Why AI Agents Can Beat the Incumbents
Elena Burger, Seema Amble, Vladimir Keil
AI-assisted summary of Why AI Agents Can Beat the Incumbents — original source. Check the original for full context.
Source published:
Key ideas
- Incumbent software giants like SAP and Salesforce own static enterprise records, but miss the 80% of unstructured operational labor—spanning hundreds of emails, CAD files, and spreadsheets—required to reach a final transaction price. (4:55)
- Enterprise trust in autonomous agents is earned incrementally, starting with unmanaged tail spend under $50,000 where downside risk is low, before advancing to complex direct sourcing negotiations. (16:51)
- Internal enterprise AI builds face a severe 'DIY trap': achieving 70% to 80% baseline accuracy does not equal 70% automation, because human teams must still audit every transaction until edge-case integration is complete. (40:00)
- A 1% savings in enterprise procurement flows directly to the bottom line, delivering the equivalent P&L impact of a 10% increase in gross revenue across capital-intensive physical supply chains. (55:00)
Detailed notes
## Background Published on October 2, 2026, via RSS podcast distribution, this episode is hosted by a16z investors Elena Burger and Seema Amble, who interview Lio co-founder and CEO Vladimir Keil. The episode examines how AI-native startups can outmaneuver enterprise software incumbents by orchestrating the complex, unstructured workflows that take place outside traditional ERP systems of record.
## TL;DR Vladimir Keil explains that while legacy systems like SAP and Salesforce capture static transaction endpoints, up to 80% of actual procurement labor occurs across emails, spreadsheets, and cross-departmental negotiations. Lio tackles this through multi-agent architectures that scale across four tiers of autonomy, starting with low-risk tail spend to build enterprise trust before assisting in multi-million-dollar direct manufacturing sourcing. Keil also warns against the internal enterprise build trap—clarifying that 70% model accuracy does not equal 70% workflow automation—and explains how operational dependency, rather than a passive database record, forms the true defensible moat for vertical AI.
## Atmosphere & style The discussion is rigorous, highly pragmatic, and stripped of generic AI hype, resembling an insider strategic briefing between venture investors and an operator in the trenches. Keil brings an engineer's blunt candor to the conversation, contrasting the mundane reality of spreadsheets and procurement friction against the massive financial stakes of stalled production lines. The dynamic is intellectually brisk and collaborative, moving methodically from organizational psychology and software architecture to macroeconomic supply chain dynamics.
## Bulletin - `04:55` Incumbent software giants like SAP and Salesforce own static enterprise records, but miss the 80% of unstructured operational labor—spanning hundreds of emails, CAD files, and spreadsheets—required to reach a final transaction price. - `16:51` Enterprise trust in autonomous agents is earned incrementally, starting with unmanaged tail spend under $50,000 where downside risk is low, before advancing to complex direct sourcing negotiations. - `40:00` Internal enterprise AI builds face a severe 'DIY trap': achieving 70% to 80% baseline accuracy does not equal 70% automation, because human teams must still audit every transaction until edge-case integration is complete. - `55:00` A 1% savings in enterprise procurement flows directly to the bottom line, delivering the equivalent P&L impact of a 10% increase in gross revenue across capital-intensive physical supply chains.
## Key takeaways - `02:30` Systems of record record business outcomes but fail to execute the unstructured, cross-department workflows required to produce them. - `04:55` Procurement line items in legacy ERPs flatten weeks of work, dozens of meetings, and hundreds of emails into a single static number. - `06:22` AI agents divide across four functional tiers—retrieval, process, policy, and principal—with legacy software vendors largely stuck at the lower two levels. - `10:00` Incumbent software vendors face an innovator's dilemma: selling autonomous work conflicts directly with seat-based models and risks existing enterprise trust if early agents fail. - `11:51` Standard software solves the happy path of invoice syncing, which represents only 20% of the actual manual effort spent handling non-standard exceptions and fraud. - `16:51` Enterprises build operational trust by initially granting agents full autonomy over tail spend under $50,000, where manual procurement teams lack capacity to negotiate. - `20:00` Direct industrial procurement in aerospace, robotics, and data centers carries extreme physical stakes where single-component delays cause nine-figure factory halts. - `23:25` Predicting physical supply chain disruptions requires moving beyond internal ERP schedules to analyze live external signals like news and prediction markets. - `26:17` True end-to-end task automation requires multi-agent systems because human procurement workflows span multiple departments, stakeholders, and software tools. - `28:01` Requisitioners frequently bypass complex ERP interfaces, making multimodal intake of photos, PDFs, and spreadsheets essential for enterprise adoption. - `32:16` In direct procurement negotiations, up to 90% of the leverage comes from analytical pre-computation and live copilots benchmarking commodity indices and bill-of-materials costs. - `34:04` Closing the final 20% accuracy gap on proprietary industrial engineering tasks requires custom evaluation harnesses and outcome-based fine-tuning. - `36:34` Defensibility for vertical AI startups stems from operational dependency, as replacing an agent actively performing work creates far higher friction than swapping a static database. - `40:00` Achieving 70% accuracy on a prototype does not deliver 70% automation; closing the final 20% gap requires 80% of the total engineering effort. - `44:30` Bilateral automation between buyers and suppliers streamlines pre-negotiation spec analysis and compliance tracking where both parties share an interest in reducing cycle times. - `50:00` Forward-deployed engineering models in vertical AI succeed only when engineers are measured on KPIs that automate their own deployment workflows into the core product. - `55:00` A 1% reduction in corporate procurement spend delivers the same bottom-line P&L impact as a 10% increase in revenue, creating a massive economic opportunity in physical industries.
## Detailed walkthrough ### `00:00` Introduction: The High Stakes of Enterprise Procurement
In complex manufacturing environments, a single missed message can trigger massive financial fallout. If an aerospace enterprise is building an aircraft, it relies on thousands of external suppliers. Should a supplier email a routine notice that a single component will arrive two weeks behind schedule, overlooking that message can stall assembly lines and cascade into hundreds of millions of dollars in losses. While procurement was historically treated as an isolated back-office function, modern supply chains intersect directly with legal, finance, enterprise software platforms, and human operators across every department.
```mermaid flowchart TD accTitle: Incremental Trust Progression for Procurement AI Agents accDescr: Diagram illustrating how procurement AI agents evolve from human-in-the-loop oversight to handling larger autonomous negotiations.
subgraph S1["Human-in-the-Loop Foundation"] A["Supplier Interactions & RFQs"] --> B["Human Review & Feedback Loop"] B --> C["Agent Model Refinement"] end
subgraph S2["Autonomous Negotiation Milestones"] C --> D["$10K Negotiations"] D --> E["$20K Negotiations"] E --> F["$100K+ Negotiations"] end ```
Deploying autonomous agents into these environments requires navigating steep enterprise trust hurdles. Large organizations will not hand commercial negotiations over to an algorithm on day one. Instead, systems must operate with a human in the loop, absorbing feedback and operational nuances until stakeholders gain confidence in the technology's judgment. As Vladimir Keil explains: “No company and no enterprise starts with fully autonomous negotiation agents from day one.” Through this supervised learning curve, agents gradually graduate from low-stakes tasks to handling consequential negotiations valued at $10,000, $20,000, and eventually upwards of $100,000.
- Procurement mistakes in heavy industries like aerospace jeopardize hundreds of millions of dollars over single component delays. - Enterprises demand human-in-the-loop governance to calibrate agent responses before granting operational autonomy. - Trust accumulates incrementally, scaling deal thresholds as agents demonstrate reliable commercial judgment.
---
`02:30` Systems of Record vs. The End-to-End Job
The rapid advancement of foundation models has intensified the strategic debate between enterprise incumbents and AI-native challengers. Historically, startups competed directly against legacy platforms in an arms race of engineering and market capture. Today, market dynamics are complicated by incumbent software providers embedding foundation models directly into their existing products—such as Salesforce integrating Claude into its suite. Because incumbents already possess the customer relationship, proprietary database access, and deeply entrenched employee workflows, application-layer startups face urgent skepticism regarding their defensibility.
Despite these incumbent advantages, a system of record fundamentally documents business outcomes rather than executing the messy workflows required to produce them. As Seema Amble argues: “The legacy incumbent is limited to their system of record and that record that they have, and they're not completing the end-to-end job.” For instance, when a customer calls support after being billed following an account cancellation, solving the issue cannot be accomplished simply by referencing a single database entry. The resolution requires synthesizing chat logs, cross-referencing contract provisions, verifying billing ledgers, and issuing credits. Incumbent software remains compartmentalized within its specific database silo, whereas AI-native startups have the opportunity to orchestrate the entire workflow from initial customer inquiry through back-office remediation.
- Incumbents wield massive structural moats in data custody, established customer trust, and workflow adoption. - Bundling foundation models into existing enterprise suites (e.g., Salesforce with Claude) raises the competitive bar for outside software. - The primary strategic wedge for AI startups lies in end-to-end workflow ownership across fragmented systems rather than maintaining static database records.
---
`04:55` The Hidden Iceberg: Why Procurement Lives Outside the ERP
Enterprise resource planning (ERP) platforms present a deceptive illusion of simplicity. An ERP entry might display a single, tidy line item—such as an aluminum order finalized at $8,000. In reality, that final number represents the tip of an operational iceberg. Behind a single signed purchase order sit weeks of unstructured, collaborative labor that legacy software never captures.
```mermaid flowchart LR accTitle: The Procurement Workflow Iceberg accDescr: A comparison showing the visible ERP record versus the extensive unstructured labor happening outside of it.
subgraph Invisible["Unstructured Work Outside ERP"] direction TB M["30 Stakeholder Meetings"] E["500 Email Exchanges"] S["20 Spreadsheets"] CAD["3 Weeks of 3D CAD Modeling & Cost Engineering"] P["Supplier Pushback ($10K asking price)"] end
subgraph Visible["ERP System of Record"] R["Final Entry: Aluminum @ $8K"] end
Invisible -->|"Culminates in"| Visible ```
To secure that $8,000 price, internal teams may have navigated roughly 30 internal stakeholder meetings, exchanged more than 500 emails, iterated across 20 distinct spreadsheets, and countered a supplier pushing for $10,000. Furthermore, specialized cost engineers often spend up to three weeks analyzing 3D CAD models and financial projections to determine what a component ought to cost before any commercial negotiation even begins. Vladimir Keil highlights this disparity between the record and the reality: “So like most of the work in procurement actually happens outside of this ERP or any system of record.” Because incumbent systems only log the final transaction, they remain blind to the communication, engineering context, and negotiation dynamics where AI agents can deliver the greatest operational value.
- Final ERP entries flatten weeks of cross-functional friction into static prices. - Sourcing decisions routinely involve hundreds of emails, dozens of spreadsheets, and extensive engineering modeling before entering enterprise systems. - The vast majority of procurement context and negotiation takes place entirely outside of existing software records.
---
`06:22` The Four Tiers of AI Agents and the Incumbent Bottleneck
To evaluate how software platforms evolve beyond rudimentary chatbots, Seema Amble outlines a taxonomy of four distinct agent architectures: retrieval agents, process agents, policy agents, and principal agents. Each tier demands a progressively higher degree of autonomy, contextual synthesis, and qualitative judgment.
```mermaid flowchart TD accTitle: The Four Tiers of AI Agents accDescr: Hierarchical progression of agent capabilities from basic information retrieval to strategic commercial judgment.
Tier1["1. Retrieval Agent<br>• Chatbots & Search<br>• Pulls contract terms & verifies service outages"] Tier2["2. Process Agent<br>• Deterministic Execution<br>• Executes predefined refund workflows without human intervention"] Tier3["3. Policy Agent<br>• Interprets Ambiguity<br>• Determines if a 20-minute downtime constitutes a significant outage"] Tier4["4. Principal Agent<br>• Strategic Decision-Making<br>• Exceeds standard policy to protect strategic client relationships"]
Tier1 --> Tier2 --> Tier3 --> Tier4 ```
Using a telecommunications service outage as an illustrative benchmark, the functional differences across these four tiers become clear:
* **Retrieval Agents:** Function as conversational search layers over a database. When a customer calls regarding an outage, the retrieval agent pulls up contract provisions, confirms the incident occurred, and summarizes relevant account details without taking autonomous action. * **Process Agents:** Execute deterministic, rule-based tasks without human discretion. Once outage criteria are verified, the process agent automatically triggers the billing credit specified in the contract. * **Policy Agents:** Navigate ambiguity and exercise interpretation. If an outage lasted 20 minutes, the policy agent assesses whether the disruption qualifies as "significant" under ambiguous service-level agreements and determines whether compensation is warranted. * **Principal Agents:** Exercise high-level strategic reasoning to maximize overarching business objectives. A principal agent might decide to offer restitution far beyond contractual requirements to preserve a multi-million-dollar enterprise relationship after a severe service failure.
While major software vendors market agentic tools like Agentforce by emphasizing their potential to transform business processes, incumbent products remain predominantly constrained to retrieval and basic process execution. Incumbents benefit from vast distribution channels and pre-existing customer trust—allowing them to deploy features at minimal marginal cost—yet they face deep structural limitations. They are fundamentally held back by legacy system boundaries and internal revenue incentives that discourage them from venturing into subjective, high-liability decision-making tiers.
`10:00` Incumbent Conflicts and Lio's Spectrum of AI Agents
When legacy software vendors attempt to build advanced agents, they run directly into internal structural conflicts. Selling a product that provides workflow tools for a human support agent is fundamentally different from selling a solution that resolves the customer support issue end-to-end. As Seema Amble and Elena Burger point out, those two offerings target completely different buyers, pit different internal VP-led organizations against each other, and create conflicting incentive structures over what product is actually being sold.
```mermaid flowchart TD accTitle: Agent Spectrum by Complexity and Human Involvement accDescr: Progression from simple retrieval agents to high-risk autonomous agents with human-in-the-loop oversight. subgraph LowRisk["Low Risk / High Volume"] A["Retrieval Agents<br/>(Extract & Search Documents)"] --> B["Process Agents<br/>(Match Records, Update ERPs)"] end subgraph HighRisk["High Risk / Complex Reasoning"] B --> C["Policy & Exception Agents<br/>(Flag Fraud & Discrepancies)"] C --> D["Autonomous / Long-Running Agents<br/>(Tail Spend & Multi-Million RFQs)"] D -.->|"Oversight & Feedback"| E["Human-in-the-Loop<br/>(Engineers & Procurement Leads)"] end ```
When asked where Lio sits along the spectrum ranging from retrieval agents to process, policy, and principal agents, Vladimir Keil explains that the company spans across all of them depending strictly on risk and complexity. In lower-risk scenarios, Lio operates fully autonomously, whereas high-stakes workflows retain a human in the loop based on budget approval thresholds and operational risk. Addressing incumbent trust, Vlad contrasts existing internal reputation with external trust:
> “Incombents already have to trust, but this also means they like they can destroy the trust if they ship something like too early and the product maybe doesn't work or it's bad or it's like works decisions in a bad way.”
- **Organizational Friction in Incumbents:** Selling workflow software to augment human teams conflicts directly with selling agentic software that replaces or resolves the task, forcing different internal VPs to compete for budget and territory. - **Spectrum Coverage:** Lio deploys retrieval, process, policy, and autonomous principal agents, gating autonomy by budget authorization limits and deal complexity. - **The Trust Dynamic:** While startups must painstakingly earn customer confidence through proven reliability, incumbents already possess that trust—putting them at constant risk of squandering it if their initial agentic features fail or make damaging decisions.
---
`11:51` Invoice Processing: Why Software Covers Only 20% of the Work
Examining standard accounts payable and invoice matching highlighted the stark gap between legacy software coverage and actual enterprise labor. Traditional invoice software performs a predictable, mechanical process: it retrieves data from an incoming document, matches line items across corresponding purchase orders and receipts, and syncs the reconciled transaction back into systems like SAP or Oracle. While incumbent enterprise software was constructed to automate that straight-line flow, that "happy path" represents only a fraction of the real-world operational burden.
> “And then we figure out, okay, that's 100% of the software market. So that's invoice software. That's how you build it today. But it's actually only 20% of the work of the job to be done, of the problem.”
The remaining 80% of human effort is consumed by edge cases and exception handling: investigating fraudulent invoices, resolving pricing and quantity mismatches, and handling non-standard vendor terms. Lio quickly evolved from basic process automation into handling those complex exceptions. Vlad reflects that Lio benefited from being slightly ahead of the curve, starting three years earlier just before the breakthrough release of ChatGPT. While basic document retrieval seemed revolutionary to enterprise buyers three years ago, Lio used that initial foothold to pitch subsequent generations of process agents, autonomous agents, and long-running workflows.
- **The 80/20 Disconnect:** Standard invoicing software automates 100% of the predictable mechanical path, but that represents just 20% of the actual human labor required to manage invoices end-to-end. - **The Real Workload:** Enterprise teams spend the vast majority of their time on exception handling—uncovering fraudulent billings, reconciling missing purchase orders, and managing contract discrepancies. - **Rapid Iteration Cycle:** Lio tested market appetite by pitching advanced agent capabilities directly to enterprise buyers, confirming commercial demand, and shipping production deployments weeks later.
---
`14:04` Convincing Enterprises to Trust AI-Led Negotiations
Procurement is far more intricate than consumer purchasing; it is an economy-driving engine involving cross-functional stakeholders spanning finance, legal, cost engineering, and supply chain operations. Trusting an agent with internal back-office data is one hurdle, but allowing an external-facing AI agent to conduct negotiations directly with third-party suppliers represents a massive leap in enterprise risk.
```mermaid flowchart LR accTitle: Lio Progressive Enterprise Trust Model accDescr: Stepwise deployment from human-in-the-loop pilot to high-value autonomous negotiations. A["Initial Pilot<br/>(Human-in-the-Loop Evals)"] --> B["Low Spend Autonomy<br/>($10K – $50K Tail Spend)"] B --> C["Mid-Tier Spend<br/>($100K+ Sourced Deals)"] C --> D["Multi-Million Complex RFQs<br/>(Long-Running Agent + Expert Loop)"] ```
Enterprises are deeply skeptical of software vendor promises due to past disappointments with legacy tools that failed to deliver on marketing claims. Vlad notes that Lio manages this apprehension by implementing rigorous evaluations and an incremental, human-in-the-loop ramp-up:
> “No company and no enterprise starts with fully autonomous negotiation agents from day one. No one does that. Why? Because they don't trust us and they don't trust the technology from day one.”
Lio seeds each deployment with a standardized negotiation baseline, but adapts to the idiosyncratic operational practices of individual Fortune 10 enterprises by routing decisions through human operators. As procurement managers review and refine the agent's proposed counter-offers, the system captures institutional feedback. This gradually earns clearance to autonomously run $10,000, $20,000, and $100,000 negotiations. For multi-million-dollar strategic deals requiring hours of compute to inspect intricate 3D CAD models, Lio pairs the long-running agent directly with human cost engineers at critical milestones.
- **Cross-Functional Complexity:** Enterprise procurement decisions are rarely isolated to purchasing leads; they demand input and sign-off from legal, finance, and specialized engineering divisions. - **Overcoming Enterprise Skepticism:** Buyers carry significant fatigue from legacy vendors overpromising AI capabilities, requiring Lio to prove functionality via strict evaluation metrics rather than marketing claims. - **Escalating Financial Scope:** Trust expands incrementally—progressing from low-risk $10,000 orders to $100,000 purchases as the agent absorbs company-specific organizational patterns through human feedback. - **Expert-in-the-Loop for Large Contracts:** Multi-million-dollar deals deploy multi-hour, long-running agents that analyze complex 3D technical files, pausing to pull in specialized cost engineers before executing the next negotiation phase.
---
`16:51` Autonomous Tail Spend vs. Human-in-the-Loop High-Stakes Sourcing
The degree of autonomy Lio deploys is dictated by the category of spend and the underlying supplier relationship. Historically, large enterprises ignored all purchases below roughly $50,000 because procurement organizations simply lacked the headcount to review them. This created a recognized loophole in corporate spending:
> “Like this is maybe a hack for like other startups. You can just send an enterprise an invoice for 40K. They will probably not negotiate because they don't have the capacity to do so. Except they have Leo agents, then we are going to negotiate against you.”
For unmanaged tail spend under $50,000, the downside of utilizing an AI agent is virtually zero; because enterprises never negotiated those invoices in the first place, any savings the agent captures represents net-new value. If an agent occasionally miscalculates a low-value transaction, the business is no worse off than it was under manual neglect.
However, financial cost is not the sole variable in supplier management. Certain specialized expenditures—such as high-touch marketing services or dedicated audio engineers setting up a podcast studio for a venture firm like Andreessen Horowitz—involve institutional context and delicate vendor relationships. In those situations, buyers care deeply about tone of voice, partnership goodwill, and contract design rather than extracting every dollar of margin. For those sensitive interactions, Lio keeps procurement personnel actively in the loop to review messaging and protect strategic rapport, rather than letting the agent optimize purely for cost reduction.
- **The Unmanaged Tail Spend Hack:** Enterprising vendors often send $40,000 invoices knowing legacy enterprises will pay them without contest due to severe procurement capacity bottlenecks. - **Asymmetric Risk in Tail Spend:** Deploying autonomous agents against sub-$50,000 spend carries almost zero downside risk because companies previously left those transactions completely unnegotiated. - **Relationship-Driven Procurement:** Low-dollar expenditures tied to strategic partners, creative contractors, or specialized operational teams (e.g., podcast production) require human oversight to maintain goodwill and preserve the appropriate tone of voice. - **Contract Design Nuance:** Beyond top-line price concessions, agent workflows must navigate legal safeguards, service-level agreements, and long-term contract structures.
`20:00` Beyond Back-Office: Cross-Functional Complexity in High-Stakes Procurement
Procurement is far from an isolated administrative silo; it directly intersects with legal contract structures, finance, cost engineering, and commercial supplier relations. Vladimir Keil emphasizes that the people running these operations are not low-level back-office staff, but deeply specialized operators possessing intricate knowledge of proprietary workflows, enterprise constraints, and industry-specific regulations. These operators actively feed their domain insights into Lio’s long-running agents. Seema Amble and Elena Burger note how procurement has expanded from a historical back-office box into a cross-cutting nerve center touching legal, finance, and diverse software systems. When asked to ground this complexity in a tangible vertical—such as manufacturing humanoid robotics, drones, or commercial aircraft—Vlad makes it clear that Lio is not built for routine office supplies.
Vlad dismisses commodity procurement as long-solved table stakes: “When we as Lio, when we talk about procurement of purchases, like we think, don’t talk about laptops and pencils. We think that’s solved also by Lio agents, but that’s easy. We solved this three years ago.” The true challenge lies in capital-intensive physical builds: assembling an aircraft, standing up a robotics line, or building out the hyperscale data centers required for the AI boom. In these environments, procurement requires coordinating thousands of suppliers. A delay on a single component can stall a factory line and jeopardize hundreds of millions of dollars.
Much of this risk hides inside mundane operational exchanges. A supplier might send an email buried among 500 unread messages in a procurement manager's inbox stating that a critical part will arrive two weeks late. In an ERP system of record, that update registers merely as a date shift from this Wednesday to next Wednesday. The software cannot discern whether that date slip is benign or a catastrophic disruption. Lio’s agents are designed not merely to extract text from emails, but to evaluate operational downstream impact and initiate resolution strategies autonomously.
- **Cross-functional scope**: Procurement bridges legal terms, financial cost engineering, and commercial supply chain strategy across multiple specialized stakeholders. - **High-stakes scale**: Building aircraft or AI data centers demands synchronizing thousands of vendors where missing a single component can trigger nine-figure project delays. - **System of record blind spot**: Legacy ERPs store flat metadata (such as an updated delivery date) without interpreting operational context or calculating financial exposure.
---
`23:25` Predicting Physical World Friction and Managing Supply Chain Risk
When software interacts with the physical world, certain disruptions—such as cargo falling off a container ship or unexpected road closures—are impossible to prevent directly. However, Vlad explains that mitigating these failures comes down to probabilistic reasoning and holistic context. While an AI agent cannot alter maritime weather or road accidents, it can quantify supplier reliability and forecast risk based on historical operational data.
Vlad illustrates this trade-off with a concrete sourcing scenario: a company may evaluate two suppliers, one of which fails to deliver 20% of the time but is ten times cheaper, while the other maintains a 99% fulfillment rate at a premium price. For a non-critical component, the cheaper option might be acceptable; for a mission-critical line-stopper, paying 10x more to ensure delivery is an obvious operational necessity. To make these determinations, Lio synthesizes internal enterprise history with external context, including global news, supplier-side demand signals, and even macroeconomic betting markets like Polymarket to detect impending supply chain shocks.
Vlad sees this contextual awareness expanding beyond internal procurement workflows: “And this is what we are building at Lio, right? So it’s much bigger than just procurement into a company, it’s more like intra-company and how enterprises are doing business with each other so like buyer and supplier side.” By understanding the objectives and constraints of both buyers and suppliers, autonomous agents can negotiate and clear transactions across corporate boundaries.
---
`26:17` Orchestrating End-to-End Workflows: The Architecture of Multi-Agent Systems
Elena asks Vlad to clarify what Lio’s multi-agent architecture looks like under the hood. Vlad explains that Lio maps its agents across different levels of risk and complexity, as outlined in Seema Amble's enterprise AI taxonomy. Completing an enterprise workflow end-to-end requires multiple agents because the underlying human process naturally involves around eight stakeholders across three departments and at least five distinct software applications.
Vlad reflects on Lio's technical evolution over the past three years. The platform initially began as a retrieval-focused copilot before shifting toward single-agent automation. The team quickly realized that a lone agent cannot reliably execute complex tasks like commercial negotiation without supporting agents monitoring legal contracts, assessing operational inventory, and checking external news feeds. Multi-agent coordination became essential for true end-to-end execution.
Vlad outlines the core design requirement for full workflow autonomy: “When we talk about like a multi-agent system, this is essentially what we are doing: we are solving like the task end-to-end and because the human-level task involves eight people, eight stakeholders, and maybe like three departments and five different software tools, we need to cover like all of those to do like the job end-to-end and those agents need to then communicate with each other.”
- **Historical progression**: Lio evolved from an initial retrieval copilot to single-task agents, and ultimately to specialized multi-agent systems. - **Enterprise integration footprint**: Tasks span an average of eight cross-functional stakeholders, three departments (procurement, engineering, legal/finance), and five software platforms. - **Inter-agent communication**: Individual agents pass structured state sequentially and conditionally, allowing contract agents, news monitors, and price benchmarkers to inform negotiation strategies.
---
`28:01` The Life of a Bolt: Anatomy of an Autonomous Procurement Cycle
To illustrate the concrete mechanics of Lio's multi-agent system, Elena asks Vlad to walk through the end-to-end procurement of a physical part—such as an airline or aircraft manufacturer like Boeing needing to procure a specialized bolt. Vlad explains that the journey begins with demand capture, a phase notorious for user friction. A construction worker or shop-floor engineer who only logs into a computer every other week usually despises navigating complex ERP tools like SAP. They do not know or care about general ledger (GL) accounts, spend categories, or framework master agreements.
Lio removes this barrier by allowing the user to submit demand through multimodal inputs—snapping a photo of the part, uploading a supplier quote, or dropping in an unformatted spreadsheet. Lio's intake agent automatically parses the demand, infers the relevant accounting codes, and queries internal inventory across neighboring manufacturing plants to see if the bolt can be transferred internally. If internal stock is unavailable, the sourcing agent queries both approved and external suppliers, drafts a request for quotation (RFQ), and dispatches it over email.
When supplier responses arrive, Lio ingests an unstructured mix of raw email bodies, attached PDFs, and messy spreadsheets. The retrieval agent normalizes the bid data and benchmarks pricing against historical and market baselines. If a savings opportunity exists, the orchestrator triggers a negotiation agent. Depending on the dollar value and strategic risk, the system can execute an autonomous negotiation, launch an electronic auction (e-auction), or initiate a human-in-the-loop workflow where the agent drafts negotiation parameters for human review.
Vlad summarizes the fundamental disconnect between front-line requisitioners and legacy ERP software: “No one cares outside of the procurement department about categories, GL accounts, framework contracts, no one cares. We in the procurement world care, but no one outside that cares.”
```mermaid flowchart TD accTitle: End-to-End Autonomous Procurement Workflow accDescr: Diagram illustrating Lio multi-agent workflow from demand intake to automated negotiation.
subgraph Intake ["Intake & Internal Check"] A["User Input: Photo / Quote / Excel"] --> B["Intake Agent (Assigns GL & Category)"] B --> C{"Check Internal Inventory"} end
subgraph Sourcing ["External Sourcing & RFQ"] C -->|"Not Available Internally"| D["Sourcing Agent (Selects Vendors)"] D --> E["Draft & Dispatch RFQ via Email"] E --> F["Ingest Unstructured Quotes (PDF / Excel / Email)"] end
subgraph Analysis ["Benchmarking & Resolution"] F --> G["Price Benchmarking & Impact Analysis"] G --> H{"Negotiation Strategy"} H -->|"Autonomous"| I["Autonomous Negotiation / E-Auction"] H -->|"Strategic"| J["Human-in-the-Loop Negotiation"] end ```
- **Intake without friction**: Requisitioners capture demand via photo uploads, vendor quotes, or spreadsheets, bypassing manual entry into SAP or other legacy ERP systems. - **Automated accounting categorization**: The agent infers GL accounts, spend categories, and contract frameworks without requiring requisition training from front-line staff. - **Internal vs. external routing**: The system checks regional plant inventory before engaging external vendors to avoid unnecessary spend. - **Unstructured parsing**: Responses spanning diverse formats (freeform email text, attachments, PDF spec sheets, Excel matrices) are ingested and structured automatically. - **Adaptive negotiation paths**: Depending on value, category risk, and benchmarked savings opportunities, Lio dynamically routes between autonomous negotiations, competitive e-auctions, or human-in-the-loop approvals.
`30:00` Indirect vs. Direct Procurement: Where Agents Deliver P&L Impact
Vlad explains that procurement workflows can run end-to-end autonomously—encompassing everything from order confirmation to shipment tracking and invoice processing. When thinking about where agents belong, Lio categorizes enterprise tasks across two axes: what teams already do thousands of times a day that they wish they could do zero times, and what they currently do zero times a day that would yield a massive P&L impact if done thousands of times. The latter involves autonomous negotiations on spend categories companies have historically ignored due to a lack of headcount.
```mermaid flowchart TD accTitle: Indirect vs Direct Procurement Workflows accDescr: Comparison of indirect procurement automation versus direct procurement deep analysis. subgraph Indirect["Indirect Procurement (High Volume, Broad Tail)"] direction TB I1["High volume: MRO parts, laptops, pencils, studio builds"] I2["50,000+ suppliers across tail spend"] I3["Goal: Autonomous negotiations on spend never negotiated before"] I1 --> I2 --> I3 end
subgraph Direct["Direct Procurement (Mission-Critical Strategic Spend)"] direction TB D1["Core inputs: Airplane parts, drones, robotics"] D2["100 to 2,000 strategically vital suppliers"] D3["Months-long negotiations, $1B+ spend per supplier"] D1 --> D2 --> D3 end ```
Vlad contrasts indirect procurement (MRO parts, factory buildouts, bolts, laptops, pencils, podcast studio construction) with direct procurement: - Indirect procurement deals with long-tail suppliers (often upwards of 50,000 vendors) where automated micro-negotiations uncover margin that was previously left on the table. - Direct procurement involves core materials directly needed to build a physical product, like an airplane, drone, or robot. Here, enterprises deal with only 100 to 2,000 vendors, but spend on a single supplier can easily exceed $1 billion. - Because the stakes are so high in direct procurement, an autonomous agent does not simply take over the negotiation. Instead, it acts as deep preparation support across months-long negotiations involving cost engineers who analyze aluminum indices, oil prices, and CAD drawings.
Vlad points out: “90% of the work is preparation. The end result that you see in your system of record is like, oh, instead of like one billion, I paid 900 million dollars.”
---
`32:16` Real-Time Agent Copilots in High-Stakes Negotiations
Elena asks whether direct procurement means agents operate behind the scenes or sit directly across the table like an automated negotiator shaking hands. Vlad clarifies that it is primarily back-of-house, doing the heavy analytical lifting that ten people would normally spend three months preparing. However, real-time in-negotiation assistance represents an emerging frontier for direct procurement:
> “Imagine while we are negotiating, I would have real-time insights on my screen popping up where you tell me the indice for oil changed 10%. So it's increased by 10%. So that's why we need to increase the prices by 10%. And I would have an immediate pop-up of like, that's true, oil increased by 10%, but the product has only 30% of oil contained. So your overall price should only increase maybe only by 4%.”
By pulling live index data, contract clauses, and raw material bill-of-materials breakdowns during the conversation itself, human negotiators can instantly counter supplier cost-increase claims with hard math rather than conceding ground or pausing the meeting to run the numbers.
---
`34:04` Foundation Models, Custom Harnesses, and Outcome-Based Training
When asked how Lio handles fine-tuning and model selection relative to companies like Harvey or Decagon, Vlad notes that Lio treats underlying foundation models across various providers largely as commodities for generic enterprise tasks—reading documents, generating PDFs, and drafting spreadsheets. In many standard workflows, combining an off-the-shelf foundation model with a robust harness can solve 100% of the job to be done.
```mermaid flowchart LR accTitle: Model Architecture for Procurement accDescr: Pipeline showing commodity models versus outcome-based fine-tuning. subgraph Commodity["Standard Tasks (100% via Harness)"] F1["Foundation LLMs (Any Provider)"] --> H1["Custom Harness"] H1 --> O1["PDF & Excel Processing, Admin Tasks"] end
subgraph Specialized["Complex Engineering & Benchmarking (Needs Fine-Tuning)"] F2["Proprietary Enterprise Data"] --> M1["Outcome-Based Training"] M1 --> O2["Should-Cost Modeling (CAD Drawings)"] M1 --> O3["Price Benchmarking (BCG vs. McKinsey Quotes)"] end ```
For specialized industrial workflows, however, the standard "foundation model plus harness" paradigm caps out at around 80% accuracy. Bridging the final 20% requires fine-tuning on proprietary data: - **Should-Cost Modeling:** Cost engineers examine technical drawings to calculate the bottom-up target price of an engineered part based on machine time, materials, and labor. - **Price Benchmarking:** Benchmarking proprietary B2B service quotes that are not publicly available. Vlad compares this to assessing two identical-looking consulting proposals from BCG and McKinsey that differ by 10x in price, or pricing out professional video content creation: an experienced procurement lead uses gut instinct and implicit taste rather than codifiable rules.
To capture that qualitative intuition, Lio looks beyond standard conversational text completion toward outcome-based training that maps input proposals directly to an accurate target price.
---
`36:34` Moats, Dependency, and the Durability of Vertical AI
Elena turns to Seema to discuss how durable vertical AI companies build defensibility as AI research labs partner directly with incumbents. Seema notes that trying to plot out a rigid multi-step roadmap to a moat from day one is a trap:
> “If you look back at all the best businesses, at the early stages, they were just thinking about, okay, I'm winning customer trust, I'm selling more to them, and there's a lot of opportunity versus, okay, I'm going to do these six steps and then get to the seventh step, and then we'll have them vote.”
Seema argues that true defensibility stems from creating structural operational dependencies: - Traditional software (like an old CRM) served as a passive database or log of closed deals. - Vertical AI agents actively execute work: owning deal preparation, running outbound, triaging inbound, and completing analysis. - Once a business hands over end-to-end execution of a core function to an agent, replacing that product becomes far more disruptive than swapping out a static system of record. - Moat characteristics like high switching costs, stickiness, and proprietary data flywheels are natural byproducts of owning the workflow, not things that can be designed in isolation before achieving product-market fit.
`40:00` The DIY Trap: Why 70% Performance Does Not Equal 70% Automation
Three years ago, Lio’s very first product tackled a straightforward yet high-impact workflow: retrieving unstructured data from vendor quotes and entering it cleanly into SAP. Originally built and sold by a team of just three or four engineers, that exact workflow has now become a standard eight-hour applicant case study for prospective engineers applying to work at Lio. This dramatic shift highlights how trivially easy it has become to construct basic AI prototypes, which inevitably prompts enterprise procurement departments to wonder why they should not simply build their own internal solutions in a matter of months. However, Vladimir Keil explains that while an internal team can assemble a prototype in an afternoon, hitting roughly 70% accuracy, that baseline metric falls critically short in real-world operations. As Vlad points out: “And the problem is 70% of the performance or accuracy or however you measure it, it depends really on the task, doesn’t mean 70% automation, right?” When a system operates at only 70% reliability, employees still have to review and verify 100% of the output, creating more busywork than existed before.
Seema Amble notes that while surging enterprise familiarity with foundational models and ChatGPT is a net positive that builds trust for vertical AI applications, do-it-yourself software initiatives almost always stall out. Recounting a conversation with the management team of a Fortune 500 company two or three weeks prior, Seema describes their failed attempt to construct an in-house cash collection application. The internal team struggled with poor-quality organizational context, messy screen grabs, and inconsistent meeting recordings. The project completely unraveled when confronted with real-world infrastructure complexity: the company was running two distinct ERP instances and preparing to acquire a third, leaving them paralyzed over who would maintain mappings, manage ongoing system integrations, and update complex edge-case rules. Enterprise leaders eventually remember the historic lesson of IT: custom internal tooling distracts from a company's core business competency.
Vlad reinforces this dynamic using the Pareto principle, explaining that bridge from an 80% prototype to a production-ready system absorbs 80% of the total engineering effort. That final stretch demands deep vertical domain data, persistent agent memory, robust workflow engines, and reliable enterprise integrations. Unless a Fortune 150 enterprise intends to turn managing an autonomous procurement workforce into its primary strategic competency, attempting to build and maintain that infrastructure internally is an inefficient use of resources.
- **The Hiring Benchmark**: Lio’s initial commercial product—a retrieval agent that ingests vendor quotes directly into SAP—is now given to engineering candidates as an 8-hour take-home challenge. - **The Automation Fallacy**: Achieving 70% model accuracy does not equal a 70% reduction in workload; human staff must still audit 100% of incoming records to catch errors, eliminating the efficiency gains. - **The Maintenance Overhead**: Internal projects stumble when mapping messy operational exceptions across multiple legacy ERP systems, especially during corporate mergers and acquisitions.
```mermaid flowchart TD accTitle: Internal AI Build vs Production Reality accDescr: Diagram contrasting an internal prototype hitting an accuracy plateau with the full requirements needed for production deployment.
subgraph Internal["Internal Build Attempt"] A["8-Hour Prototype"] --> B["70% - 80% Accuracy"] B --> C["100% Manual Human Verification"] C --> D["Net Increase in Operational Burden"] end
subgraph Production["Production Enterprise Agent Harness"] E["The Final 20% Hurdle"] --> F["Custom ERP Integrations"] E --> G["Persistent Agent Memory"] E --> H["Edge-Case Exception Handling"] E --> I["Multi-System Synchronization"] end ```
---
`44:30` Bilateral Agents: When Both Buyers and Suppliers Automate
Looking ahead to market maturity, Elena Burger asks how dynamics shift when suppliers also deploy AI agents, and whether procurement will soon feature software agents negotiating directly against each other. Vlad confirms that bilateral agent interactions are inevitable, but notes an unexpected dynamic in today's industrial sectors. While software sales teams are traditionally seen as tech-forward compared to "unsexy" procurement teams, suppliers servicing major Fortune 5000 industrial accounts are surprisingly behind on real automation. They frequently use transcription apps like Granola or record sales calls, but lack deployed autonomous agents. Crucially, industrial enterprise buyers—such as automotive manufacturers—command massive leverage over their suppliers. They dictate product quality tolerances, response formats, and RFQ protocols, creating a clear opening for buyers to mandate that suppliers interface directly through Lio agents, allowing Lio to coordinate both sides of the transaction.
```mermaid flowchart LR accTitle: Bilateral Agent Coordination Model accDescr: Flowchart illustrating how buyer and supplier agents share unified incentives across upstream tasks despite adversarial pricing negotiations.
subgraph Upstream["Aligned Upstream Workflows (5,000 Shared Tasks)"] direction TB T1["RFQ Specification & Intake"] T2["Engineering & Compliance Reviews"] T3["Data Center & Aircraft Timeline Tracking"] T4["Frictionless Documentation Hand-offs"] end
subgraph Parties["Bilateral Parties"] B["Enterprise Buyer Agent"] S["Industrial Supplier Agent"] end
subgraph Downstream["Adversarial Zero-Sum Stage"] P["Final Contract Price Negotiation"] end
B --> Upstream S --> Upstream Upstream --> Downstream ```
Elena and Seema draw a parallel to the legal sector, where opposing law firms representing adversarial clients still benefit tremendously from shared platforms that track current draft versions, log open issues, and verify agreed-upon terms without tedious human back-and-forth. Vlad emphasizes that while final price negotiation is inherently a zero-sum contest, the actual pricing agreement is merely the end result of thousands of cooperative upstream operational tasks. In massive manufacturing projects—such as constructing aircraft or building hyperscale data centers—buyers and sellers have perfectly aligned incentives: buyers want fast time-to-market, and suppliers want frictionless sales cycles without spending six months parsing complex RFQ responses. As Vlad summarizes: “Sales wants to have as little friction as possible. Buyers want to have a really fast time to market.” Deploying agents on both sides to automate those preliminary tasks benefits both parties, regardless of who negotiates the final number.
- **Procurement Leverage**: Industrial enterprises hold significant sway over vendor behavior, allowing them to dictate the software interfaces, compliance standards, and submission protocols that suppliers must use. - **Shared Incentives Over Zero-Sum Friction**: Although final price margins represent conflicting financial interests, the preceding 5,000 steps—including scope definitions, spec analysis, and delivery schedules—share an identical goal: maximizing execution speed and removing friction. - **The Legal Precedent**: Adversarial entities in high-stakes fields already rely on shared coordination software to catalog agreed terms and outstanding deal points without requiring constant administrative friction.
---
`48:30` Forward-Deployed AI and Customization at Scale
Elena raises a traditional rule of enterprise B2B software: vendors should avoid heavy product customization for individual enterprise clients to preserve software margins and avoid becoming professional services firms. She questions whether large language models and autonomous agents fundamentally break that rule by enabling hyper-customization at software-like speeds. Seema agrees that the paradigm has shifted, observing that early-stage enterprise AI companies rely heavily on forward-deployed engineering teams to bridge current product gaps. This high-touch deployment model exists because legacy enterprise data environments are notoriously disorganized. As Seema explains: “And part of that is because the state of the customer data and understanding customer N is a lot harder than understanding N plus one.”
This forward-deployed dynamic creates an asymmetric advantage for nimble AI startups over entrenched legacy software incumbents. Traditional enterprise software giants rely on rigid implementation teams whose client work is treated as an afterthought rather than a real-time feedback channel for the engineering roadmap. In contrast, modern AI startups use forward-deployed engineers to build robust automated evaluation frameworks and capture critical user edge cases directly from production environments. By feeding operational nuances straight back into the agent harness, the model continuously learns, refining agent performance across each successive enterprise rollout.
- **Forward-Deployed Advantage**: Early enterprise AI deployments require hands-on technical work to untangle idiosyncratic customer data environments, but each deployment systematizes learnings that make onboarding subsequent customers faster. - **Structural Incumbent Weakness**: Legacy enterprise vendors isolate their services and implementation teams from core engineering, while AI-native startups treat implementation as an active product feedback and evaluation engine. - **Learning Loops**: Vertical AI systems leverage automated evals and real-time operational feedback to refine agent performance and custom system behavior over time.
`50:00` Automating the Forward-Deployed Engineer
```mermaid flowchart TD accTitle: Forward-Deployed Engineering to Self-Service Evolution accDescr: Three evolutionary stages showing how enterprise software implementation transitions from manual consulting to self-service AI tooling.
subgraph Legacy["Legacy Model"] L1["Accenture / Systems Integrator"] -->|"Manual Customization"| L2["Core ERP / SAP Deployment"] end
subgraph Hybrid["FDE Transition Model"] F1["Lio Forward Deployed Engineers"] -->|"Spec Nuances & Build Integrations"| F2["Multi-Agent Workflow"] F1 -->|"Automate Own Role / FDE KPI"| F3["Turn Work into Code"] end
subgraph SoftwareDriven["End State: Pure Software"] S1["Automated Deployment & Self-Service"] -->|"Enterprise Knobs & Levers"| S2["Scalable Enterprise Rollout"] end
Legacy -.-> Hybrid Hybrid -.-> SoftwareDriven ```
Seema Amble and Vladimir Keil explore how enterprise deployment is shifting away from heavy professional services. In the legacy era, deploying and tailoring systems like SAP required bringing in an army of consultants from firms like Accenture. Modern software startups replaced that with forward-deployed engineering (FDE) teams, but the next frontier is automating the deployment itself. Seema notes that portfolio companies are moving quickly toward software where customization is handled automatically, enabling customers to adjust their own operational parameters via software levers rather than relying on humans.
Vladimir confirms that this philosophy dictates Lio's organizational design: 85% of their personnel are engineers, and even those without official engineering titles typically hold engineering backgrounds. The explicit objective is to prevent Lio from morphing into an IT consultancy while tackling intricate enterprise nuances across direct procurement, indirect procurement, and finance. To scale without expanding headcount linearly, Lio gives its FDEs an unusual performance objective: automate their own day-to-day workflow. As Vladimir explains, their task is to convert repetitive manual onboarding and configuration into self-service product features, freeing them to move on to the next set of challenges.
- 85% of Lio's staff are engineers, and non-engineering roles predominantly require engineering backgrounds to avoid becoming a consulting practice. - Forward-deployed engineers at Lio are explicitly evaluated on automating their own deployment responsibilities into self-service tooling. - Vladimir Keil on engineering incentives: “Their KPIs, you're seeing this happening multiple times. Literally, your job is to automate yourself. And then if you automate yourself, you go to the next task.”
`51:30` Bots and Buyers: The Reality of Procurement
Elena Burger shifts the conversation to the "Bots and Buyers Summit," an event hosted by Lio a few blocks from their New York office, contrasting it humorously with Salesforce’s Dreamforce. Vladimir notes that this was their third summit, gathering over 100 enterprise C-level leaders, chief procurement officers (CPOs), and vice presidents. Despite the appearance of more than 500 specialized tools across landscape graphics on LinkedIn over the past 26 to 27 years, procurement remains universally dreaded. Requesters despise navigating it, suppliers find it painful, and procurement specialists themselves dislike working within their own departmental constraints.
Vladimir diagnoses the root problem: existing software vendors merely slapped digitizing layers over legacy processes without changing how people actually spend their days. Regardless of how many point solutions a company licenses, employees continue to coordinate through unstructured channels—endless email threads, Microsoft Teams messages, static PowerPoint decks, and sprawling Excel spreadsheets—all running without AI assistance. Rather than unveiling incremental standalone tools like an isolated invoice feature, Lio designs its workflows cross-departmentally, tracking an item from the instant a team member needs it until it arrives on site across direct materials, indirect goods, logistics, and corporate finance. To bridge the gap between abstract agent architectures and practical operations, Lio erected physical booths at their New York office where executives could test and inspect the agents hands-on, leading to a planned 700-person gathering in Munich.
- Lio’s New York summit convened over 100 enterprise procurement executives (C-level, CPOs, and VPs), with an upcoming summit targeting around 700 attendees in Munich. - Vladimir Keil on procurement tooling's failure: “You can find them on LinkedIn and you will see like they're like 500 procurement tools... But they never changed how those people actually work. And it's like crazy to see that they work in emails and on Microsoft Teams and in Excel sheets and PowerPoints.” - Vladimir Keil on universal frustration: “I challenge someone like to find someone who like loves to work with procurement like no one does. Like you can really state people hate working with procurement.”
`55:00` The Trillion-Dollar Nexus of Emotion, Boredom, and P&L Impact
Seema Amble highlights that procurement has historically had notoriously low Net Promoter Scores, recalling customer discovery calls where frustrated executives stated they were locked into systems like Coupa and refused to even discuss software evaluations. Vladimir explains that this deep-seated friction is precisely what makes procurement an extraordinary startup opportunity. While perceived from the outside as boring back-office B2B SaaS, the emotional charge surrounding broken workflows gives it intense urgency. The category has suffered a total lack of fundamental innovation for twenty years, with the only notable updates being occasional user interface refreshes roughly a decade ago. Consequently, delivering tangible, agentic automation quickly amazes buyers who are unaccustomed to real technical progress.
Beyond emotion and neglect, Vladimir emphasizes procurement's disproportionate leverage on enterprise balance sheets and the broader physical economy. Procurement dictates the execution, cost, and speed of building high-stakes infrastructure, including data centers, commercial aircraft, automotive manufacturing, and drones. Financially, operating leverage dictates that gaining a 1% margin expansion through top-line growth demands roughly a 10% increase in gross revenue; by contrast, capturing a 1% reduction in procurement spend flows directly to the bottom line, delivering that same bottom-line outcome instantly. This combination—unmatched operational frustration, a neglected software category, and extreme financial leverage—creates what Vladimir characterizes as a massive market opportunity before Elena closes the episode.
- A 1% cost savings in procurement yields the equivalent bottom-line P&L impact of driving a 10% increase in corporate sales revenue. - Procurement dictates real-world industrial infrastructure, directly governing data center rollouts, aircraft assembly, vehicle production, and drone manufacturing. - Vladimir Keil on the enterprise leverage of procurement: “If you just manage to get 1% savings, it's equals 10% of sales that you have to do to get the same outcome in your P&L.” - Vladimir Keil on why the category is poised for massive venture outcomes: “And you combine all of those three things a trillion dollar business opportunity. That's my opinion.”
## Notable quotes > `02:30` The legacy incumbent is limited to their system of record and that record that they have, and they're not completing the end-to-end job. — Seema Amble
> `00:00` No company and no enterprise starts with fully autonomous negotiation agents from day one. — Vladimir Keil
## Mentioned Elena Burger, Seema Amble, Vladimir Keil, Lio, Salesforce, Claude, Agentforce, SAP, Oracle, ChatGPT, Andreessen Horowitz, Polymarket, Boeing, Harvey, Decagon, BCG, McKinsey, Granola, Accenture, Coupa, Salesforce Dreamforce, Bots and Buyers Summit
## Video description a16z’s Seema Amble and Elena Burger sit down with Lio co-founder and CEO Vladimir Keil to ask where AI-native startups have an advantage when incumbent software companies already own the customer, the data, and the system of record. Their answer comes down to the work that happens outside those systems. In procurement, a final price in an ERP can hide hundreds of emails, spreadsheets, supplier conversations, engineering analyses, and decisions across legal, finance, and operations. Vlad explains how Lio uses multi-agent systems to take on more of that end-to-end work, from sourcing and RFQs to negotiation, shipment tracking, and invoices. They also discuss how enterprises learn to trust agents with increasingly consequential decisions, why the last 20% of an internal AI build can require most of the effort, and what happens when both buyers and suppliers have agents working on their behalf.
[Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/63755eda-ff93-4095-8822-e8bcb95a06a6/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=63755eda-ff93-4095-8822-e8bcb95a06a6&feed=JGE3yC0V)