a16zWhy AI Is Reinventing How Businesses Buy Everything
CHAPTERS
- 0:00 – 0:58
Why procurement failures can cost hundreds of millions
Vlad opens with a vivid example: a single missed supplier email can delay critical parts and cascade into massive financial impact. This sets the stage for why “deploying” AI into real operations is harder—and more valuable—than building a demo agent.
- •Procurement in manufacturing relies on thousands of suppliers
- •Operational details (like delivery changes) can create outsized losses
- •Information often lives in email, not in systems of record
- •The real risk is missed context and missed decisions
- 0:58 – 4:19
Why AI-native startups can beat incumbents: owning the end-to-end job
Seema frames the new competitive landscape: incumbents can bolt models onto existing systems, but that doesn’t complete the full job. Startups win by owning the entire workflow across systems, people, and unstructured context—not just a single system of record.
- •Incumbents pair distribution with AI add-ons, raising the bar
- •Systems of record capture only a slice of the workflow
- •The ‘job to be done’ spans multiple tools and departments
- •Startups can design around the full end-to-end arc (e.g., legal, procurement)
- 4:19 – 5:34
The hidden work behind an ‘$8K’ line item
Vlad explains that procurement outcomes look simple in ERP (a price), but the work is complex and mostly invisible: stakeholder meetings, emails, spreadsheets, engineering analyses, and negotiations. This “dark matter” outside the system of record is where agents must operate to create real value.
- •ERP/records show the outcome, not the effort or rationale
- •Procurement involves stakeholder alignment and extensive back-and-forth
- •Engineering work (e.g., 3D modeling, cost analysis) informs pricing
- •Most critical context lives outside ERP in email and files
- 5:34 – 10:59
Retrieval vs. process vs. policy vs. principal agents (and why incumbents stall)
Seema lays out a four-level taxonomy of agents, from simple retrieval to principal-level judgment calls. She argues incumbents started with ‘chat on top’ retrieval, and face organizational and product incentives that make it hard to move into higher-judgment agents that actually resolve work end-to-end.
- •Retrieval agents: find/summarize info from records
- •Process agents: execute deterministic workflows with low judgment
- •Policy agents: apply interpretation when rules are fuzzy
- •Principal agents: optimize for broader goals/relationships and tradeoffs
- •Incumbents face product cannibalization, org conflict, and sales complexity
- 10:59 – 15:02
Trust, autonomy, and the human-in-the-loop ladder for negotiations
Vlad describes how Lio spans the agent spectrum depending on risk and complexity, with autonomy increasing as trust grows. Human-in-the-loop feedback becomes a learning loop: it adapts agents to each enterprise’s unique ways of working and safely expands usage from thousands to hundreds of thousands of negotiations.
- •Autonomy varies by budget, risk, and deal complexity
- •Enterprises don’t start with fully autonomous negotiation agents
- •Human-in-the-loop accelerates learning and enterprise-specific adaptation
- •Long-running tasks (e.g., complex negotiations) keep experts in the loop
- 15:02 – 21:42
What procurement actually looks like across legal, finance, engineering, and relationships
The conversation broadens procurement beyond “price” into cross-department coordination and relationship management. Vlad distinguishes low-risk spend where agents can negotiate autonomously from cases where vendor relationships and tone require human oversight.
- •Procurement touches legal, finance, cost engineering, and operations
- •Many negotiations were never done before due to capacity limits (e.g., sub-$50K spend)
- •Relationship-sensitive vendors require human review of tone/approach
- •Negotiation includes terms, collections, commercials—not just price
- 21:42 – 26:26
Procurement at Boeing-scale: physical-world constraints and risk prediction
Using aircraft/data center builds, Vlad shows how small delays can derail massive projects. Agents can’t prevent every disruption, but with enough context they can predict risk, select more reliable suppliers, and proactively mitigate downstream impacts.
- •Large programs depend on thousands of coordinated deliveries
- •A single delayed part can postpone projects and destroy value
- •Agents can model supplier reliability and trade cost vs. risk
- •External context (news, signals) can be fused with internal data to predict disruptions
- 26:26 – 28:03
Why Lio is multi-agent: coordinating an end-to-end job across tools and stakeholders
Vlad explains that end-to-end procurement requires multiple specialized agents that share context and act in sequence. A single agent can’t cover contracts, sourcing, external signals, and execution; multi-agent orchestration mirrors how multiple humans/departments execute procurement today.
- •End-to-end work requires sequencing and shared state across agents
- •Agents cover specialized domains (contracts, sourcing, external context, etc.)
- •Human workflows span many tools (email, ERP, docs), so agents must too
- •Multi-agent coordination is necessary to replicate cross-functional execution
- 28:03 – 30:39
A bolt order, end-to-end: from demand capture to RFQ, negotiation, shipping, and invoices
Vlad walks through a concrete example: procuring a bolt autonomously. The key insight is demand intake and messy inputs (photos, PDFs, Excel, emails) are the norm; agents must normalize inputs, source internally/externally, run RFQs, handle negotiation modes, and manage downstream execution.
- •Demand capture is often the hardest step for non-procurement employees
- •Agents translate messy artifacts into procurement structure (GLs, categories)
- •Workflow includes inventory checks, internal sourcing, RFQs, response parsing
- •Negotiation may be autonomous, human-guided, or auction-based
- •Downstream tasks include order confirmation, shipment tracking, and invoicing
- 30:39 – 34:26
Indirect vs. direct procurement: where autonomy ends and ‘should-cost’ begins
Vlad distinguishes indirect procurement (high-volume, automatable) from direct procurement (strategic, high-stakes, slower). In direct procurement, most value comes from months of preparation—analysis of drawings, indices, quality, and cost structure—where agents augment expert teams rather than replace them.
- •Indirect: MRO, laptops, services—optimize for automation and coverage
- •Direct: airplane/drone components—fewer suppliers, strategic spend concentration
- •High-stakes negotiations require deep prep and multi-month cycles
- •Agents help with preparation and can assist in real-time negotiation with fact-checking
- 34:26 – 37:58
Models, harnesses, and when fine-tuning matters (pricing and should-cost modeling)
Vlad argues foundation models are becoming commoditized; differentiation comes from harnessing (integrations, workflows, memory, evals) and selective specialization. Certain tasks—like price benchmarking and should-cost estimation—may require fine-tuning or outcome-trained models because the needed data is proprietary and intuition-heavy.
- •Lio uses multiple models; providers are treated as interchangeable where possible
- •Harnessing can get many workflows to ‘production-grade’ performance
- •Some domains cap at ~80% without specialization (e.g., should-cost from drawings)
- •Price benchmarking needs proprietary data not available to general models
- •Outcome-based training may be better than pure text generation for some tasks
- 37:58 – 40:07
What a durable vertical AI company looks like: moats emerge from dependency, not slideware
Seema describes durability as hard to predict upfront; the best early signal is whether customers become dependent because the product does real work. As vertical AI expands from logging work to performing it, stickiness and defensibility become downstream effects of value delivered.
- •Moats are difficult to forecast early; execution and trust matter first
- •Owning more of the workflow increases dependency and lock-in
- •Vertical AI shifts from systems of record to systems that do the work
- •Defensibility follows from usage, switching costs, and learning loops
- 40:07 – 44:47
Why ‘just build it in-house’ and ‘just add a model to ERP’ usually fails
Vlad and Seema explain that building a prototype is easy, but reaching production reliability is not—especially with constant exceptions, data quality issues, and ERP changes. The last 20% (integrations, evals, mappings, exception handling) drives most of the effort and determines whether automation is real.
- •Quick builds often reach ~70–80% accuracy but still require full human checking
- •Partial automation can create more work if error handling isn’t solved
- •Enterprise environments change (multiple ERPs, acquisitions), breaking DIY tools
- •Production requires evals, memory, workflows, and continuous maintenance
- •Enterprises should focus on core competencies vs. maintaining AI agent stacks
- 44:47 – 48:55
When both sides deploy agents: buyers, suppliers, and shared incentives beyond price
Vlad predicts agents on both buyer and supplier sides, even if price is adversarial. He argues most underlying tasks (RFQ processing, coordination, speed, reducing friction) are aligned incentives, making a shared platform plausible—where procurement can even push suppliers to adopt compatible automation.
- •Supplier-side agent adoption appears behind procurement in many industries
- •Procurement can dictate standards/tools to suppliers, enabling network expansion
- •Even with adversarial pricing, thousands of coordination tasks are win-win
- •Platforms may coordinate drafts, open issues, and status across parties (incl. legal)
- 48:55 – 52:22
Customization and forward-deployed engineering: automating the deployment itself
Seema and Vlad discuss why early vertical AI requires forward-deployed work to understand messy customer data and workflows. Lio’s approach is to productize and self-serve customization over time—measuring FDE success by how effectively they automate their own deployment tasks.
- •Early-stage deployments require humans due to data/process variability
- •Incumbents struggle to integrate deployment learnings into product cycles
- •Learning loops and evals enable progressively more complex automation
- •FDEs should ‘automate themselves’ to reduce customization cost over time
- •High engineering concentration supports rapid iteration and deployment automation
- 52:22 – 59:13
Bots and Buyers: why procurement leaders are suddenly paying attention
Vlad shares learnings from Lio’s events: procurement has many tools but remains email-and-spreadsheet driven, with low satisfaction across requesters, suppliers, and procurement teams. Hands-on demos that show end-to-end outcomes (not isolated features) change perception and create momentum in a historically stagnant category.
- •Despite hundreds of tools, procurement work still lives in email/Excel/Teams
- •Legacy tooling improved efficiency but didn’t change how work is done
- •End-to-end demos make ‘agents’ concrete and physical-world relevant
- •Procurement is ‘boring but emotional’ with massive P&L leverage
- •Small savings translate to large profit impact; category has trillion-dollar upside
