Procurement's invisible 80%: why Lio's AI agents own the end-to-end job
Key points
Key takeaways from the a16z podcast interview with Lio co-founder and CEO Vladimir Keil, joined by a16z general partner Seema Amble (October 2026):
The incumbent's advantage and its limit. A legacy vendor pairing a model with its system of record has distribution and customer trust — Salesforce signed users up to AgentForce at near-zero incremental cost — but it is confined to that record and cannot complete the end-to-end job: resolving a cancelled-but-charged customer dispute spans billing, chat history and contracts, knowledge no single system of record holds.
Four agent archetypes. Seema Amble walks through retrieval agents (chatbots over a system of record), process agents (apply fixed policy, no judgement), policy agents (apply judgement where definitions are fuzzy) and principal agents (weigh trade-offs, such as offering more compensation to preserve a relationship).
Incumbents are held back, not holding back. Internal incentive conflicts stop enterprise software vendors from moving past retrieval and light process: workflow products and end-to-end resolution are different products to different buyers, fought over by different VPs in different orgs.
The ERP hides the real work. A procurement record such as 8k for aluminium conceals 30 stakeholder meetings, 500 emails, 20 spreadsheets, supplier pushback and weeks of cost-engineer modelling on 3D drawings — most procurement work happens outside the system of record.
Owning the whole arc. Lio ranges across retrieval, process, policy and principal agents depending on complexity and risk, running some use cases fully autonomously while keeping humans in the loop for high-risk, high-budget approvals.
Trust via the human in the loop. No enterprise starts with fully autonomous negotiation agents; the in-the-loop phase feeds the agent feedback until customers trust it for 10k, 20k and 100k negotiations — and cost engineers stay deliberately in the loop for multi-million-dollar deals involving complex technical drawings.
Invoice processing was a side quest. A clean invoice agent addresses 100% of the existing software market but only 20% of the job to be done; the other 80% is exceptions — fraud, mismatches, off-happy-path cases — which pushed Lio to build exception-handling agents that sync back to SAP and Oracle.
Savings below the radar. Enterprises never negotiated anything under roughly 50k for lack of capacity; Lio's agents capture that overlooked value, and the risk of a bad negotiation agent is near zero where no negotiation would have happened at all.
Ahead of the curve. Lio started weeks before ChatGPT's breakthrough and pitches each next-generation agent — retrieval, then process, now autonomous and long-running — to customers early, confirming real problems and shipping fast.
Read more: a16z Podcast — YouTube
