Why AI Agents Can Beat the Incumbents

AI-native agents can beat incumbent software when they take responsibility for completing a job across systems, rather than stopping at the boundaries of a system of record. Vlad said procurement decisions often depend on supplier emails, stakeholder conversations, spreadsheets, engineering data, and negotiation history that an ERP does not capture. A record showing an $8,000 price, for example, may omit supplier pushback and cost-engineering analysis. An agent able to assemble that context and act on it can address the work itself, not merely retrieve or store its information.

Sima distinguished retrieval agents, which find and synthesize information; process agents, which execute defined steps; policy agents, which exercise judgment when rules are unclear; and principal agents, which weigh broader interests such as a customer relationship. Incumbents have largely delivered retrieval and simple process automation, she said. More capable agents can collide with internal product boundaries: a workflow tool and a product that resolves a job end to end may have different buyers, teams, and sales incentives. Leo’s procurement system coordinates agents and people across software; a bolt purchase, for instance, can involve checking inventory, contacting suppliers, issuing requests for quotes, processing email and attachments, evaluating options, and either buying autonomously or routing the case to a human negotiation or auction.

Leo started with invoice-processing “happy paths,” which Vlad said represented about 20% of the job, then moved to exceptions such as mismatches and possible fraud, which made up about 80%. The company begins negotiations with human oversight and adapts from enterprise feedback before increasing autonomy. Vlad described customer progression from 10,000 to 20,000 and eventually 100,000 negotiations as trust grows. Automation can make previously uneconomic work actionable: enterprises often left negotiations below $50,000 alone for lack of capacity, including a $40,000 invoice that might otherwise go unchallenged. Vlad said the downside of a poor negotiation in such cases can be nearly zero compared with doing nothing, but tone and continuity still matter when a vendor relationship is important.

The highest-stakes work calls for more human involvement. Leo focuses on complex procurement for aircraft, robots, drones, and data centers, where a component arriving two weeks late could cause hundreds of millions of dollars in damage or delay. Agents can extract delay information from procurement emails, assess project impact, and help determine a response, while a manager may be dealing with 500 messages. For multimillion-dollar negotiations involving 3D models and technical drawings, experts remain involved as agents work for hours. In direct procurement, a buyer may deal with 100 to 2,000 suppliers and spend up to $1 billion with one supplier; a major negotiation can take three months, involve 10 full-time participants, and devote 90% of effort to preparation. Agents can prepare and analyze, while staff and engineers retain a role in strategic decisions.

Vlad said Leo uses models from multiple providers and treats foundation models as commodities for general business tasks, combining them with a job-specific harness. Some tasks may reach full completion, while drawing analysis for should-cost models and price benchmarking may remain at 80%. The company’s enterprise benchmarking data is proprietary and unavailable to general-purpose models for training, and it has begun fine-tuning models for should-cost modeling with a stated goal of reaching “200%.” Vlad also cautioned that a promising prototype is not a production system: a 70%-accurate tool may still require people to check every data point. Integrations, memory, workflows, and sometimes vertical data make the final 20% of performance, in his estimate, account for 80% of the effort.

That implementation burden helps explain both the opportunity and the uncertainty around durable vertical AI companies. Sima argued that they need to own work end to end, build a data asset, and become important enough that customers depend on them, though the eventual moats are hard to predict. Vlad said one Fortune 500 company spent at least three or four months building a cash-collection product but ran into inadequate context and ongoing work mapping and testing across multiple ERPs and exception processes. Enterprise deployments are still early, with humans involved as vendors learn each customer’s data and processes. Leo says its forward-deployed engineers aim to make work self-service and automate their own jobs before moving on; 85% of employees are engineers, including people without engineering titles. Vlad expects agents eventually to work on both buyer and supplier sides, reducing friction around RFQs, contract drafts, open issues, and agreed terms even when the parties disagree on price.