> ## Content Index
> Fetch the complete content index at: https://www.signalful.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# The 500 Emails Behind the Price
- URL: https://www.signalful.com/the-500-emails-behind-the-price/
- Published: 2026-10-06T15:59:00.000Z
- Updated: 2026-10-07T05:15:39.000Z
- Description: The ERP says aluminum costs 8K. It doesn't show the three weeks of spreadsheets behind that number. Lio's agents do the work that never gets recorded.
- Author: Signalful Editorial
- Tags: Builders, Lio, Vladimir Keil, AI Agents, Procurement, Enterprise Software, #sf-0051, #yt-OTQ-lFsq7zA, #Import 2026-10-06 15:50

Vladimir Keil is the co-founder and CEO of Lio, which builds AI agents for enterprise procurement. He started the company about 3 years ago, a few weeks before the ChatGPT breakthrough, and Lio's first product pulled supplier quotes into SAP, a use case he says went viral in the procurement world. Lio now runs agents across sourcing, negotiation, ordering and invoicing for both indirect spend and the direct parts that go into aircraft, drones and data centers. About 85% of the company are engineers, and Lio hosts a recurring procurement leaders event in New York, with the next one planned for Munich at around 700 people. These signals are from Vladimir Keil's interview with Seema Amble and Elena Burger on the a16z podcast.

1. **The 500 Emails Behind the Price.** Most procurement work happens outside the ERP, Keil says. The system of record shows "8K for aluminum," but not that the supplier pushed for 10K or that a cost engineer spent 3 weeks on spreadsheets and 3D models to find the right price. He compares it to a signed sales contract that hides 30 stakeholder meetings, 500 emails and 20 Excel sheets. Amble says this is where an AI-native startup can beat an incumbent that adds a model to its own data: by owning the "end-to-end arc" of a job that crosses billing, contracts, chat history and people.
2. **Four Kinds of Agents.** Amble sorts agents into retrieval, process, policy and principal, using a customer asking for credit after a service outage. A retrieval agent confirms the contract terms and the outage, a process agent applies the credit, a policy agent judges whether 20 minutes counts as significant, and a principal agent decides to pay more to protect the relationship. She says most incumbents started at retrieval and have barely reached process. Their incentives hold them back: a product that resolves the work competes with one that sells workflow to human agents, the two have different buyers, and "two VPs" in different orgs can't agree on what to sell.
3. **The 80% the Software Missed.** Keil says invoice software is a clear process: pull the data, match it against other documents, push it into SAP or Oracle. That process is "100% of the software market" and only about 20% of the job. The other 80% is exceptions: fraudulent invoices, mismatches, anything off the happy path. Lio moved quickly to agents that handle those exceptions, and Keil treats process agents as "kind of like a side quest."
4. **Trust by the Thousand.** "No company and no enterprise starts with fully autonomous negotiation agents from day one," Keil says, because customers trust neither Lio nor the technology yet. Lio starts with a human in the loop, which teaches its general negotiation agent how a specific Fortune 10 company operates. As the feedback builds, customers hand over 10,000, then 20,000, then 100,000 negotiations. Keil adds that incumbents already hold customer trust, so shipping an agent too early puts that trust at risk in a way a startup doesn't face.
5. **The $50K Blind Spot.** Before agents, enterprises ignored most spend under $50,000 because they lacked the people to negotiate it. Keil's tip for other startups: send an enterprise a $40,000 invoice and it will probably just pay, unless it runs Lio. On that spend, the risk of a mediocre negotiation agent is "nearly zero," since the alternative was no negotiation at all. Relationship vendors are the exception, and his example is the studio crew that already knows how a16z records podcasts, where agents run mostly autonomously but procurement still controls the tone.
6. **The Two-Week Email.** Building an aircraft or a data center means coordinating thousands of suppliers, and Keil says one part arriving 2 weeks late can cost hundreds of millions of dollars. The warning often comes as one of 500 emails in a procurement manager's inbox, while the ERP only records the new date. Lio's agents read that email and decide whether the delay matters and how to fix it. For events no one can stop, like cargo lost at sea, he argues context turns the problem into probability: a supplier that loses 20% of goods may be worth dropping for one that loses 1% at 10x the price, and agents can also watch the news and even Polymarket bets on disruptions.
7. **Ordering a Bolt.** Keil walks through a part like a bolt for Boeing, which Lio can buy fully autonomously. A construction worker who opens a laptop every second week can't file a request in SAP, so Lio lets him upload a photo, a quote or a spreadsheet, since "no one cares outside of the procurement department" about GL accounts or framework contracts. Agents then check inventory, ask another plant, find internal or external suppliers, send RFQs by email, and parse replies that arrive as email text, PDFs and Excel files. After price benchmarking, an agent picks the next step (human-led negotiation, autonomous negotiation or an e-auction) and then handles the order confirmation, shipment tracking and invoices.
8. **Direct Spend Stays Back of House.** Direct procurement covers the 100 to 2,000 suppliers that actually make the airplane or robot, sometimes with $1 billion of spend on one supplier. There Keil wants a 3-month negotiation with engineers tracking aluminum and oil indices, and he says "90% of the work is preparation," often 10 people full-time to turn $1 billion into $900 million. Lio's agents work behind the human negotiators, with some real-time help at the table. His example: a supplier asks for 10% more because oil rose 10%, and a pop-up notes the part is only 30% oil, so the increase should be closer to 4%.
9. **Models for Gut Feel.** Lio uses models from every provider and treats them as a commodity for general work like reading PDFs and building spreadsheets. Keil says a strong model plus good tooling gets some tasks to 100% but stalls others at 80%, including should-cost modeling and price benchmarking. Those depend on private pricing data from one or many enterprises that general labs can't train on, so Lio has started fine-tuning and is looking at models trained to output a price. A quote from BCG and one from McKinsey can differ 10x for the same work, and only an experienced procurement manager has the "gut feeling" to tell which is right, a judgment that person can't write down as rules.
10. **The 70% Trap.** Lio's first product, the quote-to-SAP agent, is now an 8-hour take-home for engineering candidates. Keil says anyone can build that, and a procurement team could build much more in 2 months, but it reaches about 70% performance, and 70% accuracy can still leave people checking 100% of the work. Amble adds a Fortune 500 management team that spent at least 3 or 4 months building its own cash collection tool, found the context too thin, and faced two ERPs plus a third coming through an acquisition. Keil's rule: the last 20% decides whether anything reaches production, and building it should only happen if AI procurement is meant to be a core competency.
11. **Agents on Both Sides.** Keil expects agents for buyers and suppliers, and he finds Fortune 500 suppliers less advanced than he assumed: heavy users of meeting recorders, light on deployed agents. In industries like automotive, procurement dictates how suppliers answer RFQs, so Lio sees a path to put its agents on the supplier side too. Price is the one adversarial point, he says, and it is the result of about 5,000 other tasks where both sides want the same thing: buyers want speed to market and sellers want a sale with little friction. Amble extends the idea to law firms, where opposing counsel would both benefit from a shared record of drafts, open issues and agreed terms.
12. **Boring, Emotional, Huge.** Keil says nobody likes working with procurement, including requesters, suppliers and procurement staff, even though LinkedIn market maps list some 500 tools. Those tools made the process more efficient without changing the work, which still runs on email, Teams, Excel and PowerPoint. One of the hosts recalls a Coupa user who gave her "the most disgruntled customer call" she has ever done. Keil sees the frustration as an opening, since the last real change in the category was about 20 years ago, and the payoff is large: a 1% margin gain needs about 10% more revenue, while 1% in savings gets the same P&L result.

Watch the full video at [https://www.youtube.com/watch?v=OTQ-lFsq7zA](https://www.youtube.com/watch?v=OTQ-lFsq7zA&ref=signalful.com).