A first pass, not a final answer
The useful way to think about a language model in a business process is as a fast, tireless junior who has read everything and is occasionally, confidently wrong. You would not let that person send invoices unsupervised. You would let them do the first read of a fifty-page document and tell you what matters.
That is where AI earns its place: work where producing a draft is slow and checking a draft is quick. Most of the value sits in that gap.
Where it tends to be worth it
- Reading long documents. Tender packs, contracts, specifications. For a construction and M&E group we built AI-assisted tender analysis, so an estimator can get through a document set that used to take days. The estimator still makes the call; they just start from a summary instead of page one.
- Pulling structure out of paperwork. Invoices, statements and catalogues arrive as PDFs and get retyped into a system. At an accountancy practice, two document extraction services now take the first pass at invoices and catalogue data, because a meaningful share of the admin was people retyping information that already existed.
- Triage. Sorting incoming requests, emails or tickets into the right queue with a suggested priority, for a person to confirm.
- Drafting. First versions of replies, summaries and descriptions that somebody edits rather than writes from scratch.
Where it is not
If the task can be described as a rule, a rule should do it. Rules are cheaper,
faster, and give the same answer every time. A lot of what gets pitched as an
AI project is ordinary process automation with
a model bolted on where an if statement would have done.
It is also the wrong tool anywhere a mistake cannot be caught before it matters: payments going out, legal commitments, anything sent to a customer with nobody reading it first. And it is no substitute for having the data in order. A model will not fix a process that nobody can describe.
How it goes in
Every AI step we build follows the same rules:
- A person reviews before it counts. The output lands as a suggestion, a draft or a pre-filled form, and somebody accepts, corrects or rejects it.
- It shows its working. Where it extracted a figure from a document, the reviewer can see where; where it summarised, the source is one click away.
- It is logged. What went in, what came out, who approved it, and what they changed. That record is how you find out whether it is actually helping.
- The system works without it. If the model is unavailable, slow or switched off, the work falls back to the manual route rather than stopping. The AI is a step in the process, not the process.
Corrections are the most valuable data you will collect. They tell you where the model is reliable and where it is not, and they are what you look at before deciding to trust any step a little further.
Data and suppliers
Sending business documents to an AI provider is a data protection decision as much as a technical one. The scope sets out which provider and model are used, what data is sent to them, what is kept and for how long, and what never leaves your own systems. Those are your decisions to make, with the trade-offs laid out in plain English, and they are recorded before anything is built.
Scope and price
AI work is scoped and priced like everything else we do: fixed, in writing, from £4,000. Where it is not yet clear whether a model can do a task well enough, the first phase tests it on a sample of your real documents, so you find out before paying for the whole system. The running cost of the model itself is separate, and we estimate it in the scope. If you are writing the request up, the brief guide covers what helps.