AI takes the typing off your hands, not the decision
Read data out of documents, sort the inbox before anyone opens it, find knowledge buried in old filing. If you prefer, all of it runs on a machine in your own server room.
Book an intro callIn a mid-sized company, AI replaces nobody in the back office today. It does not chase overdue payments; it negotiates nothing; and it makes no decision that somebody has to answer for. What it can do is narrower and far less spectacular: pull the notice period out of a maintenance contract, sort the morning inbox by urgency, or track down the one equipment description nobody can place, buried in filing that has grown for years. That is the work that eats the hours.
Our first step is therefore not about AI at all, but about the task that costs you the most time. Often it turns out that a folder with a few mailbox rules is enough and no model is needed. If a task is left over that is worth the effort, we build it as the first use case, connect it to Microsoft 365 or your line-of-business application and, if you want, put the model on a machine in your own building.
What is included.
Reading data out of documents
Invoices, delivery notes and contracts run through recognition that pulls out amounts, deadlines and contracting parties and hands them over to your accounts department. Fields we are unsure about are flagged for a person to check, rather than waved through quietly.
Sorting the inbox in advance
Incoming mail and form submissions are filed by topic and urgency and routed to the mailbox responsible for them. For enquiries that keep coming back, we prepare draft replies that a person reads, completes and sends.
Making knowledge findable
Your filing, manuals and project folders become searchable. Ask how handover was agreed for a particular site and you get the answer along with the file name and the page number, so you can read it in the original.
Running on your own hardware
If no document is to leave the building, the language model runs on a server at your premises. We size the machine, set it up and keep it current. The analysis stays inside your network.
Setting the boundaries
Before the first use, we agree which documents the model may see, who signs off on results and at which point a person countersigns. Personnel files and payroll data are usually left out.
Briefing at the desk
The people who work with it every day walk us through their process, and we show them what changes for them. That is usually when the small snag in the workflow surfaces, the one that still needs putting right.
How we work.
Sitting in on the daily routine
We sit down with the departments where things get stuck and watch: which papers get typed up by hand, what people search for, what has to be handled twice. Out of that comes a list of tasks with an estimated share of working time.
Choosing the first use case
From that list we take one case that comes up often and where a mistake does no real harm. You know in advance which model we use, where it does its computing and which data it gets to see.
Build and trial run
The case goes live, supervised at first. For a few weeks the old way of working carries on alongside it, so you can put both results side by side. If the case does not carry its weight, we stop it and say so plainly.
Support afterwards
After that, you still have a point of contact. Models age, forms change, a supplier switches its invoice layout. We adjust, and we add the next case if you want it. Access credentials, configuration and documentation stay with you.
Common questions.
My people have been using ChatGPT in the browser for ages. What do I need you for?
For a lot of things that really is enough, and if so you should carry on doing it that way. The difference starts where the model has to reach your own data: the contract archive, the mailbox, the line-of-business application. That needs a connection, properly governed access rights and a storage location where nobody pastes customer data by accident. That is the piece we build.
I have heard of AI projects that fizzled out. What if that happens to us?
Then you have lost one use case, not a programme. That is why we start with a single task and let the existing way of working run alongside it for a while. If the automation brings nothing, we switch it off. What is left is the effort spent on that one case, not on a platform you keep paying for over years.
I do not want to upload our contracts to an American provider. Is that still possible?
Yes, in that case the model runs on a server in your own building. Open models are now good enough for analysis, summarising and searching your own holdings, although they do need a graphics card and power. Where the demands on the result are higher, a service in a European data centre is often the more sensible choice. We set out both routes side by side.
What happens if the thing misreads an invoice and the payment goes out?
Sign-off stays with your accounts department, not with the model. Recognised amounts and bank details appear on screen for confirmation before anything is posted, and uncertain fields are flagged. Models do misread, particularly with skewed scans and unfamiliar layouts. So the only things that go through without a person looking at them are those that hurt nobody if they go wrong, filing by supplier, for example.
If you set this up for us, am I then tied to you for good?
No, because everything that matters sits with you. The documents, the server, the access credentials and the configuration belong to you, the steps are clearly documented, and we build on widely used components rather than a platform of our own. Another provider can take it over. It will not run entirely without support, though, because formats and models change.
Show us the job nobody wants to do
Name two or three processes that cost a lot of typing. In the conversation we work out which of them suits a first use case, what technology it would take and whether it can run on your own hardware.
Book an intro call