves, CFOs ranked prioritization as their single greatest hurdle, followed closely by talent and organizational capacity. Remarkably, technology and AI enablement were cited as the primary obstacle by less than 4% of respondents. Entering the year, only one in eight CFOs felt highly confident that AI initiatives would deliver measurable ROI. Consequently, executive leadership is keeping experimental budgets on a short leash, with Wall Street analysts actively grilling institutions like JPMorgan Chase on earnings calls regarding their specific AI "token burn." Furthermore, corporate leadership remains intensely anxious about enterprise data leakage and where proprietary numbers migrate once exposed to vendor pipelines.

``` WHY FINANCE AI UNDERPERFORMS: PERCEPTION VS. REALITY

Assumed Bottlenecks Actual Operating Obstacles (Ranked <4% by CFOs) (Grant Thornton / CFO Studies) ┌─────────────────────────┐ ┌───────────────────────────────┐ │ • Weak AI Model │ VS │ • Governance & Compliance │ │ • Immature Tooling │ │ • Prioritization Deficits │ │ • Tech Enablement Gaps │ │ • Talent & Capacity Scarcity │ └─────────────────────────┘ │ • Untrusted Data / No Owners │ └───────────────────────────────┘ ```

- 92% of senior finance executives report AI underperformance in their organizations. - Less than 4% of mid-market CFOs view technology or enablement tools as their primary blocker. - Only 1 in 8 CFOs (12.5%) enter the current operating year highly confident in AI delivering positive ROI. - Executive scrutiny has shifted from pilot vanity metrics toward managing raw AI token burn and enterprise data exposure.

[05:40](https://www.youtube.com/watch?v=pG2FNIJViz8&t=340s) Operating on Legacy Debt vs. Clean Foundations

Generative artificial intelligence does not inherently replace or redesign the business engine beneath it; it sits directly on top of legacy infrastructure. When underlying workflows are clearly articulated, the core data is trusted, and human review boundaries are established, AI accelerates turnaround times and produces reliable leverage. However, when an underlying process is structurally broken, layering AI on top simply accelerates the propagation of errors through the corporate pipeline—packaging flawed conclusions into visually convincing, highly authoritative deliverables far more dangerous than the clumsy output of past automation scripts.

Compounding this hazard is the reality that corporate finance remains anchored to legacy manual systems. Over half of enterprise finance organizations operate with zero core automation, leaning almost entirely on manual data entry and brittle spreadsheets. Furthermore, 86% of financial leaders report that legacy tech debt severely restricts their readiness to integrate modern AI workflows. DeFabrizio, currently leading a major transformation as CFO of a private equity-backed workforce solutions business implementing Workday Financials and Workday Adaptive Planning, warns against using tech debt as an excuse for inaction. Rather than waiting years for an ERP overhaul or attempting overly ambitious, blue-sky moonshots, finance executives must initiate AI deployments exclusively where teams understand the underlying workflow, trust the data lineage, and maintain strict human review mechanisms.

[07:42](https://www.youtube.com/watch?v=pG2FNIJViz8&t=462s) Proven Use Cases: Reconciliations, Journal Entries, and Scenario Modeling

Examining real-world finance functions reveals that high-ROI artificial intelligence deployments cluster around three functional domains: operational workflows, financial controls, and forward-looking strategic analysis. On the operational side, high payoffs appear in invoice matching, month-end close acceleration, and account reconciliations. Within controls, AI proves exceptional at analyzing journal entries and flagging anomalous general ledger activity. On the strategic side, models ingest historical seasonality and operational trends to generate multi-variable scenario modeling. While DeFabrizio jokes about burning corporate tokens to run millions of hypothetical operating scenarios that will likely never materialize, the common denominator behind every winning use case is consistent: recurring cadence, known source data, and explicitly defined review gates.

To demonstrate how this functions in practice, DeFabrizio highlights an account reconciliation workflow. Traditionally, staff accountants manually match two large, established data streams—such as internal general ledger entries against bank statements, or ERP subledger transaction details against external settlement files. In the automated architecture, an AI pipeline standardizes file formats, executes matching logic, isolates routine transactions, and bifurcates anomalies into an exception log. Crucially, the system does not produce an unauditable black-box conclusion; it outputs a structured reconciliation file with exceptions surfaced for human inspection. By letting AI handle data processing while finance professionals retain sole approval authority, the implementation recaptured over 200 hours of staff capacity in a single organization.

``` RECONCILIATION WORKFLOW PATTERN

┌──────────────────────┐ ┌──────────────────────┐ │ Internal Ledger / │ │ External Statement/ │ │ Subledger File │ │ Settlement File │ └──────────┬───────────┘ └──────────┬───────────┘ │ │ └────────────┬─────────────┘ ▼ ┌───────────────────────────────────┐ │ AI INGESTION ENGINE │ │ • Standardize formatting & schema│ │ • Execute automated matching │ └────────────────┬──────────────────┘ │ ┌───────────┴───────────┐ ▼ ▼ ┌───────────────────┐ ┌───────────────────┐ │ Routine Matches │ │ Unresolved Items │ │ (Auto-Cleared) │ │ (Exception Queue) │ └───────────────────┘ └─────────┬─────────┘ │ ▼ ┌───────────────────┐ │ Human Finance │ │ Investigation, │ │ Approval & Sign-│ │ Off │ └───────────────────┘ ```

- Core operational use cases: high-volume account reconciliations, automated three-way invoice matching, and ledger close tasks. - Control enhancement use cases: real-time journal entry risk-scoring and transaction anomaly detection. - Strategic forecasting use cases: multi-factor scenario simulation incorporating historical seasonality and operational trend data. - Quantified operational impact: 200+ annual labor hours directly returned to the accounting team on a single reconciliation deployment.

[09:59](https://www.youtube.com/watch?v=pG2FNIJViz8&t=599s) Rebuilding a Global Financial Deck in Four Days

To demonstrate what AI execution demands from executive leadership, DeFabrizio shares a personal turnaround story from his private equity-backed HR technology company. The enterprise required a consolidated global financial deck to serve both internal management and the governing private equity board across all operating regions. A senior, highly capable external finance contractor had been leading the project for three months utilizing AI tooling, yet the initiative remained stalled without a finished, functional deliverable. Stepping in, DeFabrizio recognized that the bottleneck was not a matter of tweaking slides; the underlying structure and reporting logic were fundamentally flawed.

Treating the project as a greenfield rebuild, DeFabrizio reconstructed the entire reporting framework from scratch using an AI-first methodology. Four days later—powered by late nights, an intensive weekend, and more than 30 hours of rigorous manual labor—the global deck was completed, deployed, and adopted across every operating region. Emphasizing that AI is never an easy "one-click" magic solution, DeFabrizio identifies three prerequisites that made the turnaround possible: 1. Deep business context: DeFabrizio understood precisely where raw data lived, which internal sources could be trusted, and the exact strategic signals executive leadership and the private equity sponsor required to drive decisions. 2. Hands-on AI operational expertise: Rather than dumping a messy, monolithic prompt into an LLM, he decomposed the deliverable into discrete analytical components, provided targeted context, aggressively challenged the model's interim drafts, and iterated through successive refinements. 3. Proper tool fit: He selected generative tools specifically designed for corporate presentation synthesis rather than forcing analytical workflows into poorly suited models.

[12:17](https://www.youtube.com/watch?v=pG2FNIJViz8&t=737s) Democratized Software Development: The Five-Component Diagnostic Stack

The democratization of generative AI has eliminated the barrier between strategic leadership and technical execution. Non-technical executives no longer need to submit tickets to engineering backlogs to build working enterprise applications. To prove the point, DeFabrizio highlights a live diagnostic software product he engineered independently for an upcoming business book—bypassing corporate development resources entirely.

His lightweight production stack combines five interoperable, off-the-shelf tools: - Tally for dynamic input form generation and front-end user interactions. - Zapier for business process workflow orchestration and data routing. - Claude with Sonnet for analytical reasoning, contextual scoring, and generative text synthesis. - Kit (formerly ConvertKit) for customer segmentation and digital audience management. - Resend for transactional output delivery and programmatic client communication.

While admitting that an enterprise data science team could have written a cleaner, more scalable codebase, DeFabrizio notes that speed to execution is what matters. He used AI as an interactive pairing partner to debug API integrations, configure webhook connections, and resolve deployment edge cases. However, AI could never supply the strategic intent: it did not determine which operational diagnostic questions to ask, how to weight business scoring parameters, or what conclusions would actually provide actionable value to a reviewing CEO. The business judgment and domain logic had to originate entirely from the executive architect, transforming what could have been generic AI slop into a functional business diagnostic.

[13:42](https://www.youtube.com/watch?v=pG2FNIJViz8&t=822s) The Five-Step Framework: From Initial Mapping to Enterprise Scale

Whether architecting an external diagnostic assessment or evaluating an automated finance pipeline on a Monday morning, DeFabrizio adheres to a strict five-step operating sequence. The core mistake of failed corporate pilots is jumping immediately to software procurement before understanding the underlying workflow. By enforcing operational clarity at the foundational stages, finance leaders guarantee that automations solve authentic enterprise friction rather than institutionalizing chaos.

``` DELEGATION VS. CONTROL: THE 5-STEP ORDER OF OPERATIONS

Step 1: SEE IT ──► Map the actual, ground-truth business workflow. Step 2: CLEAN IT ──► Audit data integrity and assign a named owner. Step 3: FRAME IT ──► Define decision rights ("Autonomy is not Authority"). Step 4: BUILD IT ──► Select specialized tooling; execute targeted pilot. Step 5: REFINE IT ──► Review edge exceptions, audit changes, and scale. ```

The framework demands rigorous progression across five linear phases: - Step 1: See It. Map the actual workflow taking place on the ground, utilizing process intelligence tools rather than relying on idealized documentation. - Step 2: Clean It. Restrict execution exclusively to trusted data streams, and require that a named individual within the leadership team personally owns the data output. Meetings that lack an explicit owner for an outcome inevitably stall. - Step 3: Frame It. Establish deterministic decision boundaries, operating policies, and strict approval thresholds. Underscoring that "autonomy is not authority," DeFabrizio relates an anecdote of an acquaintance's young daughter ordering burritos via Amazon Alexa; assuming the device failed to register her voice, she repeated the order until nine burritos showed up at the front door. Guardrails must enforce that autonomous systems cannot execute unapproved commercial actions. - Step 4: Build It. Initiate targeted builds using modular, specialized tooling rather than sprawling, monolithic platforms. - Step 5: Refine It. Measure quantitative delta against baseline performance, conduct human audits of all exception logs, remediate structural edge cases, and institutionalize the workflow as a repeatable standard.

For vendors selling to enterprise finance, DeFabrizio offers a critical piece of positioning advice: software building begins at Step 4, but business value is generated at Step 1. Product pitches that fail to address the first three stages will consistently fail to gain traction with operating CFOs.

[15:11](https://www.youtube.com/watch?v=pG2FNIJViz8&t=911s) The Apprenticeship Dilemma and the Imperative for Reskilling

The rapid automation of entry-level accounting and finance workflows creates a profound, under-discussed organizational consequence: the erosion of junior corporate apprenticeship. AI excels at executing low-level repetitive tasks—querying reports, executing ledger matching, formatting data extracts, and producing baseline forecasting drafts. Yet historically, those mundane assignments served as the primary training ground where junior analysts learned how their company actually operates.

By manually generating ad-hoc extracts, an analyst uncovered the fundamental economics of customer contracts. By reconciling messy GL accounts, an accountant learned precisely where enterprise internal controls broke down. By compiling the baseline inputs of an operating model, a financial planner discovered which commercial assumptions genuinely moved the needle. If algorithmic workflows swallow these entry-level duties without structural redesign, the institutional learning curve is severed: - 73% of enterprises have deployed or actively piloted artificial intelligence tools. - Fewer than 10% of enterprise workforces have received formal training on how to use AI workflows. - Only 20% of corporate organizations have redesigned functional job descriptions to accommodate automated pipelines.

Preserving manual, inefficient busywork simply to train entry-level staff is an untenable strategy. However, modern enterprises face an urgent obligation to replace the tacit institutional knowledge those tasks once provided. DeFabrizio stresses that despite being a CFO responsible for driving P&L performance, he actively advocates for expanding corporate budgets for learning, development, and workforce reskilling to cultivate the next generation of strategic decision-makers.

[16:54](https://www.youtube.com/watch?v=pG2FNIJViz8&t=1014s) Redesigning the Office of the CFO

The central strategic question facing finance executives is not "which finance tasks can AI do?" Treating AI as an ad-hoc plug-in to accelerate existing legacy tasks misses the broader structural opportunity. Enterprise finance must be radically redesigned from first principles around artificial intelligence, purposefully dividing workflows between machine computation and human judgment. Algorithms should ingest, harmonize, analyze, and draft; human operators must inspect, challenge, approve, and carry ultimate accountability for corporate results.

This architectural shift profoundly alters the identity of the Chief Financial Officer. The historical mandate of finance—reporting past results, safeguarding accounting accuracy, and maintaining statutory internal controls—remains non-negotiable. Yet those duties are no longer sufficient on their own. Modern CFOs must step into the role of enterprise system architects who actively design the operational engines producing the future, rather than serving as administrative historians who merely tally the past. Teams achieving genuine value from AI do not begin by browsing enterprise vendor halls for frontier models; they map baseline operations, enforce data hygiene, mandate personal ownership, establish strict review gates, and systematically iterate toward durable ROI.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=pG2FNIJViz8&t=4s) You can't automate what you can't see. — Thomas DeFabrizio

> [03:10](https://www.youtube.com/watch?v=pG2FNIJViz8&t=190s) The gap is almost never the model; it is the machine the model is bolted onto, which means the data, the process, and who owns the decision. — Thomas DeFabrizio

> [00:04](https://www.youtube.com/watch?v=pG2FNIJViz8&t=4s) bad guys and gals — Thomas DeFabrizio

> [03:10](https://www.youtube.com/watch?v=pG2FNIJViz8&t=190s) token burn — Thomas DeFabrizio

> [09:59](https://www.youtube.com/watch?v=pG2FNIJViz8&t=599s) one-click — Thomas DeFabrizio

> [13:42](https://www.youtube.com/watch?v=pG2FNIJViz8&t=822s) Autonomy is not authority — Thomas DeFabrizio

> [13:42](https://www.youtube.com/watch?v=pG2FNIJViz8&t=822s) see it, clean it, frame it, build it, refine it — Thomas DeFabrizio

> [16:54](https://www.youtube.com/watch?v=pG2FNIJViz8&t=1014s) which finance tasks can AI do? — Thomas DeFabrizio

> [16:54](https://www.youtube.com/watch?v=pG2FNIJViz8&t=1014s) Come if you would rather architect the system that produces the future than report the past. — Thomas DeFabrizio

## Mentioned Thomas DeFabrizio, Impellam Group, Ai4, Deloitte, Bank of New York, Bloomberg, Grant Thornton, JPMorgan Chase, Workday Financials, Workday Adaptive Planning, Tally, Zapier, Claude, Sonnet, Kit, ConvertKit, Resend, Amazon Alexa

## Video description Thomas DeFabrizio - CFO, Impellam Group

Finance teams deployed AI at scale: 63% now use it, and only about 1 in 5 can prove it paid off.

The gap is almost never the model; it is the machine the model is bolted onto, which means the data, the process, and who owns the decision.

I will show you the pattern behind the finance AI use cases that are actually working right now, then go up close on a few, from reconciliations rebuilt in an afternoon to a company-wide reporting deck I rebuilt in days, not months. Then I will give you a five-step order of operations (see it, clean it, frame it, build it, refine it) that makes any of them repeatable and keeps AI from simply making the wrong number look official.

I build these tools myself, including a diagnostic that runs live on a model's API, so expect a practitioner field report rather than a sales pitch. Come if you would rather architect the system that produces the future than report the past.

[Source](https://www.youtube.com/watch?v=pG2FNIJViz8)

中文简报

Impellam CFO：财务AI失败根源在于不良工作流而非模型企业财务AI表现不佳，是因为机构在破碎的流程与不可信的数据上部署算法，且未指定专职单一责任人来监管输出结果。 Thomas DeFabrizio认为，企业财务中AI表现欠佳并非源于模型缺陷，而是由于破碎的工作流程、不可信的数据以及缺乏专职单一责任人（single-threaded ownership）。他以2008年金融危机作类比，强调若缺乏结构性可见度与严格的人工治理，单纯获取原始数据毫无意义。DeFabrizio展示了经过验证的实际应用案例（如账目核对与报告幻灯片的快速重组），并提出了一套五步操作方法，旨在不让渡最终决策权的前提下安全实施AI。他最后警告称，初级财务人员的学徒培养机制正在受到侵蚀，并敦促现代CFO从历史记录员转变为主动的企业系统架构师。 CFO在企业AI应用中的首要职责是将原始技术能力转化为持久的经济价值、运营风险缓解以及经过验证的实际业绩成果。 (0:04) 尽管有63%的财务部门已部署AI，但仅有21%能够证明可衡量的ROI，凸显出跨越“应用门槛”与跨越“价值门槛”之间存在巨大差距。 (1:22) 获取海量实时数据并不等同于具备结构性可见度，这是2008年金融危机期间形成的核心运营教训。 (1:22) AI表现欠佳极少由模型局限所致；在表现不佳的财务AI项目中，92%的失败源于治理崩溃、不可信数据以及缺乏专职单一责任人。 (3:10) 仅有不足4%的中端市场CFO将技术赋能视为主要障碍，他们认为真正的阻碍在于战略优先级排序、人才匮乏以及组织能力不足。 (3:10)

Reading summary

Impellam Group CFO Thomas DeFabrizio argues that enterprise finance artificial intelligence initiatives stall not because of immature models, but because organizations fail to establish workflow visibility, data integrity, and strict human governance before automating. The Adoption-Value Gap and Governance Bottlenecks (1:22) Corporate finance functions face a distinct performance paradox: Deloitte research indicates that while 63% of finance departments have deployed AI in some capacity, only 21% report measurable return on investment. Furthermore, a Grant Thornton survey of nearly 1,000 senior financial leaders revealed that 92% have experienced AI underperformance, pointing to governance and compliance breakdowns rather than algorithmic limitations as the primary culprit. Drawing on his experience at the Bank of New York during the 2008 financial crisis, DeFabrizio observes that massive volumes of data do not inherently provide visibility or prevent catastrophe without structural clarity and assigned accountability. This operational disconnect is echoed among mid-market leaders. In a survey of 230 CFOs, less than 4% identified technology or enablement tools as their main obstacle, citing prioritization, talent, and organizational capacity instead. Entering the operating year, only one in eight CFOs felt highly confident that AI projects would yield positive ROI. With institutional scrutiny intensifying over raw token burn and proprietary data leakage, enterprise finance struggles primarily with operational hygiene rather than model sophistication. Legacy Tech Debt and High-Yield Finance Workflows (5:40) Over half of enterprise finance organizations operate with zero core automation, relying heavily on spreadsheets and manual entry, while 86% of finance leaders report that legacy tech debt restricts their ability to adopt AI workflows. DeFabrizio warns that layering generative models over broken processes merely accelerates the propagation of errors, packaging flawed numbers into authoritative-looking deliverables. Rather than waiting years for sweeping ERP overhauls, he recommends deploying AI strictly where underlying data lineage is understood and human review boundaries are enforceable. High-return finance deployments consistently cluster around three operational areas: high-volume operational workflows (such as three-way invoice matching and account reconciliations), internal controls (including real-time journal entry risk-scoring and anomaly detection), and forward-looking strategic simulations. In an account reconciliation pipeline, an automated engine standardizes files and matches entries while segregating anomalies into an exception log for human approval. By handling repetitive data ingestion while preserving human sign-off authority, this single workflow returned over 200 annual hours to an accounting team. A Five-Step Framework for Controlled Execution (13:42) To prevent pilot failures, DeFabrizio advocates a sequential five-step operating methodology: See It (mapping actual workflows on the ground), Clean It (verifying data integrity and mandating single-threaded executive owners), Frame It (setting explicit decision rights and deterministic thresholds), Build It (initiating targeted pilots using modular tools), and Refine It (auditing exception logs and quantifying operational delta against baselines). He emphasizes that autonomy does not equal commercial authority, requiring hard guardrails to prevent unvetted automated decisions from executing in production. Demonstrating that domain context and strategic direction matter far more than heavy engineering resources, DeFabrizio highlights rebuilding a stalled global financial deck for a private equity board in four days using targeted prompts and presentation tools, after an external contractor spent three months without completing it. Similarly, he engineered a live business diagnostic product using an off-the-shelf stack—Tally, Zapier, Claude with Sonnet, Kit, and Resend—proving that business judgment, scoring logic, and workflow design must originate from the executive rather than the software vendor. The Junior Apprenticeship Crisis and Structural CFO Redesign (15:11) The rapid automation of foundational finance tasks creates an unintended organizational vulnerability: the erosion of junior corporate apprenticeship. Entry-level assignments such as ledger reconciliation, report formatting, and ad-hoc data extraction historically taught analysts the commercial mechanics of customer contracts and the operational vulnerabilities of internal controls. Industry metrics show that while 73% of enterprises have piloted or deployed AI tools, fewer than 10% of workforces have received formal training on AI workflows, and only 20% of companies have redesigned job descriptions accordingly. DeFabrizio argues against preserving inefficient manual tasks purely for training purposes, urging finance executives instead to expand budgets for workforce reskilling and structured development. He asserts that the Office of the CFO must evolve from an administrative historian that aggregates past performance into an enterprise systems architect. Under this model, finance functions intentionally assign data aggregation and baseline drafting to algorithmic engines while reserving oversight, critical evaluation, and ultimate accountability for human operators. DeFabrizio concludes that enterprise value from AI stems from operational discipline, clear ownership, and workflow governance rather than vendor demos. However, the presentation leaves open the question of how quickly mid-market enterprises can bridge legacy tech debt and talent scarcity to achieve reliable financial returns.

精炼阅读摘要

Thomas DeFabrizio 提出，企业人工智能自动化的成败不在于模型能力的先进与否，而取决于底层运营纪律，即“你无法自动化你看不到的东西”。首席财务官（CFO）的核心职责并非阻碍技术创新，而是严格把控运营风险与资本回报，确保算法潜力在清晰的业务流程与治理结构之上转化为持久的经济价值。价值鸿沟与治理合规的现实瓶颈 (1:22) 企业在财务 AI 的落地应用中正遭遇明显的投入产出悖论。德勤（Deloitte）的研究数据显示，尽管已有 63% 的财务部门在业务中部署了 AI 工具，但仅有 21% 能够证明其带来了明确且可衡量的投资回报率（ROI）。Grant Thornton 对近千名高级财务主管的调研进一步表明，92% 的受访者遭遇过 AI 表现不及预期的问题，其中被列为首要原因的并非模型成熟度不足，而是治理与合规机制的缺失。在另一项针对 230 位中型市场 CFO 的同行调研中，受访者将战略优先级混乱以及人才与组织能力不足列为最大瓶颈，仅有不到 4% 的人认为技术工具本身是主要阻碍；同时，仅有八分之一（12.5%）的财务首脑对 AI 产生正向 ROI 持高度信心。 DeFabrizio 结合自身在 2008 年金融危机前夕任职于纽约银行的经历指出，当年终端屏幕上充斥着海量的交易数据与复杂的定价模型，却未能阻止顶级金融机构的系统性崩溃，这证明单纯占有信息并不等于具备运营可视性。当前，管理层对 AI 实验性预算的审查日趋严格，华尔街分析师已开始在财报电话会议中直接质询摩根大通（JPMorgan Chase）等机构的 AI“token 消耗”情况；加之对专有数据泄露风险的担忧，企业对 AI 的评估已从概念验证的虚荣指标彻底转向对治理合规与底层数据的严格约束。遗留系统技术债与高回报落地场景 (5:40) 企业财务流程普遍受到遗留系统与手动操作的严重制约。数据显示，超过半数的企业财务部门核心业务自动化程度为零，高度依赖人工录入与电子表格；86% 的财务主管直言，遗留技术债务严重制约了组织整合现代 AI 工作流的能力。DeFabrizio 警告称，生成式 AI 并不能自行修复底层破损的业务引擎，若直接在混乱的流程上套用 AI，只会以极高的权威感加速错误扩散。因此，财务团队不应以等待 ERP 系统重构为由停滞不前，而应优先在业务流程清晰、数据血缘可信且设有严格人工审核的环节推行自动化。在实际业务场景中，高 ROI 的财务 AI 应用高度集中于三类标准化领域：一是运营工作流（如三方发票自动匹配、月末关账加速和账户对账）；二是内部控制（如日记账实时风险评分与总账异常交易识别）；三是战略分析（如结合历史季节性数据的多变量情景模拟）。以高频账户对账为例，AI 引擎能够统一数据格式并自动核销绝大部分常规条目，同时将未决差异自动分流至异常队列，保留人类财务人员的排查与最终签字权。在具体落地案例中，该模式成功为单个机构的会计团队每年释放了超过 200 小时的人工工时。五步实施框架与决策控制权界定 (13:42) 为防止企业自动化项目陷入盲目采购软件的误区，DeFabrizio 建立了一套五步实施框架：第一步“看清”（See It），利用流程智能测绘地面真实业务链路，而非依赖理想化的制度文档；第二步“清洗”（Clean It），严格限定受信任的数据源，并要求管理层指定专人承担数据输出责任；第三步“界定”（Frame It），划定明确的确定性决策边界与审批阈值，强调“自主权不等于审批授权”（正如儿童通过智能音箱重复指令导致误购 9 个卷饼的案例所警示的失控风险）；第四步“构建”（Build It），采用专用模块化工具开展针对性试点；第五步“优化”（Refine It），量化评估成效并建立长效审计机制。对于软件供应商而言，技术采购始于第四步，但实际商业价值早在一至三步就已经决定。这种业务逻辑先行的原则同样体现在去中心化的轻量级开发中。DeFabrizio 曾结合 Tally、Zapier、Claude（搭配 Sonnet）、Kit 和 Resend 这五款现成商用工具，在完全绕过传统 IT 排期的情况下独立构建出一套面向企业的商业诊断应用；同时，他曾通过深度拆解业务指标与多轮迭代，仅耗时四天便重构了一套私募控股企业跨区域使用的全球财务汇报框架。他指出，非技术高管完全可以借助 AI 承担 API 联调和脚本编写工作，但诊断题目的设计、评分权重的配置以及战略结论的推导，必须完全由具备业务判断力的管理者主导。初级学徒制断层与财务长角色重塑 (15:11) 初级财会工作的大规模自动化正在给企业组织结构带来深层次的次生挑战。调查数据显示，虽然已有 73% 的企业部署或试点了 AI 工具，但接受过正式 AI 技能培训的员工比例不足 10%，仅有 20% 的企业针对自动化流程重新设计了岗位职责说明。在传统财务体系中，手动拉取报表、核对杂乱账目和搭建基础模型等繁琐工作，是初级分析师深入理解商业合同本质、发现内控漏洞和掌握核心业务动因的关键学徒平台。若任由算法直接吞噬这些基础工作而不做组织培训补救，企业将面临人才梯队培养断层的风险。面对这一困局，保留低效的人工作业已不可取，管理层必须积极扩大针对员工再培训与学习发展的预算投入，以重构人才培养机制。在此背景下，CFO 自身的定位也正在经历根本性演变：维护历史账目准确与坚守法定内控依然是不可妥协的底线，但现代 CFO 已不能仅仅扮演记录过去的“行政历史学家”，而必须转型为主动设计人机协作引擎、统筹运营架构并对最终商业结果负责的“企业系统架构师”。 DeFabrizio 的核心论点指出，企业财务 AI 的成功并非模型竞赛，而是一场建立在业务流程可视性、数据可靠性与清晰责任制之上的系统工程。然而，在基础岗位被算法快速替代的趋势下，企业如何在消除低效工时的同时有效重建年轻人才的专业培训闭环，以及面对持续的遗留技术债务与高昂的运行成本如何建立可持续的 ROI，仍是横亘在财务决策者面前的核心未决挑战。

Detailed notes — zh

## Background 本视频由 Ai4 频道于 2026 年 8 月 18 日发布在 YouTube 平台上。内容由 Impellam Group 首席财务官（CFO）Thomas DeFabrizio 主讲，围绕企业财务团队部署 AI 时面临的“高采纳率与低 ROI”断层，结合其一线实操经验与自研诊断工具，深入解析如何打通数据、流程与权责体系以实现真正的商业价值转化。 ## TL;DR 尽管目前已有 63% 的财务团队引入了 AI，但仅有 21% 能够证明其明确的投资回报率，其核心症结不在于大模型本身，而在于模糊的流程、不可信的数据以及缺失的产出所有权。演讲者通过自动化账户对账以及耗时 4 天重构全球财务汇报材料等实战案例，提炼出“看见业务、清洗数据、确立框架、动手构建、持续精炼”的五步工作法。此外，针对自动化导致初级员工丧失业务学习路径的隐性代价，他呼吁加大员工技能重塑投入，并指出现代 CFO 正从传统的历史数据记录员转变为面向未来的工作流架构师。 ## Atmosphere & style 整场演讲风格务实、风趣且极具一线实战的硬核底气。Thomas DeFabrizio 以掌管企业损益表的 CFO 视角切入，既有幽默自嘲的开场，又有结合 2008 年金融危机教训与近期熬夜攻坚案例的深度反思，彻底剥离了行业展厅中对 AI 技术的盲目狂热。整场分享节奏紧凑、逻辑严密，充满审慎的风险控制意识与务实的商业判断力，给观众带来极具启发性与实操价值的专业复盘体验。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=pG2FNIJViz8&t=4s) 财务 AI 采纳率与真实价值转化之间存在严重断层：Deloitte 调研显示 63% 的财务团队已部署 AI，但仅有 21% 能证明其带来了清晰可衡量的 ROI。 - [02:24](https://www.youtube.com/watch?v=pG2FNIJViz8&t=144s) 财务 AI 落地的瓶颈几乎从不是大模型算法本身，而是模糊的业务流程、缺乏信任的数据源以及缺失的产出所有权等周边支撑体系。 - [02:24](https://www.youtube.com/watch?v=pG2FNIJViz8&t=144s) CFO 面对 AI 的首要阻碍是业务优先级排序与人才资源，而对 AI 模型及技术工具本身的担忧提及率不足 4%。 - [05:40](https://www.youtube.com/watch?v=pG2FNIJViz8&t=340s) AI 无法替代底层流程，如果基础流程本身有缺陷，AI 只会加速传递错误并以更权威的格式包装错误（Making the wrong number look official）。 - [07:42](https://www.youtube.com/watch?v=pG2FNIJViz8&t=462s) 当前真正产生商业价值的财务 AI 场景具备三大共性：高频重复的工作、已知可信的数据源以及清晰的人工复核机制。 - [07:42](https://www.youtube.com/watch?v=pG2FNIJViz8&t=462s) 在账户对账场景中，AI 负责格式标准化和常规匹配预处理，财务人员专注于异常调查与最终确认，单团队即可节省约 200 工时。 - [09:59](https://www.youtube.com/watch?v=pG2FNIJViz8&t=599s) 重构耗时 3 个月未果的全球财务汇报体系仅用 4 天完成，依托的是业务上下文认知、模块化提示词经验与专用工具契合度的结合。 - [12:17](https://www.youtube.com/watch?v=pG2FNIJViz8&t=737s) 现代轻量级工具链让非技术背景的业务高管能够亲自搭建 AI 工具，核心差异在于业务逻辑与价值定义而非代码技术本身。 - [13:42](https://www.youtube.com/watch?v=pG2FNIJViz8&t=822s) 财务 AI 落地遵循五步法：看见业务（See it）、清洗数据（Clean it）、确立框架（Frame it）、动手构建（Build it）、持续精炼（Refine it）；构建始于第四步，但价值始于第一步。 - [13:42](https://www.youtube.com/watch?v=pG2FNIJViz8&t=822s) 必须严格区分自主性（Autonomy）与决策授权（Authority），明确 AI 的操作边界与人工审批机制，防止失控风险。 - [15:11](https://www.youtube.com/watch?v=pG2FNIJViz8&t=911s) AI 接管初级财务任务可能阻断新人的业务学习路径，企业在推进自动化的同时必须加大在员工技能重塑（Reskilling）上的投入。 - [16:54](https://www.youtube.com/watch?v=pG2FNIJViz8&t=1014s) 现代 CFO 的职能正在发生转变，不仅要当好历史财务业绩的记录员，更要成为面向未来 AI 工作流的系统架构师。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=pG2FNIJViz8&t=4s) CFO的现实视角与“采纳率”到“价值转化”的断层作为站在听众与午餐之间的演讲者，Impellam Group 首席财务官（CFO）Thomas DeFabrizio 开场便幽默地表示自己背负着在这20分钟内提供极高投资回报率（ROI）的压力。他直言，展厅里令人眼花缭乱的 AI 演示向人们展示了技术潜力的“可能性”，但 CFO 的核心职责却是审视“为了将这些可能性转化为企业内部的真实价值，必须满足哪些前提条件”。这种审慎的态度往往让财务管理者显得像是在“踩刹车”的“恶人”，但这并非出于阻碍业务的本意，而是因为 CFO 必须对企业的经济效益、风险控制和最终成果承担终极责任。 DeFabrizio 指出，财务 AI 的核心现实是“你无法将自己看不见的事物自动化”（You can't automate what you can't see）。德勤（Deloitte）的数据清晰揭示了当前业界的断层：高达 63% 的财务部门已经部署了 AI 工具，但其中仅有 21% 能够证明其带来了明确且可衡量的 ROI。这表明企业已经跨过了 50% 的技术采纳门槛，却迟迟未能跨越真正的“价值转化线”。真正获得可衡量价值的团队绝不是靠盲目做更多实验，而是选择了明确的问题场景、依托可信的数据、确立了责任人、设立了控制机制，并严格衡量前后的业务变化。这是在业务中产生实质改变与仅仅在行业酒会上讲个好故事之间的本质区别。 - **Deloitte 行业调研数据**：63% 的财务职能团队已部署 AI，但仅 21% 能够报告清晰、可衡量的投资回报率（ROI）。 - **核心管理逻辑**：跨越“价值线”的关键在于五个要素——明确的问题场景、可信的数据基础、明确的责任人、健全的控制流程，以及对产出变化的严格衡量。 --- ### [02:24](https://www.youtube.com/watch?v=pG2FNIJViz8&t=144s) 2008年危机的教训：为何财务 AI 的瓶颈不在于模型本身 DeFabrizio 分享了一段在生成式 AI 出现之前的职业往事：2008年金融危机爆发前夕，他在纽约银行（Bank of New York）开启职业生涯，每天坐在彭博终端前处理海量的头寸、价格、交易、估值和风险报告。然而，即便身处数据海洋之中，他依然目睹了大型金融机构在危机面前措手不及。“数据本身并不等于可见性”（Data by itself is not visibility），这一深刻教训构成了理解当今财务 AI 困境的核心。根据 Grant Thornton 对近千位资深财务高管的调研，92% 的受访者表示自己的 AI 项目存在至少一项导致表现不佳的原因。出人意料的是，被提及最多的问题并非缺乏优秀的 AI 模型，而是“治理与合规”（Governance and Compliance）。在现代企业环境中，大语言模型本身极少成为真正的瓶颈，系统的崩溃往往发生在模型周围的支撑体系：业务流程模糊不清、组织内部数据缺乏信任度、产出结果无人明确认领。把先进的模型直接套在薄弱的运营流程之上，注定无法发挥作用。 ```mermaid flowchart TD accTitle: 财务AI落地瓶颈分析图 subgraph Bottlenecks["财务 AI 落地的三大外围瓶颈"] P["模糊的业务流程 (Unclear Process)"] D["不可信的数据源 (Untrusted Data)"] O["缺失的产出所有权 (Unassigned Ownership)"] end subgraph ModelLayer["模型层（非主要瓶颈）"] M["先进的大语言模型 / AI 算法 (SOTA Models)"] end subgraph Outcome["最终结果"] F["错误加速放大与负 ROI (Failed Implementation)"] end P --> M D --> M O --> M M --> F ``` 在另一项针对 230 位中端市场 CFO 的董事会调研中，CFO 们将 AI 落地的头号障碍归结为“优先级排序”（Prioritization），随后是人才储备与交付能力，而技术本身与 AI 赋能能力的提及率分别不足 4%。更严峻的是，进入今年仅有八分之一（12.5%）的 CFO 高度确信 AI 能带来可衡量的 ROI。虽然 AI 获得了预算支持，但管理层对其容忍度明显收紧，CFO 们正紧密监控模型的“Token 消耗速度”（Token Burn），这在摩根大通（JPMorgan Chase）等大型机构的财报电话会议中已成为分析师关注的焦点；同时，数据资产在企业级模型内部的流向与合规性依然是 CFO 最大的担忧。 - **Grant Thornton 调研**：92% 的受访财务高管指出 AI 项目存在表现不佳的问题，首要痛点为治理与合规，而非模型能力。 - **中端市场 CFO 调研（230人）**：首要阻碍为业务优先级排序，其次是人才与资源容量；技术与工具本身作为阻碍的比例均低于 4%。 - **CFO 信心指数**：仅 1/8（约 12.5%）的 CFO 对 AI 实现可衡量的 ROI 持高度信心。 --- ### [05:40](https://www.youtube.com/watch?v=pG2FNIJViz8&t=340s) 财务底层流程的现状与错误加速风险 AI 在当前阶段并不能直接替代企业底层的业务流程，而是作为一层能力叠加在现有流程之上，借助现有的流程与数据来提升运转速度。当底层流程清晰、数据可信且复核机制健全时，AI 能够带来更快的执行效率和可靠的产出；但如果底层流程本身存在缺陷，AI 只会“更快地将错误传递到后续流程中”，并且以比以往自动化工具更具欺骗性、更权威的格式将错误呈现出来（"Making the wrong number look official"）。这种风险在财务领域尤为突出，因为大部分财务部门的数字化底座依然相当脆弱。行业数据显示，超过一半的财务团队至今仍未实现自动化，依旧严重依赖手工录入和 Excel 电子表格；高达 86% 的受访者直言传统遗留系统（Legacy Systems）严重制约了企业接入 AI 的成熟度。 DeFabrizio 以自身在私募股权（PE）投资背景公司推进数字化转型的经历说明，他们正在实施 Workday Financials 和 Workday Adaptive，旨在优化财务结构以便未来出售，但遗留系统的问题绝不意味着企业要停下脚步等待整套技术架构推倒重来。对于 CFO 和业务领导者而言，切忌一上来就追求宏大的 AI 构想，而必须从财务团队完全理解的流程、完全信任的数据以及结果能够被人工有效复核的场景切入，否则必将陷入负 ROI 的泥潭。 - **财务自动化现状**：超 50% 的财务团队完全无自动化支持，高度依赖手工和表格；86% 认为遗留系统阻碍了 AI 准备度。 - **落地黄金法则**：从财务团队已完全理解的既有流程、具备信任度的数据资产，以及具备明确人工复核能力的节点切入。 --- ### [07:42](https://www.youtube.com/watch?v=pG2FNIJViz8&t=462s) 财务 AI 创造价值的真实场景与对账案例根据同行实践与亲身经历，DeFabrizio 将目前能够产生明确商业价值的财务 AI 场景归纳为三大类： 1. **日常运营类（Operational）**：包括账户对账（Reconciliations）、发票匹配（Invoice Matching）和财务关账活动。 2. **控制与治理类（Strengthen & Control）**：日记账分录分析（Journal Entry Analysis）与异常检测（Anomaly Detection）。 3. **前瞻预测类（Forward-Looking）**：利用 AI 深度分析历史趋势、季节性变化与过往表现，并生成多维度情景分析（Scenario Analysis）。DeFabrizio 调侃道，虽然情景分析中 99% 的极端情况永远不会发生，但 CFO 们依然热衷于“燃烧 Token”来穷尽各种可能性。所有这些成功场景都遵循着完全相同的底层模式：**高频重复的工作、已知且可信的数据源、清晰的复核机制**。AI 负责数据的清洗、处理与预加工，而财务人员则牢牢掌握核心假设、异常调查和最终决策的控制权。 ```mermaid flowchart LR accTitle: 财务AI人机协同对账流程图 subgraph DataPrep["数据输入与 AI 预处理"] direction TB S1["数据源 A (如总账科目)"] S2["数据源 B (如对账单/流水)"] AI["AI 自动化引擎 • 标准化数据格式 • 应用匹配规则"] S1 --> AI S2 --> AI end subgraph Split["分类分流"] direction TB M["常规匹配交易 (Routine Matches)"] E["异常与差异项 (Exceptions)"] end subgraph HumanLoop["人工复核与决策 (Human-in-the-Loop)"] direction TB Rev["财务人员人工调查 (Finance Investigation)"] App["最终审批与决策确认 (Approval & Sign-off)"] end AI --> M AI --> E E --> Rev Rev --> App ``` 在详细拆解“账户对账”这一经典案例时，该场景涉及每个周期对两组既定财务数据源进行核对（例如总账科目对照银行对账单，或内部交易流水对照外部汇总报告）。AI 工作流完成了文件的格式标准化、执行匹配规则、将常规匹配项与异常项清晰分离，并直接向团队呈现需要重点关注的差异清单。整个系统绝非不可解释的“黑盒”，而是生成清晰的对账底稿与异常列表供财务团队审核。最终，AI 负责繁重的预处理，财务团队集中精力调查异常并签字确认。仅在这一项流程中，同行团队就节省了约 200 个工时的手工核对时间。DeFabrizio 强调，这种场景虽然在宣传上并不酷炫，但“释放出来的工时”正是 CFO 最看重且最愿为之买单的实际效益。 - **成功场景共性**：周期性重复（Reoccurring work）、已知数据（Known data）、明确复核路径（Clear review）。 - **人机协同分工**：AI 负责数据标准化与常规匹配筛选，财务人员负责例外调查与决策定夺。 - **对账案例成效**：通过 AI 预处理与异常清单呈现，单个团队释放了约 200 小时的手工工作时间。 --- ### [09:59](https://www.youtube.com/watch?v=pG2FNIJViz8&t=599s) 实战案例：4天重构全球财务汇报体系为了展现 AI 在高层管理场景中的真实落地方式，DeFabrizio 分享了自己在当前私募股权投资背景人力资源科技公司的亲身实践。当时公司急需一份统一的全球财务汇报材料（Financial Deck），供各业务区域、内部管理层以及 PE 投资机构共同使用。该项目此前已推进了约 3 个月，由一位资深的外部财务顾问主导，并引入了多种 AI 工具，但数月过去依然未能交付出可用的终版成品。 DeFabrizio 介入后发现，核心症结不在于现有 PPT 页面的缝缝补补，而是底层的报告架构本身存在严重缺陷。他决定抛开历史技术包袱，采用“AI 优先”（AI-First）的方法从零重构整个报告体系。最终，在经历了整整一周的连续夜战、周末加班以及超过 30 个小时的高强度实操后，全新的全球报告在 4 天内彻底完成并正式分发给各区域使用。他特别澄清，AI 绝非“按一个按钮就能自动搞定工作”，这次突破能够实现完全依托于三大核心要素： 1. **业务上下文认知（Business Context）**：他清晰掌握公司底层数据的真实来源、哪些数据源具备高度公信力，以及管理层做决策时究竟需要看什么信息。“业务上下文在未来十年甚至二十年内都将是无可替代的核心”，缺乏业务认知，模型就无法对齐真实意图。 2. **AI 模块化驾驭经验（AI Experience）**：懂得如何将庞杂的任务拆解为独立的结构化模块，为模型提供精确的上下文指令，对生成结果保持质疑与校验，并通过多轮持续迭代逐步逼近高可用模型。 3. **工具选型匹配度（Tool Fit）**：针对演示汇报与数据呈现的工作特性，选择了最适配该特定工作流的专业 AI 工具。 DeFabrizio 总结道，作为 CFO，自己平时并不亲自下场编写大量基础报告，但正是因为深刻理解最终成果的标准，AI 才能在人类商业判断力的引导下，将数月停滞不前的工作缩短至数天内高质量完成。 - **项目背景与周期对比**：耗时 3 个月未果的全球财务汇报重构项目，经由 AI 优先工作流在 4 天（累计 30+ 真实工时）内完成交付。 - **AI 赋能成功三要素**： 1. **业务上下文（Business Context）**：掌控数据源信任链与高层决策意图。 2. **AI 操作经验（AI Experience）**：任务拆解、上下文提示、结果质疑与持续迭代优化。 3. **工具契合度（Tool Fit）**：根据具体呈现形态选用针对性的专用 AI 生产力工具。 ### [12:17](https://www.youtube.com/watch?v=pG2FNIJViz8&t=737s) 业务领导者亲自构建：AI 工具开发不再是技术团队的专利现代工具链的革新打破了软件构建的传统门槛，开发能力不再被局限在专业的技术与数据科学团队内部。作为一名非技术出身的首席财务官（CFO），演讲者分享了自己为即将出版的新书所做的个人项目——一套完整的 AI 商业诊断工具。整套技术栈由无代码与轻量级工具组装而成（包括 Tally 表单、Zapier 自动化工作流、Claude 3.5 Sonnet 核心模型以及 Resend 邮件发送系统），全程没有专业数据科学团队介入。虽然专业团队可能写出更优雅、构建更迅速的代码，但关键在于：非 IT 背景的业务决策者现在能够比以往任何时候都更直接地参与到产品与工具的实际构建中。 AI 在构建过程中扮演了排错、接口调试和工作流打通的助手角色，期间解决的问题“远比在台上愿意承认的要多得多”。然而，AI 并没有决定向 CEO 提什么业务问题、如何对反馈进行评分，更没有定义什么才是一份对管理者真正有价值的报告。所有的核心底层逻辑完全来自业务端本身，这正是产出一份能为 CFO 提供清晰洞察的实用工具、而非生成一份充斥着 AI 空话的套话报告（AI slop）的根本分水岭。 - **个人实战技术栈**：Tally（数据收集）+ Zapier（逻辑编排）+ Claude 3.5 Sonnet（AI 逻辑与文本处理）+ Resend（结果交付）。 - **人机分工核心**：AI 负责连接工具与排查 Bug，业务管理者负责核心提问框架、评分逻辑与商业价值定义。 --- ### [13:42](https://www.youtube.com/watch?v=pG2FNIJViz8&t=822s) 财务 AI 落地的五步工作法：从看见业务到规模化构建评估和落地任何财务 AI 用例都应遵循一套自底向上的五步操作法则。构建工具本身只排在第四步，但真正的业务价值早在第一步就已开始沉淀。 ```mermaid flowchart LR accTitle: 财务 AI 实施五步法 subgraph S1["准备与治理阶段"] A["1. 看见业务 (See it) 梳理实际工作流"] --> B["2. 清洗数据 (Clean it) 确认可信数据并指派责任人"] B --> C["3. 确立框架 (Frame it) 界定决策规则与审批边界"] end subgraph S2["构建与迭代阶段"] C --> D["4. 动手构建 (Build it) 小步起步，选对工具"] D --> E["5. 持续精炼 (Refine it) 评估差异、修复薄弱点并规模化"] end ``` 1. **看见业务（See it）**：完整绘制出企业内部真实发生的业务与工作流程，而不是停留在理论上的标准作业流程。 2. **清洗数据（Clean it）**：使用真正可信的数据，并且必须明确指定数据的业务负责人（Owner）。在管理层会议中，一旦缺少明确负责人，决策就会陷入停滞，因此每一个关键产出都必须绑定具体负责人。 3. **确立框架（Frame it）**：清晰界定决策范畴、业务规则、输出形式以及哪些环节仍需人工审批。演讲者引用了一个生动的例子：一位父亲的女儿通过 Alexa 语音点墨西哥卷饼，因为以为设备没听清而重复指令，结果门口送来了九个卷饼——这说明必须时刻牢记“自主性（Autonomy）并不等同于决策授权（Authority）”。 4. **动手构建（Build it）**：选择合适且轻量适用的工具，从小切口场景起步快速落地。 5. **持续精炼（Refine it）**：量化落地后的业务变化，复盘异常情况，修补系统薄弱环节，最终将其固化为可复用的标准化流程并推向规模化。对于向企业推销 AI 解决方案的厂商而言，唯有贯穿这五步逻辑进行沟通才能真正打动 CFO，因为“构建始于第四步，但价值始于第一步”。 --- ### [15:11](https://www.youtube.com/watch?v=pG2FNIJViz8&t=911s) 自动化背后的隐性代价：初级任务消失与组织学习断层作为一家人力资源解决方案公司的 CFO，演讲者指出了 AI 落地带来的一个不可忽视的运营隐患：AI 极为擅长替代财务的基础层工作——如拉取报表、对账、生成第一版财务分析草稿。然而，这些基础事务恰恰是初级财务人员“学习业务本质”的核心途径。正是通过拉取底层数据，分析师才搞懂公司到底是如何盈利的；正是通过手工对账，员工才察觉内部控制在哪里存在漏洞；正是通过搭建第一版预测模型，新人才能理解哪些业务假设真正撬动着公司的业绩大盘。当 AI 接管了所有底层任务，原本伴随手工操作而产生的业务认知与学习效应并不会自动留存。数据残酷地反映了这一断层鸿沟：高达 73% 的企业已经部署或试点上线了 AI，但接受过系统 AI 培训的员工不足 10%，仅有 20% 的企业针对 AI 重构了岗位角色。企业绝不应为了培养新人而刻意保留低效的手工作业，但必须主动设计机制来填补失去的学习通道。尽管掌管着公司的损益表（P&L），演讲者依然强烈主张企业必须加大在员工学习与发展（L&D）以及技能重塑（Reskilling）上的资金投入。 - **AI 渗透与培训断层数据**： - **73%** 的企业已部署或正在试点 AI； - 仅有 **<10%** 的员工接受了相关培训； - 仅有 **20%** 的企业重新设计了岗位职责。 --- ### [16:54](https://www.youtube.com/watch?v=pG2FNIJViz8&t=1014s) 财务重构而非修补：从汇报过去转向架构未来业界过度聚焦于“AI 能做哪些财务任务”这个狭隘的命题，真正巨大的机会在于从源头上重新设计工作的完成方式。这种重构并非简单地为了裁减财务人员，而是明确划定职责边界：让 AI 负责信息的准备、分析与自动化执行，而将商业判断、终审把关与结果问责牢牢保留在人类手中。“AI 改变了工作的完成方式，但人类依然承担决策和结果的最终责任。” 财务的未来绝非“在旧流程之上缝合 AI”，而是“围绕 AI 重新设计财务工作流”。这一变革直接重塑了 CFO 自身的核心职能。传统的历史数据汇报、财务准确性、内控与合规完整性不但没有被削弱，反而要求更高；但现代 CFO 不能再仅仅充当历史业绩的记录员，更必须成为面向未来系统与工作流的架构师。资本市场与投资者电话会议的关注点也印证了这一点：真正从 AI 获得可衡量回报的团队，从不会把时间浪费在盲目挑选“最先进的模型”上，而是扎扎实实地梳理流程、清洗数据、明确责任人、设定人工审批节点，随后才进入构建与规模化阶段。 --- ### [18:42](https://www.youtube.com/watch?v=pG2FNIJViz8&t=1122s) 总结与资源分享在演讲尾声，演讲者再次提醒在场管理者，不要被“为了快而快”的说辞所误导，前期梳理业务与治理数据的准备工作绝非拖延时间，因为“价值始于第一步，构建始于第四步”，这一认知决定了企业究竟是获得切实的商业价值，还是仅上线了一个看似炫酷的玩具。演讲者在现场屏幕上留下了二维码，公开了本次演讲的全部材料，并在附录中提供了大量极具操作性的落地指南： - 如何挑选第一个财务 AI 落地场景； - AI 与传统自动化的边界区别； - ROI 计算模型、权限控制与安全规范； - 自研（Build）与采购（Buy）的决策权衡； - 现场演示的业务可见度诊断工具。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=pG2FNIJViz8&t=4s) 你无法将自己看不见的事物自动化（You can't automate what you can't see）。 — Thomas DeFabrizio > [02:24](https://www.youtube.com/watch?v=pG2FNIJViz8&t=144s) 数据本身并不等于可见性（Data by itself is not visibility）。 — Thomas DeFabrizio > [05:40](https://www.youtube.com/watch?v=pG2FNIJViz8&t=340s) 以比以往自动化工具更具欺骗性、更权威的格式将错误呈现出来（Making the wrong number look official）。 — Thomas DeFabrizio > [09:59](https://www.youtube.com/watch?v=pG2FNIJViz8&t=599s) 业务上下文在未来十年甚至二十年内都将是无可替代的核心。 — Thomas DeFabrizio > [13:42](https://www.youtube.com/watch?v=pG2FNIJViz8&t=822s) 自主性（Autonomy）并不等同于决策授权（Authority）。 — Thomas DeFabrizio > [13:42](https://www.youtube.com/watch?v=pG2FNIJViz8&t=822s) 构建始于第四步，但价值始于第一步。 — Thomas DeFabrizio > [16:54](https://www.youtube.com/watch?v=pG2FNIJViz8&t=1014s) AI 改变了工作的完成方式，但人类依然承担决策和结果的最终责任。 — Thomas DeFabrizio > [16:54](https://www.youtube.com/watch?v=pG2FNIJViz8&t=1014s) 现代 CFO 不能再仅仅充当历史业绩的记录员，更必须成为面向未来系统与工作流的架构师。 — Thomas DeFabrizio ## Mentioned Thomas DeFabrizio, Impellam Group, Deloitte (德勤), Bank of New York (纽约银行), Grant Thornton, JPMorgan Chase (摩根大通), Workday Financials, Workday Adaptive, Tally, Zapier, Claude 3.5 Sonnet, Resend, Alexa ## Video description Thomas DeFabrizio - CFO, Impellam Group Finance teams deployed AI at scale: 63% now use it, and only about 1 in 5 can prove it paid off. The gap is almost never the model; it is the machine the model is bolted onto, which means the data, the process, and who owns the decision. I will show you the pattern behind the finance AI use cases that are actually working right now, then go up close on a few, from reconciliations rebuilt in an afternoon to a company-wide reporting deck I rebuilt in days, not months. Then I will give you a five-step order of operations (see it, clean it, frame it, build it, refine it) that makes any of them repeatable and keeps AI from simply making the wrong number look official. I build these tools myself, including a diagnostic that runs live on a model's API, so expect a practitioner field report rather than a sales pitch. Come if you would rather architect the system that produces the future than report the past. [Source](https://www.youtube.com/watch?v=pG2FNIJViz8)

Source quotes

[0:04] Thomas DeFabrizio: 你无法将自己看不见的事物自动化（You can't automate what you can't see）。 [2:24] Thomas DeFabrizio: 数据本身并不等于可见性（Data by itself is not visibility）。 [5:40] Thomas DeFabrizio: 以比以往自动化工具更具欺骗性、更权威的格式将错误呈现出来（Making the wrong number look official）。 [9:59] Thomas DeFabrizio: 业务上下文在未来十年甚至二十年内都将是无可替代的核心。 [13:42] Thomas DeFabrizio: 自主性（Autonomy）并不等同于决策授权（Authority）。 [13:42] Thomas DeFabrizio: 构建始于第四步，但价值始于第一步。 [16:54] Thomas DeFabrizio: AI 改变了工作的完成方式，但人类依然承担决策和结果的最终责任。 [16:54] Thomas DeFabrizio: 现代 CFO 不能再仅仅充当历史业绩的记录员，更必须成为面向未来系统与工作流的架构师。

Source quotes

[0:04] Thomas DeFabrizio: You can't automate what you can't see. [3:10] Thomas DeFabrizio: The gap is almost never the model; it is the machine the model is bolted onto, which means the data, the process, and who owns the decision. [0:04] Thomas DeFabrizio: bad guys and gals [3:10] Thomas DeFabrizio: token burn [9:59] Thomas DeFabrizio: one-click [13:42] Thomas DeFabrizio: Autonomy is not authority [13:42] Thomas DeFabrizio: see it, clean it, frame it, build it, refine it [16:54] Thomas DeFabrizio: which finance tasks can AI do? [16:54] Thomas DeFabrizio: Come if you would rather architect the system that produces the future than report the past.

Original transcript

[0:04] Good afternoon everybody. Uh I understand I'm the CFO standing between you and lunch. So I know I've [0:10] got a high ROI for the next 20 minutes. All right. So let's get into it. All right. The AI demos that [0:17] we've seen similar to the ones in exhibit floor tell us what is possible. The CFO's job is to [0:27] ask what has to be true for that possibility to turn into value inside the business. And often [0:35] that makes us the bad guys and gals. Okay? And it's it's not because we want to slow work down. [0:42] All right? It's because we are accountable for the economics, the risk, and the result. [0:51] In the next 20 minutes, I'm going to show you where finance AI is producing measurable value, [1:00] what those use cases have in common, and the order I would follow Monday morning if I was [1:05] sitting in your chair. I'm not here to sell AI. I use it. I build with it. And as a CFO, [1:15] I'm accountable for controlling it. You can't automate what you can't see. [1:22] I'm in finance. So, we're going to start with two numbers. Okay. First one here. [1:26] Deote said 63% of finance functions have deployed AI. Yet, only 21% report clear, measurable ROI. [1:43] That second number is the whole story. Client has crossed the adoption line. Hey, we're over 50%, [1:49] right? But we are not yet crossed the value line. The teams getting measurable value [1:57] are not simply running more experiments. Okay? They're the ones choosing a clear problem, [2:04] using data they trust, naming an owner, setting the controls, and measuring what changed. [2:14] That's the difference between AI that actually changes the work [2:18] and AI that only is a good story at a conference reception happy hour. [2:24] And the reason why I see it this way is a lesson I learned way back in 2008. [2:30] And I learned this lesson well before generative AI. All right? I started my career at the Bank of [2:35] New York few years before the financial crisis. We were surrounded by data. Okay. uh positions, [2:44] prices, transactions, valuations, risk reports. I mean, I was a Bloomberg hero over here typing [2:51] away. And still, I watched major institutions get blindsided in real time. The information existed. [3:02] Data by itself is not visibility. And that gaps this whole talk. [3:10] All right. So, okay. The question here is why finance AI underperforms. [3:15] Okay, Grant Thornton surveyed nearly 1,000 senior financial leaders. 92% said, "Hey, I [3:23] I have at least one cause of AI underperformance." And we said, "All right, well, what is that?" [3:29] Somewhat surprisingly, at least to me, the most common sighted issue was actually governance [3:35] and compliance. Not they we didn't have the right AI model. [3:41] The model is rarely the constraint in today's world. All right? In finance, [3:45] the breakdown happens all around it. I know some of you in this room are here solving it. [3:51] So where is it breakdown? You have a unclear process you're trying to put AI on. [3:56] You have data that's not trusted within organization. Nobody clearly owns that output. [4:03] So you pop in a better model on top of a weak operating discipline, it just won't work. [4:10] Okay. Okay. Well, what do CFOs specifically in their role have to say? I don't want to hear what [4:16] senior finance leaders say. Okay. Uh, one of the CFO organizations that I'm on the board on, [4:21] we surveyed 230 mid-market CFOs and we say, "Hey, where do you think your biggest obstacle is with [4:30] AI?" Their number one response was prioritization. Okay, that was a surprise. Second and third talent [4:39] and capacity might be think oh what about technology and AI enablement both those technology [4:44] and I enable individually both below 4% of why the biggest obstacle for AI success and this is a [4:54] a little bit of a sad stat here only one in eight of CFOs entering this year are highly confident [5:01] that AI will deliver measurable ROI not good so AI is funded. We've seen here we're at the exhibit [5:08] hall. Uh but it's on a shorter leash. I know from my seat I got my eyes on the AI token burn. [5:14] Okay. And I've been listening to a couple uh earnings calls particularly JP Morgan Chase [5:18] where analysts are asking them now tell us about your your token burn. And so that will be a story [5:22] as we continue over the next earnings releases upcoming year. And then CFOs their biggest concern [5:29] with AI was actually you know where the data is going even inside enterprise models. So again, [5:36] it's not really, hey, what AI model should we use? It's more around that governance and ownership. [5:40] And so, all right. Okay. Well, tell me where is AI actually scale inside a finance process. [5:48] All right. So, AI does not, as we all know, replace the process underneath in the business, [5:54] at least for now. It sits on top, majority speaking. And so, it uses that existing process, [6:01] existing data to move faster. And I've talked to a few folks here where they've said, "Hey, [6:06] I've got folks that just rather start their own company, start AI fresh than buy a company [6:10] for the name and customers because it's too much tech debt." So, it's definitely [6:14] an interesting conversation there. But when the process is clear, the data is trusted, [6:19] review is defined, you get faster execution and more reliable output. Perfect world, right? [6:25] Or we're going to see that's not really the case. So, but when those processes are weak, [6:31] AI can move an error faster through the process. Okay? And it can spit out that error [6:35] in a way that is much more credible than it and than any automation in the f in the past. [6:40] And that matters because this this crazy sad here is that [6:44] in finance, we're starting from a messy messy point here. Over half of finance teams [6:50] still have no automation. I mean, they're relying on manual entry and spreadsheets. [6:54] I'm a little bit of a of a problem child as well. Okay. 86% say legacy systems limit AI readiness. [7:03] Okay. Now, I'm I'm a CFO for a private uh backed company and I came on board to help transform it. [7:08] We're doing workday financials, workday adapt to help get the company a better position to sell, [7:14] but it's not a reason to wait. Okay? I'm not saying this is reason to wait or replace your [7:18] whole tech stack, but it tells you where to start. If you're a CFO or business leader, [7:23] do not begin with the biggest AI ideas. All right? Start with a process finance already understands, [7:30] data it trusts, and an output someone can review. Otherwise, you'll be in a world of negative ROI. [7:42] So, this is where finance AI is already producing value. [7:46] All right, this came from a recent CFO peer study and a little bit of my experience in here as well. [7:51] Some of it's oper operational. I've got some reconciliations in there, invoice matching [7:56] and closed activities. You can see the value of the ROI that came along with it. [8:00] Some strength and control. You've got journal entry analysis, anomaly detecting. [8:05] And some of this more forward-looking stuff, stuff I really like is using AI to analyze trends, [8:10] seasonality, prior performance, and then spitting out scenario analysis. I I don't know about y'all, [8:16] but I love seeing all the millions of scenarios that could possibly ever happen. [8:21] 99% never do, but burn those tokens so I can see them, right? It's kind of crazy. Uh, but the [8:26] pattern across all these are the same. All right, so reoccurring work, known data, clear review. [8:32] AI it processes it prepares the work but you always have finance someone in there that owns the [8:38] assumptions the exceptions and the final decision and let me share one of these like let's go and [8:45] eat the hood on one of these uh what do we have account reconciliation here all right classic team [8:50] comparing two established financial data sources every period okay that could be a ledger account [8:56] against a statement or an internal transaction file against an external summary report [9:01] the workflow The AI workflow in this example or this peer study here that I spoke to the [9:06] CFO about standardized the files, applied matching rules, separated routine matches from exceptions [9:13] and show the team what they needed to look at and the output which is very important was not a [9:19] black box. Okay, it's it was a reconciliation file that the exception list of finance could review. [9:26] AI prepared the work. finance investigated the exceptions and approved the result and what it [9:31] came out to be was about 200 hours of returned uh work to the to this gentleman's team. I've seen on [9:37] my own team and I'm telling you this this AI use case is not sexy but you want to know what to CFOs [9:44] for yourself or the people who you talk to and try to sell to this is the stuff okay hours returned [9:49] is the stuff among other things so doesn't have to be the hottest topic you just got to get get [9:53] the right path and I'll kind of go how I think about and how maybe you can you can sell to a CFO [9:59] um and you know I'll make it personal here I've done examples before but this one in my current [10:04] company I mentioned a pback HR technology ology company. We needed one global financial [10:09] deck across all regions for our reporting P firm and internally. The project had been underway [10:16] for about 3 months and we had a contracted senior financial leader on it very capable. [10:21] We had AI involved and AI to involved but we did not have finished product after months and months [10:25] and months. So I had to step in. All right. I realized the remaining issue wasn't really just [10:31] hey just finish up the existing deck. All right. It was a structure itself that needed to change. [10:37] And so I started as I mentioned before some some folks here are saying I'm going to build a company [10:42] from scratch for getting on that tech debt. So I rebuilt it using AI first. Okay. Four days later [10:50] the new deck was finished and my regions were using it. And to be clear it was nights that [10:55] week and a weekend and more than 30 hours of real work. Okay. AI did not do the job for me. [11:01] And I'm I'm not advocating that they're going to do the job press of a button. But [11:04] three things made the difference. Okay. First, I understood the business context. Okay? And that's [11:10] going to be key as you going through all these meetings. Business context is key to the future. [11:14] We heard it in our keynote speak uh speech ear keynote earlier that we think over the [11:19] next decade business context will be king. We have a apparently two more decades left of of [11:24] being in the business of of having that context to help out. But I knew where the numbers came from. [11:30] I know which sources to trust in our company data sources that is and I do and I knew what [11:35] the deck needed to say for a leadership team to make a decision to make decisions. Second, [11:43] I had more experience working with AI. I knew how to break the work up into pieces, find the right [11:48] context, challenge the results, and keep refining, refining, refining until we got a very useful [11:53] model and report. Third, the tool fit mattered. Uh I'm not here to say it didn't. I used a tool that [11:58] was better suited for this type of presentation work. So AI helped me rebuild it and finish the [12:04] work much faster. Uh my CFO days I don't build a lot of reports but I knew what they needed [12:08] to look like and it did not replace the judgment required. So the business judgment is always key [12:17] and for me that made me think hey what what else can a CFO do without a software development team? [12:23] And so what I'm realizing and talking to y'all in my own business here is that the ability to build [12:28] is no longer really limited to technology teams themselves. I'm a CFO. I am not a software [12:33] developer. But I built and launched an AI business diagnostic for a personal project for an upcoming [12:39] book. All right. The whole stack tally Zapier claw with Sonnet 4.6 kit and resend. No data science [12:46] team. And okay, a real data science team would have done it way faster and cleaner. But that's [12:51] not the point. Okay. Okay, the point is that nonIT business leaders can get much closer to the build [12:58] than they ever could before. You know, but it's not magic. You know, I use AI to help me [13:03] connect the tools, troubleshoot the workflow, fix problems that came up, and there were many, many, [13:08] many more problems than I will admit on the stage. All right? But AI was, here's the important part, [13:13] but AI did not decide what questions to ask. It didn't decide how I should score those answers [13:20] back from my CEOs. and it didn't decide what a useful report would end up to be. [13:27] The logic had to come from the business side and that made all the difference in the output [13:33] being useful for CFOs to use for visibility in their business and just an AI slot report. [13:42] And the order I follow, this key, the order I followed in building that diagnostic is [13:46] the same order I would use Monday morning if I hadn't evaluated an a finance AI use case. [13:52] I would first I would see it. All right, map the work that actually happens in the business. [13:57] And there are companies out in the exhibit before that actually help you do this. [14:00] Second, clean it. So use the data you trust and name the person who owns it. I've been in [14:06] many leadership meetings where something comes up and you don't have an owner in that meeting [14:11] and it stalls a meeting. That is something I won't do anymore and I I recommend that you [14:15] don't. You always attach an owner to an outcome. Third, frame it. You want to define the decision, [14:20] the rules, the output, and what still requires approval. Yesterday, I was uh in a in a [14:28] a a presentation and the gentleman, his daughter was ordering burritos off of Alexa, and apparently [14:33] she didn't think that Alexa heard her. So he ended up with nine burritos at his door. So you always [14:37] have to make sure you you aware that autonomy is not authority. Fourth, build it. Start small. [14:43] Use the right tool for the for the job. And then fif five, fifth, refine it. Measure what changed. [14:49] Review the exceptions. Fix the weak points. And then scale. Turn that into a repeatable process. [14:57] And actually, I really like this. And this is for your folks that are selling it. If you [15:00] could talk across these five because the build starts at four, the value starts at step one. [15:06] So make sure you can express all through the process that will get CFO's ears. [15:11] And there is one operating consequence to this that we cannot ignore. I think about [15:15] a lot because I work for a workforce solutions company and we heard it earlier in the keynote [15:20] is and I'll speak just specifically from the finance side. AI is specifically good [15:25] at that first layer of finance work. Pulling reports, matching transactions, [15:31] preparing the first draft of analysis. But that is also where people learn the business. [15:39] That analyst pulling the report learned how the company makes money. That person preparing the [15:46] reconstellation learned where the controls broke. that person building the first forecast learn what [15:52] what assumptions actually mattered and move the needle forward in the business. [15:59] So when AI takes that over that task over the learning doesn't automatically survive [16:05] and the numbers show the gap 73% have deployed AI or pilot AI fewer than 10% [16:19] of the workforce is actually trained on it and then 20% have actually redesigned the role [16:27] and so I'm not Okay, we should not preserve that manual work just because it used to be [16:33] used for training people. But we do have to replace the learning that it created. Okay, [16:38] and I am standing up here as a CFO in charge of P&L saying we should spend more money in [16:42] learning and development. Uh this different conversation I had in the past few years, [16:47] but I am passionate about it. We should spend more money reskilling people. [16:54] And this is an interesting one. I like this one. So we spent a lot of time asking hey what [17:00] finance task can AI do and this is an important question okay but I think it's a bit too narrow [17:06] the bigger opportunity is to say how can redesign how work gets done in the first place okay that [17:14] doesn't mean redesign finance to remove people uh per se it means deciding where AI can prepare [17:21] analyze and execute and where human judgment approval and accountability still belong [17:27] AI changes how the work gets done, but people still own the decisions and the results. [17:32] And I I actually love this here. The future of finance is not AI added to old work. [17:37] Is finance work redesign around AI. And that is that's cool. And it's tough if you come [17:40] into an organization you're trying to reshape the transformation. It's [17:43] it is definitely a challenge, but that makes me feel like, hey, it changes the role of the CFO. [17:49] And so reporting the past still matters, okay? accuracy, control, financial integrity. They do [17:56] not go away. Probably scale up to be quite honest. But the CFO has to design the work, okay, with the [18:02] future in mind. They just not just cannot report the past and the results. And you're hearing it on [18:09] the investor calls. You'll be hearing it more and more. The teams that get measurable value from AI, [18:15] they do not begin by shopping for the best model. Although I do like the exhibit hall out there. [18:20] All right. Uh but the teams they've mapped the processes they clean the data name the [18:26] owner define the review and approval points and then they built and scaled. So that work is not [18:32] a delay. Don't let your company tell you oh we can't do this. We got to move fast. It's like we [18:36] got to do that initial work. I mentioned the value starts at step one when the build's at step four. [18:42] So it's going to it's going to be a difference between what delivers useful AI [18:46] and value. All right. Thank you AI4 CFO track and everyone in the room. Everything from today's [18:53] session is behind this QR code. Also in the deck there's practical slides in the appendix such as [18:59] choosing your first AI use case, AI verse automation, ROI, permissions, build versus buy, [19:05] security the diagnostic tool if you want to find out if your own business has some [19:10] visibility issues and my LinkedIn I talk about this sort of stuff from my operating CFO seat [19:15] every week. So scan it, follow uh love to connect. If you're CEO, CFO working on some cool stuff, [19:22] find me after. Hit me up after. Take care and enjoy lunch.