The demo answered well and nobody measured how often
A pilot handled ten chosen examples. In production it meets the other layouts, the handwriting and the coffee stain, and nobody knows the error rate because nobody counted.
Services
Business analysis and AI
A supplier invoice arrives as a photograph, and someone reads it and types the total into the system. That is a real task with a measurable cost, and it is the kind of task a model can do — up to a point, with a person checking the ones it is unsure about. The same holds for sorting support messages and suggesting a reorder list. Most other AI proposals we see are a demo with a price: convincing in the meeting, unmeasured in the workflow. The work is telling the two apart before you pay.
Who it is for
- Teams retyping documents that already exist in a readable form somewhere.
- A support inbox where the same few questions arrive daily and someone sorts them by hand.
- Buyers making reorder decisions from memory and a walk through the warehouse.
Problems it solves
A rule broke on the second supplier
A macro or a spreadsheet formula worked for one invoice layout. The next supplier changed a column and the rule produced wrong numbers without saying so.
The data cannot leave
The simplest service sends every document to a server abroad. For invoices carrying customer names and prices, that is a legal question before it is a technical one.
Use cases
Supplier invoice extraction with a human check
Supplier, date, total and the VAT line read from the scan, with anything below a confidence threshold sent to a person instead of guessed. The person checks the exceptions, not the whole batch.
Sorting support messages
Five or six categories, each routed to the person who handles it, with the original message kept and the category correctable in one click. The correction is the training data for the next version.
Answers drawn from your own documents
Contracts, procedures and past quotes indexed so that a question returns the three relevant paragraphs with the document and page beside the answer. Only worth building when that corpus already exists and somebody keeps it current — a model asked to answer from an unmaintained folder produces a confident guess in a citation-shaped wrapper.
A reorder suggestion the buyer can edit
Recent movement, supplier lead time and current stock produce a suggested order per product. The buyer edits it, and the edit is recorded: a suggestion that is always overridden should be a rule.
The rule that replaces the model
If the decision is "orders above a threshold go to the manager", that is a rule: cheaper per event, explainable to an auditor, and correct every time. We write it and skip the model.
What is in scope
Fixed scope. Fixed price.
Agreed in writing before anything is built. You approve before we start.
- A short list of candidate tasks, with how often each is done and what it costs in hours
- A rule-based baseline any model has to beat before it is bought
- A check of the data: what is scanned, what is structured, what is missing
- Retrieval over the documents you already have, with the source shown beside the answer
- Where the model runs — an external service or a local server — and what that means for your data
- A working flow with the human check built in, not added afterwards
- Measurement on a labelled sample: error rate, cost of a wrong answer, and a decision to continue or stop
How it is delivered
- 01
Do the task by hand, once
A batch of real documents or messages, timed, with the exceptions noted. If the batch takes an hour a month, no model will pay for itself.
- 02
Try the rule first
Most tasks that look like AI are a rule with three exceptions. The rule takes an afternoon, costs nothing per call, and can be explained to an auditor.
- 03
Label a sample and measure
A set of examples with the right answer written by the person who does the work. The model is judged on that set, including the ugly cases: handwriting, stamps, a total written over another total.
- 04
Keep a person on the exceptions
The flow sends uncertain answers to a human and records the correction. The error rate is reviewed monthly; a model that degrades after a supplier changes its layout is caught by that review, not by a complaint.
FAQ
Questions about this service.
What owners usually ask before requesting a quote for it.
What does it cost per document?
An external model charges per call, and that price has to be compared with the typing it replaces plus the review time it adds. A task whose model cost approaches the manual cost does not get built; there is no saving to protect.
Can it run on a local server?
Yes for some models, with trade-offs: smaller models, slower answers, and hardware you own. It is the right answer when the documents cannot leave the country, and the wrong one when a rule would do.
What if the model is wrong and nobody notices?
That is the actual risk, and it is worse than a slow process: a wrong total in a ledger looks like a correct one. That is why the human check and the monthly error review are in scope, and why we will refuse a design where the model writes to the ledger with nobody confirming the exceptions.
Do we need a data team to keep it running?
No. The person who owns the task measures the flow by counting corrections for a week. What needs a specialist is the decision to retrain the model, and that is a conversation, not a department.
Start a project
Want this scoped for your business?
Send the brief. It comes back as a written scope and a fixed quote before anything is built.