Context
Elena Burger of a16z talks with Seema Amble, an a16z partner who wrote a piece called "The Incumbents are Coming," and Vladimir Keil, co-founder and CEO of Lio, which builds AI agents for enterprise procurement. The question is one almost every AI application company now faces: if an established software vendor already owns the customer, the data and the system of record, and can plug in a strong model, where does a startup still win? Procurement is the test case because so much of its real work happens outside any system of record. For PMs, it is a close look at how to scope an agent product, earn trust, and decide what to build versus buy.
The Big Idea
Incumbents are limited to their own system of record, so AI-native startups win by owning the whole end-to-end job, including the messy work that happens in email, spreadsheets and conversations outside any system.
The record only shows the result, such as a final price. The value is in the hundreds of steps behind it, and an agent that coordinates those steps becomes something the customer depends on.
Key Insights
The record hides the real work
Vladimir's example is procurement, where the system of record looks simple.
- What the ERP shows: a single line, for example 8K for aluminum.
- What it hides: around 30 stakeholder meetings, 500 emails and 20 Excel sheets on the buyer side, plus the same on the supplier side.
- Deeper still: a supplier who pushed back and asked for 10K, and a cost engineer who spent three weeks on Excel and 3D modeling to work out what the part should cost.
His conclusion: most procurement work happens outside the ERP or any system of record. An incumbent chatbot on top of that record cannot see it.
One small email can cost millions
Procurement is not laptops and pencils. Vladimir says the hard cases are building aircraft, robots, drones and data centers.
- A single part arriving two weeks late can do hundreds of millions of dollars of damage and delay the whole project.
- The warning is often one of 500 emails in a procurement manager's inbox. The system of record usually stores only a date, not whether the delay matters.
- Lio's agents read the message, judge the impact, and work out how to resolve it.
Four kinds of agents
Seema describes four levels, using a customer asking for compensation after a service outage.
- Retrieval agent: pulls up the contract terms and confirms there was an outage. This is the chatbot most incumbents shipped.
- Process agent: applies the policy handbook, approves a credit and applies it to the bill. No judgment involved.
- Policy agent: applies judgment where the policy has no strict definition, such as deciding that 20 minutes counts as a significant outage.
- Principal agent: weighs whether to offer more than policy requires, because the outage was bad and the relationship is worth protecting.
She says most incumbents started in the first bucket and are now marketing the second and third. What they have launched is still mostly retrieval and light process work.
Why incumbents hold back
Seema gives two reasons incumbents stay shallow, even though they would like to go further.
- Distribution is a strength and a trap. Customers trust them, and turning on a feature is easy. Salesforce's Agentforce was easy to adopt, especially when offered at almost no extra cost. But a bad launch can destroy that trust.
- Internal conflict. A workflow tool for a human support agent and an agent that resolves the issue end to end have different buyers and different teams. The two products compete, and sales cannot agree on who to sell to.
Process agents were a side quest
Lio started with invoice agents, which retrieve data, match it across documents and push it to SAP or Oracle.
- Invoice software is a huge market, but Vladimir says that process is only about 20% of the work to be done.
- The other 80% is exceptions: a fraudulent invoice, a mismatch, anything off the happy path.
- So Lio moved fast to exception handling, and keeps pitching the next generation of agent to customers before it ships.
Trust grows through human oversight
No enterprise starts with fully autonomous negotiation agents, says Vladimir, because they do not trust the startup or the technology on day one.
- Lio starts with a human in the loop. Feedback from that human teaches the agent how one specific Fortune 10 company operates.
- Trust then grows, and customers let agents run 10K, 20K and 100K negotiations.
- On multi-million dollar negotiations with 3D models and technical drawings, Lio keeps experts in the loop on purpose. An agent runs for hours, then asks a cost engineer for feedback before the next step.
Start where nobody was negotiating
Vladimir calls this a hack for other startups. Before Lio, enterprises mostly did not negotiate anything below 50K because they lacked the capacity.
- A 40K invoice usually just gets paid.
- A bad negotiation agent there carries nearly zero risk, because the alternative was no negotiation at all.
- Relationship-heavy spend, like a podcast studio vendor, keeps a human in the loop because tone and the relationship matter more than price.
Direct procurement is mostly preparation
For direct procurement, meaning the 100 to 2,000 suppliers that go into a product like an aircraft, a single supplier can mean a billion in spend.
- Nobody runs an autonomous negotiation here. These negotiations take about three months with engineers analyzing aluminum and oil indices.
- Vladimir says 90% of the work is preparation. The result in the system of record is just that a billion became 900 million.
- So the agent works back of house, though Lio also tests live prompts, such as flagging that oil is up 10% but the product is only 30% oil, so the price should rise about 4%, not 10%.
Eight hours to build, 20% left
Customers ask why they cannot plug in a model and build it themselves.
- Lio's first product, putting a quote into SAP, is now a hiring exercise that engineers build in eight hours.
- But Vladimir says that gets you to roughly 70% performance, which does not mean 70% automation. People must still check everything, so it can create more work.
- The last 20% takes most of the effort, and needs integrations, memory, workflows and sometimes vertical data.
Agents on both sides
Vladimir expects agents on both the buyer and supplier side. Suppliers are further behind than expected, because in industrial companies the buyer dictates how suppliers answer an RFQ.
- Prices will not simply drop to zero. Price is the outcome of thousands of other tasks, and buyer and seller share an incentive on those, like speed and low friction.
- Seema adds that even adversarial parties, like two law firms, benefit from a shared view of the latest draft and open issues.
Mental Models & Frameworks
Retrieval, process, policy, principal agents
Seema's ladder measures how much judgment an agent applies.
- Retrieval: finds and summarizes information.
- Process: runs a defined sequence with no judgment.
- Policy: applies judgment where rules are vague.
- Principal: weighs the relationship and decides beyond policy.
Use it to place your own product honestly. Incumbents, she says, mostly sit at the first two levels. Lio spans all four, depending on risk and complexity.
Owning the end-to-end job
The test for a durable vertical AI company, per Seema: does your product own the whole arc of work, not one step in a record? Her examples are a customer cancellation that touches billing, chat history and contract, or a procurement job from request through shipment and invoice. Moats are hard to forecast, she says. The best early businesses just won trust and sold more, and stickiness and network effects followed once the customer depended on the product.
Human review as training data
Human review is not only a safety net. At Lio it is how the agent learns a specific customer's rules. The path: a human reviews each decision, the feedback feeds the agent, and autonomy expands as the customer trusts the results. Use it when the customer's own way of working is something you cannot know in advance.
Automate your own deployment job
Lio's forward deployed engineers are measured on automating themselves. Seema argues this is a structural edge: AI-native companies feed implementation learning back into the product, while incumbents treat implementation teams as an afterthought. Customization shifts from consulting projects to software the customer controls.
Decision Principles
Principle: Pick work with low downside
- When: choosing a first use case for an autonomous agent.
- Why: where the alternative is no action, like negotiating invoices under 50K that nobody negotiated before, even a mediocre agent has almost nothing to lose, and the customer sees savings fast.
Principle: Pitch the next agent early
- When: your product category moves faster than your release cycle.
- Why: Vladimir says Lio stayed slightly ahead of the curve by pitching the next generation of agent to customers, confirming it solved a real, paid problem, then shipping it weeks later.
Trade-offs & Nuance
Build internally or buy
Seema heard from a Fortune 500 management team that tried to build its own cash collection product. After at least three or four months, they found too little context and poor quality data from recordings and screen grabs.
- They also faced two different ERPs and a third on the way, with mappings to update and exceptions to remap.
- Her view: enterprises have a core competency and then internal tools, and should focus on the first. Vladimir adds that if procurement agents become a core competency, building might make sense.
Foundation model, harness, or fine-tuning
Vladimir says Lio treats models as a commodity and uses several providers. Some jobs get to near 100% with a good model and harness.
- Others, like cost engineers' should-cost modeling or price benchmarking, plateau around 80% even with the best harness.
- The data is not public and the answer is a judgment, a gut feeling about whether a quote makes sense. He sees outcome-based models, trained on text but outputting a price, as a way to close that gap.
- This is still early thinking, not a shipped result.
Incumbent trust cuts both ways
Incumbents start with the customer's trust and easy distribution. A startup has to earn trust before an agent can negotiate with third parties, but it can also ship bolder products without risking a large existing base.
Common Mistakes
Mistake: Building only the happy path
Invoice automation matched the happy path and was 100% of the software market, but only about 20% of the actual problem. Exceptions, like fraud and mismatches, are where the work is. Scope the product around those, not the demo.
Mistake: Confusing 70% performance with automation
An internal build that is 70% as good sounds close. Vladimir's point is that people must still check 100% of the output, so the team may end up with more work than before.
Mistake: Starting with full autonomy
Launching autonomous negotiation on day one asks for trust the customer has not given. Start with a human in the loop, collect feedback, and expand as results earn the trust.
Practical Application
Map the work outside the system
Pick one workflow your product touches and list every email thread, spreadsheet, meeting and approval between the start and the final record.
- Count them, like Vladimir's 500 emails and 20 Excel sheets behind one 8K price.
- The gaps between steps are the candidate for an end-to-end agent.
Place your roadmap on the ladder
Mark each agent feature as retrieval, process, policy or principal. If everything sits in the first two, you are competing with a chatbot an incumbent can add, so identify one judgment call to take on.
Plan the trust ramp
Write down the autonomy path for one agent feature.
- Start: human approves every action.
- Middle: human reviews only above a spend or risk threshold.
- Later: autonomous below that threshold.
- Always: keep experts in the loop on high-stakes work.
Scope the exceptions first
Before building, list the top exceptions, like fraud, mismatches and late deliveries. Ask what share of the job they are. If they are most of the work, as with invoices, the exceptions are the product.
Questions to Consider
- In your product, what is the final record a customer sees, and how much unrecorded email, spreadsheet and meeting work sits behind it?
- Which of your agent features use judgment, and which just retrieve or follow a fixed process that an incumbent could add with a model?
- Where could your agent act in a place with no current coverage, like Lio negotiating invoices under 50K, so that a weak first version still beats doing nothing?
- If a customer built a version of your product in eight hours, what would the missing 20% of performance cost them in manual checking?
Bottom Line
Incumbents own the record, but the real work lives outside it in emails, spreadsheets and expert judgment. AI-native startups win by owning that end-to-end job, earning trust through a human in the loop, and handling the exceptions that internal builds and chatbots do not reach.
Resources Mentioned
| Resource | Type | Why it was mentioned |
|---|---|---|
| "The Incumbents are Coming" by Seema Amble | Article | The piece that frames whether AI-native startups can beat incumbents who add a model to their existing software. |
| "Investing in Lio" by Seema Amble | Article | a16z's announcement of its investment in Lio, linked in the episode notes. |
Concepts to Explore
Should-cost modeling
Cost engineers analyze drawings to estimate what a part should cost, rather than accepting a supplier quote. Vladimir names it as a task where a foundation model plus a harness plateaus and where a fine-tuned model may help.
Direct versus indirect procurement
Indirect covers items like MRO parts, laptops and marketing services. Direct covers the parts that go into the product, such as an aircraft, from 100 to 2,000 strategic suppliers. The two need very different levels of automation.
Forward deployed engineering
Engineers embedded with customers to handle their specific processes. Both guests argue the goal is to automate that work into self-service software, so the team's job is to automate itself.
Case Studies Mentioned
Lio's bolt purchase
Vladimir walks through buying a bolt, which Lio can run fully autonomously.
- A requester uploads a photo, quote or Excel sheet, instead of working in SAP.
- Agents check inventory, ask other plants, and query internal and external suppliers.
- Another agent drafts and emails the RFQ, then reads replies in email, PDF and Excel.
- Based on price benchmarking, agents pick a negotiation route: strategic with a human, autonomous, or an e-auction.
- Then order confirmation, shipment tracking and invoices follow.
Fortune 500 cash collection build
A Fortune 500 company tried an internal cash collection product and found its recordings and screen grabs were not good enough context. With multiple ERPs and more coming, they decided the internal build did not make sense.
Lio's buyer and supplier summit
Lio hosted its third event in New York with over 100 procurement leaders at C level or vice president. The next, in Munich, is expected to have around 700. Vladimir says procurement people hate their tools because 500 products made the process more efficient but never changed that they work in email, Teams, Excel and PowerPoint.
Tools & Products
| Tool / Product | What it does | Why it was mentioned |
|---|---|---|
| Lio | Multi-agent system for enterprise procurement | Vladimir's company, covering sourcing, RFQs, negotiation, shipment tracking and invoices. |
| Agentforce | Salesforce's agent product | Seema's example of an incumbent agent that was easy for customers to adopt. |
| SAP and Oracle | Enterprise ERP systems | The systems of record that Lio's agents push data into. |
| Coupa | Procurement software | Mentioned as the legacy system a skeptical procurement buyer would not replace. |
People to Follow
Vladimir Keil
Co-founder and CEO of Lio. He started the company before ChatGPT and argues procurement is boring, emotional and high-impact, which makes it a rich place to build agents.
Seema Amble
Partner at a16z and author of "The Incumbents are Coming" and "Investing in Lio." She tracks how AI-native startups compete with incumbents.
Notable Quotes
"Most of the work in procurement actually happens outside of this ERP or any system of record." (Vladimir Keil)
"No company, no enterprise starts with fully autonomous negotiation agents from day one." (Vladimir Keil)
"It's really, really hard to forecast your moat going forward." (Seema Amble)
