Six-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 3, 4, 5 and 6 need a person. Step 4 is waiting for approval right now. Click it to open the full walkthrough.
Estimates are directional and based on stated assumptions. All names, organizations, and identifying details have been anonymized in accordance with our confidentiality agreements.
How a Financial Services Firm Built Straight-Through Invoice Processing with Exception Routing
Reads every invoice that arrives, checks it against what we ordered and what we actually received, and hands a person only the ones it cannot settle. Written for a financial services company where the work was never hard, only endless: 15,000 invoices a month across four business units, matched by three people who spent their days keying fields rather than talking to vendors. Six stations, run as a loop: what station 6 learns from the corrections changes what station 2 clears on its own.
What Was Broken
What We Built
Six stations and 13 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Invoice Reading
Nobody is asked anything hereSubagents pull every invoice off every channel it arrives on, read the document rather than the filename, and normalise it into the same shape whatever format it came as. Nothing waits on a person here. An invoice reaches a reviewer only when station 3 cannot settle where it belongs.
Record Matching
Nobody is asked anything hereSubagents compare every invoice against its purchase order, the receiving record and the contract price, and score how far the three sit apart. A person sees a score with the three records behind it, not a verdict that arrived on its own.
Variance Routing
A person answers hereSubagents read the size of each variance against the approval matrix, send it to the person whose authority actually covers it, and assemble the three records into one view before it lands. A person opens one screen and sees the delta already calculated, rather than three systems and a calculator.
Payment Approving
A person answers hereSubagents read every variance back against the contract it sits under, draft each payment at the amount the contract supports, and stop. This is the station where a human works. Everything else exists to make this decision small and well-lit.
Vendor Watching
A person answers hereSubagents count every exception by vendor and reason, separate the ones we cause from the ones the vendor causes, and name the single change that would remove the most of them. A person has the vendor conversation. The pipeline only says which conversation is worth having.
Rule Learning
A person answers hereSubagents read every reviewer decision back against the match that produced it, find the threshold that would have cleared it correctly, and propose the change. A person accepts or rejects each proposed change. Nothing about what clears on its own changes without somebody signing for it.
Where a Person Decides
Operating Model
This changes how work flows through the team.
What Transfers, What Must Be True
Failure Modes
What breaks this pattern:
An invoice arrives with no purchase order behind it, and the pipeline finds it a match anyway. That is the oldest fraud in accounts payable: a bill for goods nobody ordered, paid because a system wanted every invoice to fit somewhere.
A fifty thousand dollar variance lands with a reviewer whose limit is five thousand. They approve it because the system chose them. The person with the authority to say no never sees the case.
The invoice and the receipt disagree, and the pipeline suggests paying the smaller number. That feels safe, but it is a money decision made by a reader of records. The reviewer signs a figure no document supports, and nobody researches why the records split.
A reviewer fixes one invoice, and the pipeline proposes a new rule from it. One correction can be a typo, a one-off vendor problem, or a bad day. A rule built on it changes how every future invoice clears, on evidence that would convince no one.
Directional Outcomes
What the agent counts, and the station that counts it.
These counts are the tallies from one monitored run of the agents. They are not monthly or annual totals.
What Runs Where
Every step names the subagent that does the work, the record it writes, the thing that raises a question for a person, and what it is allowed to touch. This is drawn from the source, not from a diagram somebody kept in sync by hand.
arrival-watchap/inbound/<inv-id>.jsonreject-unreadableap/**, vendors/** read-onlythree-wayap/matched/<inv-id>.jsonno-purchase-orderap/**, erp/** read-onlyroute-varianceap/queue/<inv-id>.jsonabove-reviewer-limitap/queue/**, erp/** read-onlypayment-briefap/decisions/<inv-id>.jsonawait-human-approvalap/decisions/**, everything else read-onlyvendor-readanalytics/vendors/<period>.jsonpropose-vendor-actionanalytics/**, vendors/** read-onlyrule-readanalytics/rules/<period>.jsonpropose-threshold-changeanalytics/**, rules/** read-onlyStack
Every system this agent reads or writes.