Why Building an AI Agent Is Easier Than Deploying One
Vladimir Keil, co-founder and CEO of Lio, joins Seema Amble and Elena Burger to explain why AI-native startups can beat incumbent software vendors even when those vendors own the customer, the data, and the system of record. The discussion covers the work that happens outside systems of record in enterprise procurement, Lio's multi-agent architecture spanning retrieval through principal agents, how buyers learn to trust agents with consequential negotiations, and what happens when suppliers and buyers both run agents.
Where AI-native startups win when incumbents own the system of record
The core argument is that incumbents own the customer, the data, and the system of record, but the system of record only captures the final state of a transaction. A customer’s cancellation dispute, for example, touches billing, chat history, and contract terms, and no single system holds all of that. The opening for AI-native companies is to own the entire end-to-end arc of a job rather than one step inside one product.
The hidden work behind an $8K line item
In procurement, the ERP shows a clean outcome: an $8K aluminum purchase. Behind it sit roughly 30 stakeholder meetings, 500 emails, 20 spreadsheets, a supplier pushback from $8K to $10K, and three weeks of cost engineering with Excel models and 3D modeling. Most procurement work happens outside the system of record, which is precisely the work that incumbents’ data models never see.
Four layers of agents: retrieval, process, policy, principal
Seema Amble’s framework separates agents by the kind of judgment they perform. Retrieval agents answer questions from the system of record, the level of most incumbent chatbots launched in the first wave. Process agents execute defined rules, such as applying a credit when a documented outage meets policy thresholds. Policy agents exercise judgment where rules are fuzzy, deciding whether a 20-minute outage counts as significant. Principal agents weigh the relationship and strategic trade-offs, deciding whether a bad outage warrants compensation beyond policy.
Most incumbents remain in the retrieval tier with some process automation. Their workflow agents help get a document signed or transcribe information, but they have not moved into judgment. Part of the reason is distribution: incumbents can turn on a new agent for customers with almost no friction, as Salesforce did with Agentforce. The other part is internal conflict. Workflow products serve a human support agent, while end-to-end resolution serves a different buyer, and the two product teams and their VPs often have misaligned incentives and no clear owner for the sale.
How Lio operates across the spectrum
Lio spans all four categories, with autonomy calibrated to complexity, risk, and budget. Some workflows run fully autonomously, while others keep a human in the loop. Trust is both internal and external. Incumbents already hold customer trust, but they can lose it by shipping prematurely, and a startup has to earn that trust in the first place. Lio’s invoice agents illustrate the point. Invoice processing is commoditized and represents roughly 20% of the problem; the remaining 80% is exceptions such as fraud, mismatches, and off-path cases. Lio moved quickly into exception handling and pitched each next generation of capability to customers before shipping it.
Earning trust through human-in-the-loop negotiation
No enterprise hands a new vendor fully autonomous negotiation on day one. Lio starts with a human in the loop, feeding the agent feedback about how a specific enterprise operates. That accumulated learning is what moves customers from trusting the system with a few thousand dollars of negotiation to 10K, 20K, and eventually 100K-scale deals. For multi-million-dollar negotiations involving 3D models and technical drawings, experts remain in the loop, and long-running agents pause for cost engineers to give feedback before continuing.
The distinction between indirect and direct procurement shapes the design. Indirect spend, such as MRO parts, laptops, and marketing services, is high-volume and often never negotiated because teams lack capacity, so autonomous agents can capture savings that were previously ignored. Direct spend, the strategic suppliers behind an aircraft or a drone, involves perhaps 100 to 2,000 suppliers, each with large spend, and negotiations that run for three months with heavy preparation. Relationship-sensitive categories such as podcast studio or marketing services need a human voice even when the spend is modest. Negotiation itself is only part of the work: contract design touches legal, cost structure touches finance, and should-cost modeling touches cost engineering.
Why physical-world procurement is a coordination problem
Building an aircraft, a data center, or a factory means procuring from thousands of suppliers, and small frictions can cascade. A single part arriving two weeks late can cause hundreds of millions of dollars in damage. The critical information often lives in one of 500 emails in a procurement manager’s inbox, while the system of record shows only a date. Agents that retrieve that context and assess impact can make decisions about it rather than simply surfacing it.
Risk management is probabilistic rather than preventive. An agent cannot stop a ship from losing cargo, but with enough context about supplier reliability, it can choose a supplier whose higher price reflects a lower chance of missing goods. Combining enterprise context with external signals such as news and prediction markets extends that context further. The emphasis is on intra-company and inter-company coordination, buyer and supplier side together, rather than procurement as an isolated function.
A bolt, end to end
Take a single bolt for an aircraft manufacturer. Lio’s multi-agent system begins with a demand input that avoids SAP entirely: a user uploads a photo, quote, or spreadsheet. Agents then check inventory, ask other plants whether the part can be sourced internally, and query internal and external suppliers. A drafting agent sends the RFQ, and responses arrive in every format, from emails and PDFs to spreadsheets, which must be parsed. Price benchmarking then identifies negotiation opportunities, and the system routes to strategic negotiation with a human, autonomous negotiation, or auctions. The process concludes with order confirmation, shipment tracking, and invoicing.
A multi-agent architecture was necessary because the work involves around eight people, three departments, and five software tools. Lio’s progression ran from a copilot to a single agent to a coordinated system, since negotiation could not be solved without a contract agent or a news-monitoring agent feeding context into it. Real-time support is also in development: during a live negotiation, an agent could flag that oil prices rose 10% but that the product contains only 30% oil, suggesting a 4% increase rather than 10%.
Models, fine-tuning, and where outcomes matter
Lio uses multiple foundation models from several providers and treats them as commodities for general business tasks such as reading PDFs and producing spreadsheets. Some tasks can reach full task completion by combining a strong model with a good harness, but others stall at around 80%. Should-cost modeling and price benchmarking are examples where proprietary data and tacit expertise matter. Two identical consulting quotes can differ tenfold in price, and a procurement manager’s gut feeling about which is reasonable reflects judgment that is hard to write down as rules. That motivates outcome-based training, where a model learns the price or cost outcome rather than only the text.
Defensibility comes from ownership of the work
The durable moat in vertical AI is hard to forecast, and early businesses were rarely built around a predetermined sequence of steps. The more reliable signal is that the product owns more of the customer’s real workflow, creating dependencies and value that accumulate. A sales AI agent that handles prep, outbound, and inbound work is more defensible than a CRM that logs deals, and stickiness, network effects, and switching costs follow from that initial embedded use.
The build-versus-buy question surfaces here. A team can build a quote-to-SAP tool in about eight hours, but it will reach only around 70% of the performance, and that gap means humans must still check every output, sometimes creating more work than before. Enterprises usually conclude that internal builds are a diversion from their core competency. The last 20% of an internal build, which covers integrations, memory, workflows, and vertical data, often accounts for 80% of the effort, and that is where production readiness is decided.
Forward-deployed engineering and self-automation
Vertical AI still requires substantial forward-deployed work because customer data varies widely. The goal is to shift that work toward software. Lio’s team is roughly 85% engineers, and its forward-deployed engineers measure success by automating their own implementation tasks, so the customization burden shrinks with each deployment. The learning loop, guided by a sound evaluation process, lets systems take on more complex jobs over time.
Agents on both sides of the transaction
If buyers and suppliers each deploy agents, price becomes only one output among thousands of coordinated tasks. Both sides share incentives on most of those tasks: sales wants low friction, and buyers want speed to market for a data center or an aircraft. That shared interest suggests agents could coordinate across both sides on the 5,000 other tasks, with negotiation on price being a smaller part of the picture. Legal work offers a parallel, where opposing firms benefit from a shared view of the latest draft, agreed terms, and open issues, a coordination task that today is performed manually.
Suppliers are currently behind buyers in agent adoption, even though sales has historically led procurement. Procurement, though unglamorous, holds leverage over suppliers in industrial sectors because it dictates quality standards and RFQ formats, which creates an opening to deploy agents on the supplier side too.
Why procurement is ready for a shift
Procurement is widely disliked, and the tools that have proliferated over the past 25 years have made processes more efficient without changing how people actually work, which is still largely email, Teams, Excel, and PowerPoint. Lio’s events for procurement leaders, including one in New York with more than 100 leaders and a planned Munich event with around 700 attendees, use hands-on agent booths to show a cross-departmental workflow. The category combines emotional frustration, low innovation, and high business impact: a 1% margin improvement requires roughly 10% more revenue, so a small savings rate in procurement carries outsized weight in the P&L, which makes it a trillion-dollar opportunity.