AI Agents and the Work Behind Enterprise Procurement

a16z’s Seema Amble and Elena Burger talk with Lio co-founder and CEO Vladimir Keil about AI agents in enterprise procurement. a16z invested in Lio, so the participants have interests in the company discussed.

The hosts ask what AI-native startups can offer when incumbent software companies already have customers, data and systems of record.

The guest says important work often happens outside formal systems: emails, spreadsheets, meetings and cross-team judgments may not be fully captured in a database.

He uses a purchase order as an example: one amount in an ERP may be the result of extensive supplier communication, technical analysis and approvals.

Recording a final result and understanding how it was reached are different capabilities. An agent participating in work needs appropriate context and permissions.

They use customer support to distinguish levels: retrieving information, executing a defined process, applying policy and making more complex judgments within boundaries.

Higher-level judgment may involve customer relationships, risk and exceptions; its correctness cannot be assumed from the degree of automation.

The discussion turns to organizational incentives at incumbent software companies. The guest says different departments, products and buyers can make cross-system workflow changes difficult; this is his competitive analysis.

Enterprise customers need to trust how a new system handles data and executes tasks. Trust requires permissions, auditability, actual performance and gradual adoption, not only a sales demo.

Keil recalls starting with document processing and process agents before finding that customers needed coverage of more complete procurement work.

He describes existing systems and people in customer procurement, emphasizing that a startup must learn each enterprise’s processes rather than assume they are identical.

He describes using human participation to collect feedback and improve the system gradually. Human oversight can support learning and control but does not automatically eliminate errors.

On automated negotiation, he says risk depends on amount, relationship and task. Enterprises may not allow agents to handle their most consequential decisions autonomously on day one.

Procurement is not only about the lowest price; it can involve contract terms, supplier reliability, legal, finance, engineering and operations.

In a large manufacturing or infrastructure project, a delayed part can affect overall timing. Loss figures in the interview are illustrative scenarios, not values for every project.

The guest says a system can combine supplier information, project status and external signals to help assess delivery risk; prediction still needs quality data and validation.

Sometimes a higher unit price may buy greater delivery certainty. Whether it is worthwhile depends on the contract, project constraints and human risk judgment.

Keil describes multiple agents sharing information and handling different tasks in procurement. He says the goal is to cover a workflow, not merely answer one question.

An example begins with a parts order and involves specifications, supplier selection, requests for quotes and records in a procurement system. Different departments may provide context at each step.

The system must also handle contracts and delivery information. The guest’s description of autonomy is a company product claim, not performance independently verified by the interview.

He uses a long negotiation as an example: engineers may analyze materials, supply and prices in advance; back-office preparation and supplier-facing negotiation are both parts of the workflow.

He says models can help organize information and generate suggestions, while companies must decide which actions can be automatic and which require approval.

They discuss integration across systems and error handling in long workflows. Real deployment needs action records, permission controls and appropriate human handoffs.

On model selection, he says the company uses models from several providers and envisions combining foundation models with external workflows for different tasks.

He says some tasks need specialized domain knowledge and evaluation. General-model capability does not automatically cover every enterprise’s specific procurement policies.

On durable advantages for vertical AI companies, they discuss domain knowledge, integration, trust and sustained execution while acknowledging that moats are hard to predict.

Asked whether enterprises can build a similar system themselves, the guest says an initial prototype may arrive quickly but reliable integration, context and maintenance are often harder.

He illustrates that completing part of a workflow while people must still inspect all results makes coverage and actual labor saved different measures.

A host adds an example of insufficient context in an enterprise project. Migrating a workflow requires understanding old systems, data meaning and new responsibilities.

If buyers and suppliers both use agents, systems may need to coordinate shared files, current terms and status. More efficient communication does not mean their interests are identical.

When agents on both sides negotiate, conflicting objectives, authorization, transparency and accountability need explicit design. The future scenario is not an established industry outcome.

They discuss speeding analysis of supplier responses to avoid administrative delay; actual gains depend on data, process and system reliability.

Enterprises often need software customized to their policies and processes. The guest says AI tools may change how customization is done, while deployment and governance work remains.

He describes implementation teams helping customers configure systems and hopes to productize repeated work. This is an account of the company’s approach.

For procurement staff, adding tools does not necessarily change how work gets done. The participants discuss serving real work rather than merely adding interfaces.

The close touches on procurement communities and business impact. The guest is optimistic about the field, but company effectiveness and economic benefits require independent validation in customer data.