Invoice Agent

Facility managers approve invoices by cross-checking dozens of data points per work order, pricing, vendor history, time on site, billing codes. Invoice Agent runs those same checks automatically and recommends approve, approve with notes, or review required. The design challenge: making managers trust an agent that touches money.

The Problem

A manager already makes those checks, and makes them fast. For the agent to earn a place in that moment it has to be quicker than the manager's own check, ready when the invoice opens instead of run on request, and able to work without a long setup.

Research Findings

Sessions with facility managers revealed the core tension: active users treat the agent as "another set of eyes," a confidence check after their own review, not a replacement. They're managing client money and want to stay in control. The certainty score was misread as "how likely to approve" instead of "how confident the AI is." Every account follows unique processes.

Decisions

The agent recommends, the manager decides. This is client money, and research with facility managers was clear that they wanted a second set of eyes, not a replacement. So the output is a recommendation attached to the invoice, read at the moment of approval, and the approve button stays with the person.

A middle outcome, not a binary. The agent recommends approve, approve with notes, or review required. A binary would have turned every partial pass into a rejection and taught managers to ignore it. The middle is where most real invoices land, and where the agent earns its keep by pointing at the lines worth reading.

Every check is shown. Checks use the manager's own vocabulary and are grouped the way they already work: was the work completed, is the invoice internally consistent, do the on-site records agree, and what does this vendor's history say. The agent's reasoning is the product, and hiding it behind a verdict would have removed the one thing that built trust.

The certainty score stayed, and taught a lesson. Testing showed managers reading it as how likely the invoice was to be approved, not how confident the agent was. That lesson carried into Pricing Agent's invoice check, which lists the results and drops the score.

Path Forward

Four steps from simple to complex: simplify configurations to enable the feature, auto-run recommendations on submitted invoices, auto-approve low-risk work orders from reliable vendors below a cost threshold, and let experienced managers define their own checks and scope of automation.