Use casesBack-office workflow agents
The work that moves between systems by hand, done by agents that cannot cut corners.
Invoices matched to orders. New accounts set up across five systems. Forms read, checked and keyed in. This is where most of a business’s hours go, and it is the first thing buyers are asking agents to take. We build workflow agents on a deterministic spine: the model reads and proposes, code checks and commits, and nothing gets past a gate it has not earned.
What gets in the way
An invoice arrives as a PDF. Someone opens it, finds the purchase order, checks the amounts, keys it into the ERP and routes an exception when the numbers do not match. Multiply by a thousand a month, then by every process that looks like that.
Robotic automation broke on every layout change. Pure language models read anything and occasionally invent a line item. Neither is something a controller will sign off on.
This is the first thing most buyers ask agents to do, and the buyers who get burned are the ones who let the model commit the transaction.
What we build
Agents read and propose. Code checks and commits.
- Each agent has one job: read the document, find the matching record, propose the entry. It returns a strict structured record, never free text into a system of record.
- Plain code does everything that must be exact: matching by ID then by fuzzy name, three-way checks of amounts and quantities, tolerance rules your finance team sets.
- A gate between proposal and commit fails closed. Anything that misses a tolerance, lacks a source field or falls outside the agent’s remit goes to a person, with the reason attached.
- Every commit is written once with what the agent saw, so any entry can be audited against its evidence months later.
- The run stops itself when inputs look wrong: a batch that shrinks by a fifth, a vendor nobody has seen, a date in the future. It logs why and waits for a person.
How we know it works
We test it the way an auditor would: by trying to get a bad entry through.
- Before go-live, the agents replay months of real documents with the commits turned off, and every proposal is compared with what your team actually entered. The disagreements are the review list.
- We plant defects: a duplicated invoice, a shifted decimal, an amount split across two currencies. Each one must stop at the gate. If one gets through in testing, the test is the finding.
- This is the spine under our work for Lenovo, where two reviewer agents walk every card through the client’s own specification and fail closed, and under Winland Foods, where the bar lives outside the model so the business can move it any morning.
- What share of the volume clears is measured on your own documents at a tolerance you choose, before anything goes live. It is a number, not a promise.
Built from this work
No client runs an invoice or onboarding agent of ours under that name yet. It is built from the parts below, and we would rather say so than imply otherwise.
1,491 account cards built in September’s monthly run
24 public and purchased data sources, each with its own recognizer
188 accounts Lenovo approved and loaded in August and September
Read the story28% of claims clear automatically at the recommended bar
92% of cleared claims were real shortages, on months the model never saw
$174K in deductions cleared through the model at the first month-end close
Read the story
Where it fits
Manufacturing
AI for manufacturers, where the count has to be right.
Insurance
Insurance, from the question a customer asks to the systems that answer it.
Financial services
AI for financial services, built to pass an audit.
Questions
Which processes first?
The ones with a clear record to match against and a person already checking the result: invoice matching, order entry from email, account onboarding, claims intake. Anything where the right answer is knowable from your own systems.
What happens when a layout changes?
The agent reads, so a new layout is a new read, not a broken script. What it reads still has to pass the same checks, so a layout change can make the exception rate rise. The dashboard shows it the same day.
Can it commit to our ERP?
Only through code, through the same APIs your people use, with a limit on what it may touch. The model never holds a credential.
How much of the volume will it clear?
That is measured, not promised. The replay on your own documents tells you what share clears at a tolerance you choose, before anything goes live.
Bring us the project that matters most.
Tell us what you’re building and what’s in the way. You’ll hear back within 24 hours.