a16zWhy Building an AI Agent Is Easier Than Deploying One
a16z
The Procurement Challenge 0:00
Building something like an aircraft requires coordinating thousands of suppliers, and a missed email about a delayed part can cause hundreds of millions in damage. Procurement used to be a contained function, but now it touches legal, finance, and many different software systems and people. The opportunity for an AI native startup is to own that entire end-to-end process rather than just a piece of it. Vlad Kyle, co-founder and CEO of LEO, which builds AI agents for enterprise procurement, joins Sema Amble, a partner at A16Z, to discuss this. No company starts with fully autonomous negotiation agents from day one, because customers do not yet trust the technology, so LEO uses a human-in-the-loop approach, feeding the agent feedback until it earns trust across tens of thousands and eventually hundreds of thousands of negotiations.
Why Startups Beat Incumbents 2:00
Sema explains that incumbents can now bolt a capable model onto their existing software, which raises the question of why a startup is still needed. Her answer is that incumbents are limited to their own system of record and cannot complete the full end-to-end job. She gives the example of a customer who was charged after cancelling a service: resolving that issue means touching billing, chat history, and the contract, none of which lives in a single system of record. Vlad adds that in procurement the official record looks deceptively simple, something like an 8K price for aluminum, while the real work behind it includes a supplier pushing back for 10K, a cost engineer spending three weeks on spreadsheets and 3D modeling, and similar hidden effort on the sales side with dozens of stakeholder meetings and hundreds of emails. Most of the actual procurement work happens outside the ERP system entirely.
Four Kinds of Agents 6:01
Sema lays out four categories from her piece, using a customer service outage as the example. A retrieval agent simply pulls up stored information, like confirming a contract term or an outage date. A process agent goes further, applying a policy handbook to approve a credit without needing judgment. A policy agent exercises judgment, deciding whether a 20 minute outage counts as significant. A principal agent weighs relationship value and decides whether to offer extra compensation beyond policy. Most incumbents sit mainly in the retrieval stage, marketing movement toward process and policy but rarely reaching real judgment, partly because of internal incentive conflicts between teams selling workflow tools versus teams trying to resolve work end to end. Vlad notes LEO spans all four categories depending on risk and complexity, and that trust is both an internal matter for incumbents and an external one startups must earn, since customers must trust an AI agent enough to let it negotiate directly with a third party. He describes how LEO started three years ago with simple document retrieval, then found that 80 percent of real procurement problems lie in exception handling, like fraudulent invoices or mismatches, which pushed the company to build process and eventually autonomous agents ahead of market expectations.
Earning enterprise trust early 15:30
You win enterprise trust by moving ahead of the hype cycle, having started building before ChatGPT's breakthrough and already hearing the real problems companies faced. Once the tech bubble caught up, the team could pitch quickly, show a working chatbot or document processing use case, and ship it into production fast, which built the credibility needed for bigger deployments.
Human in the loop by design 17:31
No enterprise hands over fully autonomous negotiation on day one, so the product leans on human oversight as a trust-building step. Smaller negotiations, like a disputed 40k invoice that companies previously ignored because they lacked the capacity to fight it, can run autonomously since the downside of a bad negotiation is nearly zero. Larger or relationship-sensitive deals, such as multi-million dollar contracts involving 3D models and technical drawings, or vendor relationships like a trusted podcast studio, keep experts such as cost engineers, legal, and procurement staff actively feeding feedback into the agent over many hours of work.
Procurement across whole industries 20:30
Building something like an aircraft or a data center means coordinating thousands of suppliers, and a single part arriving two weeks late can cause hundreds of millions of dollars in damage, often because a missed email buried among hundreds in an inbox was the only warning sign. The agents work by predicting supplier reliability and shipment risk using outside context such as news and even prediction markets, sometimes favoring a cheaper, less reliable supplier or a pricier, safer one depending on the situation. Leo operates as a multi-agent system because real procurement tasks touch multiple departments, stakeholders, and software tools, requiring separate agents for sourcing, drafting requests for quotes, negotiation, and tracking shipments and invoices to all communicate and hand off work in sequence, as in the walkthrough of a company needing to procure a simple bolt from demand through delivery.
Indirect versus direct spend negotiations 31:01
The discussion separates two kinds of procurement work. Indirect spend covers things like MRO parts, laptops, pencils, and marketing services, where a business might deal with 50,000 suppliers and autonomous negotiation makes sense because volume is high and stakes per deal are low. Direct spend is different: building an airplane, drone, or robot involves only 100 to 2,000 suppliers, each strategically critical, sometimes representing a billion dollars of spend. For those, you do not want a fully autonomous negotiation; instead the process takes three months, with engineers analyzing aluminum and oil price indices, tracking price changes, reviewing drawings, and checking part quality.
Agents working behind the scenes 33:01
Agents mostly operate back of house rather than sitting across the table shaking hands. In a complex multi-million dollar negotiation, ninety percent of the work is preparation, often ten people working full time for three months before a final number like nine hundred million dollars appears in the system of record. Some uses are more real time, though: during a live negotiation an agent could flag that oil prices rose ten percent, then immediately correct that since the product only contains thirty percent oil, meaning the price should rise closer to four percent rather than ten.
Models, harnesses, and durable moats 34:30
The approach treats foundation models as a commodity used for general tasks like reading documents or building spreadsheets, while a harness, meaning the surrounding workflow and integration layer, pushes performance further. Some tasks, like cost engineers' should cost modeling or price benchmarking, rely on proprietary data no public model has seen, so the plan is to fine-tune outcome based models rather than classic language models. On durability, the view is that moats come from owning the end to end work, building a data asset, and locking in dependency, much as a new sales AI agent now owns prep and outbound work that old CRM tools never touched. On build versus buy, an internal engineering team could replicate an early retrieval agent in eight hours but would only hit seventy percent performance, meaning workers still redo all the checking, which can create more work than before. A Fortune 500 company's own cash collection build failed after three or four months due to poor context and mismatched ERPs, reinforcing that the last twenty percent of performance, built through integrations, memory, and vertical data, is what actually gets a product into production. The conversation closes by noting suppliers lag behind buyers in adopting agents, even though procurement departments often have the power to dictate tools to suppliers, raising the question of whether both sides of a transaction could eventually run on the same agent platform.
Coordination Could Become an Agent Task 47:01
Right now, tracking the latest draft of an agreement, the open issues, and what has already been settled is work done entirely by humans. That coordination role, keeping both sides aligned on where a negotiation stands, is something an agent could take over, and over time both buyer and seller sides may end up on the same platform, better coordinated for everyone rather than racing against each other.
Shared Incentives Beyond Price 47:30
Even when buyer and seller negotiation agents sit on opposite sides of a price discussion, price itself is just the outcome of roughly five thousand smaller tasks where both sides actually share the same incentive. Sales wants minimal friction, buyers want speed, especially in building things like data centers, aircraft, or cars where delay is costly. Because those underlying incentives match, agents can be deployed on both sides to automate that shared work. On customization, the discussion notes that forward deployed engineers at Leo are measured by how well they automate their own jobs, turning bespoke enterprise work into self-service product features, an approach compared to how Google operates.
Procurement's Bots and Buyers Summit 52:32
Leo's third Bots and Buyers Summit in New York drew over a hundred senior procurement leaders. The event highlighted that procurement has had around a thousand point tools built over the last twenty-five years, yet none changed how people actually work, since most still rely on email, Microsoft Teams, Excel, and PowerPoint. Leo instead shows a cross-department view, letting visitors walk through physical booths experiencing agents handling indirect, direct, logistics, and finance work together. The next event, in Munich, expects around seven hundred attendees. Procurement is described as emotional, boring and niche, yet carrying enormous business impact, since a one percent cost saving can equal the P&L benefit of ten percent more sales, making it a trillion-dollar opportunity.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
