Pay-i CEO Tepper Outlines Framework to Quantify AI Agent ROI
Ai4
Enterprise AI agents only drive productivity when task completion rates exceed human verification time relative to manual effort, as variable token costs and latency disrupt traditional financial models.
AI-assisted summary of Measuring the Business Value of Agents With Pay-I, Inc — original source. Check the original for full context.
Source published:
Detailed notes
## Background This content was published on YouTube by the Ai4 channel on August 26, 2025. It features a talk by David Tepper, CEO of Pay-I, Inc., who addresses how enterprises can move past the hype of generative AI and accurately measure, quantify, and manage the real business value, cost, and ROI of AI agents.
## TL;DR David Tepper breaks down agentic software into workflows, pipelines, and fully autonomous systems, emphasizing that evaluating their worth requires focusing on three core buckets: financial value, time savings, and productivity. He explores the hidden cost and latency complexities of Large Language Models, including token pricing discrepancies, multimodal overhead, and context-window degradation. To help organizations systematically track ROI, Tepper introduces version-bound KPIs, mathematical formulas for time savings and verification overhead, and outlines how Pay-i helps businesses govern AI spend and measure real-world performance.
## Atmosphere & style The session carries an authoritative yet accessible, pragmatically grounded tone typical of an experienced enterprise software leader. Tepper blends technical expertise with dry humor and self-deprecating remarks about the stress and fast-paced nature of the AI field. The delivery feels like a high-level executive briefing that balances rigorous mathematical frameworks with practical advice for navigating the complexities of AI adoption.
## Key takeaways - [00:04](https://www.youtube.com/watch?v=0mbItBdqlko&t=4s) The speaker, David Tepper, brings deep enterprise AI expertise, having spent 19 years at Microsoft managing internal AI platforms and advising Fortune 500 boards, the U.S. Congress, and the UN. - [01:28](https://www.youtube.com/watch?v=0mbItBdqlko&t=88s) Agentic software falls into three distinct architectures: Workflows (code calling LLMs), Pipelines (fixed multi-step sequences like standard RAG), and Agentic Systems (autonomous runtime decisions and tool selection). - [02:48](https://www.youtube.com/watch?v=0mbItBdqlko&t=168s) Enterprise business value for AI agents should be measured across three core, quantifiable buckets: Financial Value (revenue/cost savings), Time Savings, and Productivity Gains. - [04:34](https://www.youtube.com/watch?v=0mbItBdqlko&t=274s) Calculating agent costs is highly complex due to tokenizer differences, undisclosed backend vision pricing models, SDK disparities, and compound execution overhead from multi-step reasoning. - [07:25](https://www.youtube.com/watch?v=0mbItBdqlko&t=445s) Evaluating AI time savings requires analyzing specific sub-components of GenAI latency, including Time to First Token (TTFT), Inter-Token Latency (ITL), Time to First Completion Token, and Output Tokens Per Second (OTPS). - [09:27](https://www.youtube.com/watch?v=0mbItBdqlko&t=567s) Agents must be evaluated by specific use cases and versioned strictly, as minor changes to system prompts, underlying data, or models drastically alter gross margins and operational KPIs. - [12:00](https://www.youtube.com/watch?v=0mbItBdqlko&t=720s) Over 99% of enterprise production KPIs for agents fall into four structural metric types: Booleans, Numbers, Percentages, and Likert rating scales. - [13:19](https://www.youtube.com/watch?v=0mbItBdqlko&t=799s) Quantifying true time savings requires incorporating a quality/satisfaction factor to account for agent failures or tasks completed without meaningful efficiency gains. - [14:53](https://www.youtube.com/watch?v=0mbItBdqlko&t=893s) An agent genuinely increases productivity only if its probability of success exceeds the time required to verify the output relative to the time needed for a human to perform the task from scratch. - [15:36](https://www.youtube.com/watch?v=0mbItBdqlko&t=936s) Pay-i provides a GenAI FinOps platform for tracking spend, setting governance, preventing runaway agent costs, and measuring ROI, including a strategic partnership with AWS on their VALUE initiative.
## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=0mbItBdqlko&t=4s) Introduction and Background of David Tepper
David Tepper introduces the session, *Measuring the Business Value of Agents*, thanking attendees for navigating a last-minute room change. Before jumping into the core framework, Tepper establishes his long-standing background in artificial intelligence. He has been working with Generative AI since 2009 and holds a patent filed back in 2011. During a 19-year tenure at Microsoft, he managed the internal AI software platform used by Azure’s 45,000 employees. He eventually turned down an offer to report directly to Microsoft's CTO—a role overseeing internal AI adoption across the entire enterprise—in order to found Pay-i.
Throughout his career, Tepper served as a primary AI advisor to Microsoft leadership, Fortune 500 executives, the U.S. Congress, and recently spoke at the United Nations regarding AI's impact on developing nation-states. He jokes self-deprecatingly about the complexity of the space showing a picture of himself prior to starting his startup journey, joking that he lost all his hair along the way.
- **Microsoft Tenure:** 19 years leading internal AI developer tool platforms. - **Enterprise Scale:** Managed internal AI infrastructure serving 45,000 Azure employees. - **Global Policy & Executive Briefings:** Advised Fortune 500 boards, U.S. Congress, and the United Nations.
---
[01:28](https://www.youtube.com/watch?v=0mbItBdqlko&t=88s) Defining "Agent": Workflows, Pipelines, and Agentic Systems
Tepper tackles the term "agent," pointing out that it has become an overused buzzword in tech marketing where "everything seems to be agentic." To establish a rigorous baseline, he categorizes agentic software architecture into three distinct operational models:
1. **Workflows:** Traditional software programs that make discrete calls to Large Language Models (LLMs) to handle unstructured-to-structured data transformations. Once structured, conventional code executes the remaining business logic. 2. **Pipelines:** Predetermined multi-step sequences incorporating LLMs, such as standard Retrieval-Augmented Generation (RAG) chatbots or fixed data-processing pipelines. These systems lack real autonomy because they follow a fixed path without making dynamic runtime decisions. 3. **Agentic Systems:** Truly autonomous systems that evaluate choices dynamically at runtime. They select and utilize tools interactively, resulting in non-deterministic execution paths and variable response times.
``` +-------------------------------------------------------------------+ | SOFTWARE MODELS | +-------------------------------------------------------------------+ | 1. WORKFLOW: Code -> LLM (Unstructured to Structured) -> Code | | 2. PIPELINE: Fixed Multi-Step Sequence (e.g., Standard RAG) | | 3. AGENTIC: Dynamic Runtime Decisions & Variable Tool Selection| +-------------------------------------------------------------------+ ```
---
[02:48](https://www.youtube.com/watch?v=0mbItBdqlko&t=168s) The Three Buckets of Business Value
To evaluate AI deployments, organizations must move beyond qualitative feelings—such as the "warm and fuzzy feeling" in non-profit or caregiving scenarios—and focus on universally quantifiable metrics. Based on interviews with hundreds of enterprise customers, Pay-i classifies business value into three primary categories:
- **Financial Value (Raw Currency):** Direct top-line revenue expansion or bottom-line cost reduction (e.g., automation savings). - **Time Savings:** Quantifiable reduction in operational effort required to complete specific tasks compared to legacy manual baselines. - **Productivity Gains:** Expressed as output per unit of time (e.g., increase in customer service tickets closed per representative per day).
---
[04:34](https://www.youtube.com/watch?v=0mbItBdqlko&t=274s) The Hidden Complexities of LLM and Agent Costing
Calculating standard ROI appears simple (Net Profit divided by Cost), but tracking the *cost* side of an AI agent is surprisingly intricate. Unlike standard software subscriptions, AI consumption costs fluctuate based on tokenizer variations, provider pricing models, and multimodal workloads.
For instance, multimodal vision calls (such as OpenAI's GPT-4o) depend on proprietary backend image-tiling calculations with variable pricing metrics. In API billing statements, token counts for vision calls are blended together, obscuring the precise cost attribution for visual processing versus standard text generation. Tepper notes that he maintains an 80-slide presentation dedicated solely to decoding GPT-4o cost management.
``` +-------------------------------------------------------------------+ | AGENT COST CALCULATION FACTORS | +-------------------------------------------------------------------+ | * Tokenizer Differences (Variation across model vendors) | | * Multimodal Inputs (Blended visual/text API pricing models) | | * SDK Discrepancies (Pricing variance based on access method) | | * Reserved Instance Tracking (Requires multi-dimensional math) | | * Compound Execution (Multi-step reasoning and tool calling) | +-------------------------------------------------------------------+ ```
- **SDK Pricing Variations:** Calling the exact same model via different SDKs or API abstractions can yield distinct pricing profiles. - **Reserved Capacity Complexity:** Allocating static reserved instances across dynamic dynamic agent requests requires multi-dimensional modeling. - **Delayed Provider Tooling:** OpenAI posted a note promising a dedicated tracking feature for reserved instance attribution in "Q2 of 2023," a milestone that remains unreleased years later. - **Compound Execution Overhead:** A single user request to an agent triggers multiple internal reasoning steps, sub-queries, and tool invocations, each incurring variable token costs.
---
[07:25](https://www.youtube.com/watch?v=0mbItBdqlko&t=445s) Dissecting GenAI Latency Across Execution Phases
When measuring time savings, basic end-to-end response time fails to capture the user experience or system behavior. Generative AI requires analyzing four distinct sub-components of latency:
1. **Time to First Token (TTFT):** The duration until initial output streaming begins. High TTFT impacts perceived responsiveness. 2. **Inter-Token Latency (ITL):** The speed at which subsequent individual tokens stream. ITL is typically much faster than TTFT because the model has already completed its initial prompt evaluation. 3. **Time to First Completion Token (TTFCT):** In reasoning models, this represents the internal thinking time spent solving a problem before emitting the user-facing answer. 4. **Output Tokens Per Second (OTPS):** The streaming generation throughput. Long-context models like Google Gemini can slow down noticeably as generation approaches context limits due to the compute cost of processing growing context windows.
To illustrate how reasoning models operate before emitting tokens, Tepper uses a maze analogy: if an LLM is asked to output the single fastest path through a maze with multiple entry points, it cannot stream the correct initial step until it has solved the entire path internally.
---
[09:27](https://www.youtube.com/watch?v=0mbItBdqlko&t=567s) Structuring Agent### [00:04](https://www.youtube.com/watch?v=0mbItBdqlko&t=4s) Introduction and Speaker Background
David Tepper, CEO of Pay-I, Inc., introduces the session on measuring the business value of AI agents, acknowledging a last-minute room change that attendees had to navigate. He sets an agenda covering his background, definitions of agents and business value, methodologies for measuring their combination, and concrete real-world examples. Tepper shares his extensive history in the generative AI space, noting that he began working on GenAI in 2009 and filed his first patent in 2011. He spent roughly 19 years at Microsoft, where he managed the internal software platform that enabled Azure's 45,000 employees to leverage generative AI. He eventually turned down a role reporting directly to Microsoft's CTO—a position where he would have overseen internal AI adoption across the entire company—in order to launch Pay-I.
During his tenure at Microsoft, Tepper regularly briefed internal executives on how generative AI would transform products and services. He also represented the company externally, speaking on GenAI to Fortune 500 leadership, academic institutions, the U.S. Congress, and the United Nations, where he presented six months prior on how AI will impact developing nation-states. He jokes about how AI transformed from a niche area into a massive industry explosion, humorously pointing out a photo of himself with hair prior to taking on the complicated, stress-inducing field.
- **Speaker Profile**: David Tepper, CEO of Pay-I, Inc. - **Key Milestones**: Working in GenAI since 2009; first patent filed in 2011; 19-year tenure at Microsoft. - **Key Engagements**: Briefed U.S. Congress, Fortune 500 boards, and the United Nations on AI impacts.
---
[01:28](https://www.youtube.com/watch?v=0mbItBdqlko&t=88s) Defining an Agent: Workflows, Pipelines, and Autonomous Agents
Addressing the hype surrounding the word "agent," Tepper jokes that the term has become so overloaded that people treat everything—including the podium or audience members—as "agentic." To cut through the buzzwords, he establishes a taxonomy that categorizes agentic software into three distinct technical definitions:
1. **Workflows**: Traditional software applications that invoke a Large Language Model (LLM) at specific points in their process. This pattern is commonly used to parse unstructured data into structured formats so traditional code can execute subsequent tasks. 2. **Pipelines**: Predetermined sequences of execution steps where LLMs handle specific parts. A standard example is a Retrieval-Augmented Generation (RAG) chatbot; while it uses LLMs, it is not strictly autonomous or self-directing because it executes a fixed, repeatable path. 3. **Truly Agentic Systems**: Autonomous systems that make runtime decisions, dynamically select and utilize different tools, and feature non-deterministic execution times and paths.
Tepper emphasizes that Pay-I's methodology for measuring business value applies universally across all three definitions, ensuring enterprises can evaluate value regardless of which architectural pattern they implement.
``` +-----------------------------------------------------------------+ | AGENTIC TAXONOMY | +-----------------------------------------------------------------+ | 1. WORKFLOW | | [Unstructured Data] --> [LLM Parsing] --> [Structured Code] | | | | 2. PIPELINE | | [User Query] --> [RAG Retrieval] --> [LLM] --> [Response] | | (Fixed, predetermined execution path) | | | | 3. TRULY AGENTIC | | [Goal] --> [Runtime Decision] --> [Tool Choice] --> [Output] | | (Dynamic execution, non-deterministic path and duration) | +-----------------------------------------------------------------+ ```
---
[02:48](https://www.youtube.com/watch?v=0mbItBdqlko&t=168s) Defining Business Value and its Three Buckets
Moving from technical definitions to business metrics, Tepper defines business value through quantifiable components rather than subjective "warm and fuzzy feelings." Based on conversations with hundreds of enterprise customers—primarily through the Office of the CIO—Pay-I categorizes quantifiable business value into three core buckets:
- **Financial Value (Raw Currency)**: Value generated via direct monetary metrics—either driving net-new revenue streams or cutting operational costs (jokingly referencing the popular tech trope "#AI is taking all of our jobs"). - **Time Savings**: Reductions in human task duration, measured by comparing the baseline time required for a human to complete an action versus the accelerated timeframe achieved using agentic tools. - **Productivity**: Measured as output generated per unit of time (e.g., the volume of customer support tickets processed within a single work shift).
---
[04:35](https://www.youtube.com/watch?v=0mbItBdqlko&t=275s) The Complexity of Calculating Agent Costs
Tepper breaks down the standard Return on Investment equation ($\text{ROI} = \frac{\text{Net Profit}}{\text{Cost}}$), explaining that while net profit stems from savings or revenue, calculating the *cost* side for LLMs and agents is notoriously complex. Standard software pricing model assumptions break down with LLMs: two models with identical published pricing doing identical tasks can yield wildly different costs because they utilize different tokenizers, or do not charge on a per-token basis at all.
Multi-modal operations introduce hidden layers of variable pricing. Using OpenAI's GPT-4o vision pricing algorithm as an example, Tepper explains that back-end calculations vary dynamically without explicit disclosure to developers. When receiving an API response, the system returns a unified token count without breaking down how many tokens were consumed by visual processing versus text processing, making fine-grained cost allocation difficult.
``` +-----------------------------------------------------------------+ | LLM AGENT COST COMPLEXITY | +-----------------------------------------------------------------+ | MULTI-MODAL HIDDEN COSTS: | | Image Input --> [Variable Back-end Algorithm] --> API Response| | (API returns total tokens without separating vision vs text) | | | | API & SDK DISPARITIES: | | Model X via SDK A <-- Price Discrepancy --> Model X via SDK B| | | | RESERVED INSTANCES: | | Requires multi-dimensional modeling to attribute cost per call| +-----------------------------------------------------------------+ ```
Tepper notes that he maintains an 80-slide deck dedicated solely to calculating the true cost of GPT-4o calls. He highlights several billing edge cases: - Calling the exact same model via different SDKs can result in varying price points. - Model snapshots carry distinct pricing tiers that change over time. - Calculating the unit cost per API call on OpenAI Reserved Instances requires complex, multi-dimensional financial modeling. - He cites a quote from OpenAI's website promising a tracking system for reserved capacity in "Q2 of 2023"—a deadline that passed long ago without resolution.
Furthermore, an agent's operational cost cannot be computed from a single API call. Agents execute non-deterministic sequences of reasoning steps, tool calls, and model invocations, leading to fluctuating token counts and dynamic costs across every execution run.
---
[07:25](https://www.youtube.com/watch?v=0mbItBdqlko&t=445s) Understanding Generative AI Latency Metrics
To accurately calculate time savings, organizations must understand the unique nuances of GenAI latency. Unlike standard software applications where end-to-end latency is the primary metric, generative models introduce several granular latency types:
- **Time to First Token (TTFT)**: Crucial for streaming responses, TTFT measures how quickly the user perceives responsiveness. Tepper shares a counterintuitive insight: research shows LLMs frequently solve the core logic of a prompt *before* outputting the very first token, spending the remainder of the streaming time generating explanations. He compares this to solving a complex maze with multiple exits: the model must calculate the full path internally before it can begin outputting the first correct direction. - **Inter-Token Latency**: The speed at which subsequent tokens are output during streaming, which is typically significantly faster than TTFT because the core logic has already been resolved. - **Time to First Completion Token**: The duration a reasoning model spends "thinking" or processing internal step-by-step logic before it generates the first visible completion token. - **Output Tokens Per Second (OTPS)**: A metric that degrades as context windows grow. Tepper points out Google's Gemini as an example, noting that models with ultra-large context windows (e.g., 1 million tokens) can experience significant performance degradation near the end of the context window due to the computational overhead of processing vast amounts of prior context for each new token generated.
---
[09:28](https://www.youtube.com/watch?v=0mbItBdqlko&t=568s) Structuring Use Cases, Versioning, and Key Performance Indicators
Because the underlying mechanics of GenAI are complex, Tepper advises organizations to evaluate agents by **use cases and tangible outcomes** rather than underlying technology architectures. A use case defines the core objective—such as summarizing a contract or booking a travel itinerary—and must be measured by execution time, total cost, and degree of success.
Tepper cautions that generative AI systems are hyper-sensitive to minor configuration changes. Altering an English prompt, updating underlying data, swapping a model, or tweaking application code drastically shifts performance dynamics. A developer inadvertently adding a single comma to a system prompt can compromise product gross margins, increase reasoning latency, or lower response accuracy. Because of this sensitivity, every metric must be rigorously tied to a specific **version** of the agent.
``` +-----------------------------------------------------------------+ | AGENT VERSION CONTROL | +-----------------------------------------------------------------+ | [System Prompt v1.2] + [Model Version] + [Data Pipeline] | | | | | v | | Evaluated against Quantifiable KPIs (Booleans, Numbers, etc.) | | | | | v | | Yields Granular ROI Metrics (Cost, Latency, Accuracy) | +-----------------------------------------------------------------+ ```
When evaluating enterprise use cases, Tepper excludes terms like "copilot" or "agent," noting that those represent technical tools rather than outcome-driven scenarios. Based on Pay-I's study of enterprise deployments, industry-standard KPIs fall into four core structural types:
- **Booleans**: Binary success/failure flags (e.g., true/false). - **Numbers**: Discrete numeric counts (e.g., count of items processed). - **Percentages**: Accuracy or completion rates (e.g., % of tickets resolved without human intervention). - **Liker / Likert Scales**: Rating scales (e.g., 1-to-5 or 1-to-10 quality ratings).
Tepper notes that these four KPI formats encapsulate over 99% of all production metrics used by enterprises today. By continuously binding versioned execution data to these quantified KPIs, enterprises can measure business value accurately and make data-driven engineering decisions.
[13:19](https://www.youtube.com/watch?v=0mbItBdqlko&t=799s) Calculating Financial Value and Time Savings Equations
David Tepper returns to the core ROI equation to demonstrate how business logic and quantifiable KPIs—such as click-through rates and conversion rates—populate the numerator (e.g., cost savings, revenue generation) and denominator. Beyond basic ROI, evaluating real efficiency requires a more nuanced time-savings equation:
$$\text{Time Savings} = \sum_{i} \left( n_i - (s_i \cdot a_i) \right)$$
In this equation, $s$ serves as a binary satisfaction indicator ($1$ if completed satisfactorily, $0$ if not), $n$ is the standard human execution time, and $a$ is the agent's completion time. Summing these metrics across all items (such as processed documents) accounts for scenarios where an agent is performing exceptionally, completely failing, or succeeding without yielding meaningful time savings.
- **Quality Rating Impact:** Evaluates task satisfaction ($s$) across variables like document size to quantify true time savings across batch operations.
---
[14:53](https://www.youtube.com/watch?v=0mbItBdqlko&t=893s) Defining the Golden Rule of Agentic Productivity
When evaluating true productivity gains, Tepper summarizes the governing principle behind AI-driven workforce efficiency into a core mathematical rule:
$$\text{Productivity Gain Condition: } P(\text{Success}) > \frac{T_{\text{verify}}}{T_{\text{human}}}$$
An agent genuinely increases human productivity only if its probability of completing a task successfully $P(\text{Success})$ exceeds the time required for a human to verify the output ($T_{\text{verify}}$) divided by the time it takes for a human to perform the task manually from scratch ($T_{\text{human}}$). If verification overhead is too high relative to task duration or the agent's success rate is low, the system creates a net loss in productivity.
- **Key Takeaway:** If verification time relative to manual effort outweighs the probability of task success, the agent acts as a net drag on organizational efficiency.
---
[15:36](https://www.youtube.com/watch?v=0mbItBdqlko&t=936s) Summary of Agent Architectures, Business Value, and Pay-i Overview
Summarizing the talk, AI agents fall into three primary categories: simple LLM-enhanced workflows, multi-step sequential pipelines, and fully autonomous agentic systems utilizing Model Context Protocol (MCP) servers and external tools. Quantifying their enterprise value requires categorizing outcomes into direct financial value, time savings, and productivity gains. Managing these deployments in production demands tracking operational variables like failure rates and capacity planning.
``` +---------------------------------------------------------------+ | Agent Architectures | | [LLM Workflows] -> [Sequential Pipelines] -> [MCP Agents] | +-------------------------------+-------------------------------+ | v +---------------------------------------------------------------+ | Business Value KPIs | | [Financial Value] | [Time Savings] | [Productivity] | +-------------------------------+-------------------------------+ | v +---------------------------------------------------------------+ | Pay-i Platform | | [Real-Time Spend] | [FinOps Governance] | [AWS VALUE] | +---------------------------------------------------------------+ ```
Pay-i provides enterprise solutions to track and optimize these metrics, highlighted by recent recognition from *Forbes* and the FinOps Foundation as a dedicated Generative AI FinOps platform. Features include financial governance and real-time spend protection to prevent runaway costs from looping agents. Additionally, Pay-i launched an official partnership with AWS on their "VALUE" initiative (featuring an acronym structure reminiscent of Marvel's S.H.I.E.L.D.), pairing AWS teams with Pay-i tools to guide enterprises through use-case definition, KPI selection, and ROI optimization.
- **FinOps Recognition:** Identified by the FinOps Foundation as the premier GenAI FinOps solution for cost and ROI management. - **AWS Partnership:** Official collaboration via the AWS "VALUE" program to guide enterprise AI adoption and measure ROI. - **Exhibitor Details:** Pay-i demonstrated these capability suites live at Booth 246.
---
[18:06](https://www.youtube.com/watch?v=0mbItBdqlko&t=1086s) Q&A: ROI vs. Deployment Focus and Prioritizing Value Metrics
During the Q&A session, Tepper addresses questions regarding Pay-i's core focus and how organizations should prioritize competing value metrics:
* **ROI vs. Deployment Focus:** Tepper clarifies that Pay-i focuses explicitly on ROI rather than general agent deployment or building tools. While numerous vendors assist in building or deploying agents, Pay-i identifies which specific workflows yield tangible business returns. * **Prioritizing Value Metrics:** When choosing between financial value, time savings, and productivity, prioritization depends entirely on the specific use case and business objectives. For revenue-generating applications, direct financial metrics take precedence; for internal operational tools, time savings and productivity metrics dominate. Organizations should measure all three dimensions across their portfolios, as there is no single "silver bullet" metric for every AI initiative.
## Notable quotes > The hype era of “trust me, it’s working” is ending, and in its place you’ll need the tools to manage the workforce of tomorrow. — David Tepper
> [01:28](https://www.youtube.com/watch?v=0mbItBdqlko&t=88s) everything seems to be agentic. — David Tepper
> [02:48](https://www.youtube.com/watch?v=0mbItBdqlko&t=168s) warm and fuzzy feeling — David Tepper
> [02:48](https://www.youtube.com/watch?v=0mbItBdqlko&t=168s) #AI is taking all of our jobs — David Tepper
> [04:35](https://www.youtube.com/watch?v=0mbItBdqlko&t=275s) Q2 of 2023 — David Tepper
## Mentioned David Tepper, Pay-I, Inc., Microsoft, Azure, U.S. Congress, United Nations, OpenAI, GPT-4o, Google Gemini, Forbes, FinOps Foundation, AWS, Model Context Protocol (MCP)
## Video description David Tepper, CEO, Pay-I, Inc.
Abstracts
The new world of software promises to be teeming with agents and AI powered tools… but how will that impact your business? You've built the POCs and seen the demos, but scaling to production demands answers: What is the real cost per outcome? Which use cases deliver provable ROI? As enterprises race to deploy AI agents and chatbots, boards are demanding quantifiable business value metrics. The hype era of “trust me, it’s working” is ending, and in its place you’ll need the tools to manage the workforce of tomorrow. I’ve been working on AI solutions for the largest companies in the world since 2009, and in this talk I’ll show you how Pay-i is helping companies quantify the real value of every agentic workflow, align teams around the new metrics that drive business growth, and forecast margins before shipping.
[Source](https://www.youtube.com/watch?v=0mbItBdqlko)