Why AI Is Reinventing How Businesses Buy Everything
Summary
The startup advantage is not a better chatbot; it is ownership of an end-to-end job that incumbent systems of record do not capture. Procurement’s apparent “$8,000” result can conceal 30 stakeholder meetings, 500 emails, 20 spreadsheets, engineering analysis and supplier counteroffers. Seema Amble’s opportunity statement is categorical: “We’re going to take over this whole end-to-end process.”
AI agents progress from retrieving information to exercising increasingly consequential judgment. Seema separates search, process, policy and principal agents: confirm an outage, issue contractually required credit, decide whether 20 minutes constitutes a significant outage, then potentially exceed policy to preserve the customer relationship. Incumbents remain concentrated in retrieval and simple workflow execution because performing the work conflicts with software designed to support a human worker.
Trust, rather than raw model capability, determines how quickly autonomy reaches production. Lio begins with humans in the loop, absorbs company-specific feedback, and earns permission to handle “$10,000, $20,000, or $100,000” negotiations; complex multimillion-dollar cases deliberately retain engineers and procurement experts. The startup can advance incrementally, while an incumbent risks damaging an already-trusted franchise by releasing unreliable autonomy.
Procurement turns seemingly trivial information failures into material operational risk. A two-week delay buried among 500 emails can postpone an airplane or data-center project and cause hundreds of millions of dollars in losses. The agent’s valuable work is not recording a revised date but deciding “does it affect anything,” quantifying the impact and identifying a remedy.
Building a convincing AI demo has become radically easier, but deploying dependable automation has not. Lio’s original product—extracting a quote and entering it into SAP—once required its founding engineering team and can now be reproduced by a job candidate in eight hours. Yet 70% accuracy may yield “100% of the work” for human reviewers, potentially creating more work rather than automating it.
Durability comes from proprietary workflow context, specialized evaluation and customer dependence—not a confidently predeclared moat. General models can read PDFs and create spreadsheets, but they may stall at 80% on cost engineering or price benchmarking because the necessary data and tacit judgment are private. Seema argues that the strongest signal is whether the customer increasingly depends on the product to perform meaningful work.
Two-sided procurement agents could expand Lio from buyer automation into the infrastructure through which companies transact. Price remains zero-sum, but Vlad Keil argues it is merely the output of “5,000 other tasks” where buyers and suppliers share incentives to remove friction and accelerate delivery. His economic framing is equally direct: a 1% margin improvement can require 10% more revenue to replicate, making even modest procurement savings disproportionately valuable.
Deep dive
1. Systems of record capture outcomes, while the real work happens elsewhere
Seema’s framing replaces the old distribution-versus-innovation contest with a three-way fight: startups now face incumbents that can add AI models to existing data, customer relationships and installed workflows. Her answer is that a legacy vendor remains bounded by its system of record, while a startup can say, “We’re going to take over this whole end-to-end process.”
Customer service makes the gap concrete. Resolving an improper post-cancellation charge requires billing records, contract terms, cancellation history and prior conversations—not merely the data inside one application. The same logic extends from preparing legal materials through trial, or across procurement from an initial request through negotiation, delivery and invoicing.
Vlad’s procurement example exposes how misleading the recorded outcome can be: an ERP may show “$8,000” for aluminum while omitting a supplier’s $10,000 counteroffer, three weeks of engineering analysis, 30 stakeholder meetings, 500 emails and 20 spreadsheets. “Most of the work in procurement actually happens outside of this ERP” or accounting system.
Seema divides agents into four levels: search agents retrieve facts; process agents execute predefined steps; policy agents interpret ambiguity; principal agents weigh broader objectives. During an outage, that means confirming what happened, applying contractual credit, judging whether 20 minutes was “significant,” then deciding whether extra compensation is warranted to preserve the relationship.
2. Incumbents are restrained by product conflict, while startups are constrained by trust
Seema’s refinement is that incumbents are not voluntarily holding back; they are “being restrained.” Their distribution and customer trust make adoption easy—especially when an agent is bundled cheaply—but an agent that completes support work competes with software designed for human support representatives, creating conflicting products, buyers, sales incentives and internal owners.
Vlad adds external trust to Seema’s internal conflict. An incumbent can lose an established reputation by launching unreliable autonomy, while Lio must earn trust from scratch—especially when its agent communicates with third parties. “No company or enterprise starts with fully autonomous negotiation agents from day one.”
The deployment path therefore begins with human review and expands through accumulated feedback. Lio may autonomously handle lower-risk negotiations, then earn authority over “$10,000, $20,000, or $100,000”; multimillion-dollar negotiations involving drawings and 3D models intentionally pause for a cost-estimating engineer before the agent continues.
3. Exception handling—not the happy path—is where automation becomes valuable
Lio initially focused on process agents such as invoice matching and ERP entry, then concluded that this represented perhaps 20% of the problem. The remaining 80% consisted of exceptions: suspected fraud, disagreements, nonstandard cases and decisions that accounting software was never built to resolve.
Smaller purchases provide a low-risk proving ground. Enterprises historically ignored negotiations below $50,000 because they lacked the resources to bargain; Vlad jokes that vendors could send a $40,000 invoice and expect payment. If the alternative was no negotiation, even an imperfect agent has “almost zero” incremental downside.
Elena’s pushback—worth keeping—is that spend alone does not determine risk. A podcast studio may be inexpensive yet strategically sensitive because its operators understand a16z’s preferences. Lio therefore retains people where relationships, tone or supplier continuity matter, even if most of the transaction can run autonomously.
Strategic negotiations demand a different human-agent boundary. Contract structure touches legal teams, collection terms involve finance, part economics require cost engineers, and commercial terms belong to procurement; these are “highly qualified specialists” supplying company- and industry-specific knowledge to long-running agents, not interchangeable back-office reviewers.
4. Physical procurement converts context into decisions with enormous downside protection
Vlad’s sharpest example is an airplane assembled through thousands of suppliers. One vendor’s “sorry, this part will arrive 2 weeks later” may sit among 500 emails, yet missing it can delay the project and cause hundreds of millions in losses. Recording “next Wednesday” is insufficient because the same date change could be routine or catastrophic.
Elena challenges what software can actually change when cargo is damaged or a road is blocked. Vlad’s answer is probabilistic: the event may be irreversible, but broader context can make it predictable. A supplier with a 20% shortage rate may be ten times cheaper than one at 1%, yet the expensive supplier can still be rational when delivery failure threatens the whole project.
Lio’s agents therefore combine internal demand, supplier performance, seller information and external signals such as news—and, Vlad suggests, potentially Polymarket. The objective is to determine impact and remediation, expanding procurement into “how businesses do business with each other,” not merely automating an internal purchasing department.
A bolt illustrates the multi-agent chain: accept a photo, quote or spreadsheet; check inventory across factories; identify internal or external suppliers; issue an RFQ; parse emails, PDFs and Excel responses; compare prices; select autonomous negotiation, an auction or human-led strategy; then manage ordering, confirmation, shipment and invoices from A to Z. This architecture is necessary because the human process spans eight people, eight stakeholders, three departments and five software tools, which the agents must coordinate in sequence.
5. Direct procurement moves agents from transaction automation into expert preparation
Vlad distinguishes indirect spend—maintenance, factory construction, laptops, pencils, marketing and podcast services—from direct components used to build airplanes, drones or robots. Direct procurement may involve only 100 to 2,000 suppliers rather than 50,000, but one supplier can represent $1 billion of strategically important spend.
For those negotiations, “90% of the work is preparation.” The visible result might be a reduction from $1 billion to $900 million, but it follows three months of work by ten people analyzing drawings, quality, commodity markets and supplier economics. For now, agents mainly strengthen this back office rather than replace the person across the table.
Vlad nevertheless sees a live front-office role. If a supplier cites a 10% rise in oil to demand a 10% price increase, an agent could immediately flag that oil represents only 30% of the product and suggest something closer to 4%. The value is context delivered inside the negotiation, not a robot shaking hands.
6. The last 20% of deployment defines both the moat and the market
Vlad treats general models as commodities for reading, PDFs and spreadsheets, but says tools plus the best base model can still stop at 80%. Cost modeling from technical drawings and price benchmarking depend on proprietary data and tacit judgment—the purchasing manager’s “gut feeling” that cannot easily be converted into written rules—creating a case for outcome-focused specialized training.
Lio’s first viral workflow, quote extraction into SAP, once required its founding team; today a job candidate receives roughly eight hours to recreate it. But “70% productivity” does not equal 70% automation: if every output needs verification, people still perform 100% of the work and may inherit additional review overhead.
Seema recounts a Fortune 500 team concluding that in-house development did not make sense after three or four months because recordings and screenshots produced inadequate context, while two ERP systems—and a planned third—created a mapping and maintenance burden. Elena’s Pareto-style conclusion is that the final 20% required for production can consume 80% of the effort.
That deployment burden also shapes Lio’s operating model: about 85% of its people are engineers. Its FDE and implementation work is intended to reduce customization, create self-service and automate itself rather than turn Lio into a consulting firm.
Seema’s moat test is practical dependence: own the workflow, create the data asset, earn trust and perform enough work that the customer cannot casually remove the product. She also says learning cycles and the right evaluation process can let AI take on increasingly complex tasks.
Seema closes with the category economics—1% of savings can equal the income-statement effect of 10% more sales—and calls the combination of “boring,” emotionally painful and economically critical procurement a “trillion-dollar business opportunity.”