The hard part of AI in a company is deciding who may see which document.
The model is the easy half. A company with five departments already has documents that must not cross between them. It has a system of record people argue with, and a policy nobody has updated since March.
An assistant dropped into that will happily quote the March policy to a customer, or show a salaried employee somebody else’s appraisal. Which is why this kind of work is mostly about permissions, sources and who checks the output.
Enterprise AI means putting a model to work inside an organisation that already has systems, rules and people who must not see each other’s records. The model is a small part of it. The work is deciding which documents it may read for each person, who reviews what it produces, and what is written down afterwards.
Where it earns its place.
Where the difficulty actually sits
Every project of this kind meets these four before it meets a technical problem.
- Permissions: the assistant must see exactly what the person asking may see
- Sources: which document is the current one when four versions exist
- Review: which answers may go out, and which need a human first
- Record: what was asked, what was answered, and on what basis
What we put in first
Never the whole company at once. One department, one question they answer too often.
- An internal assistant over your policies, contracts and circulars
- Support drafts an agent edits rather than sends
- Extraction from documents your back office retypes today
- Classification of incoming enquiries into your own categories
- Summaries of long threads for whoever has to decide
The rules it runs under
Agreed in writing before the first screen is built, because retrofitting them is much harder.
- Access mirrored from your existing roles, not maintained separately
- Every answer showing the documents it came from
- Questions and answers logged so a dispute can be traced
- A stated position on what leaves your network and what does not
- Departments whose records are excluded outright
Where the data cannot leave your premises, models that run on your own hardware are an option. They ask more of the infrastructure and give up some quality, and we will say which way we would go.
How a rollout goes
- One department
- Real questions
- Review step
- Widen
- Retire the review
- Start where somebody can tell you the answer was wrong
- Keep a person between the answer and the customer at first
- Loosen that only when the failures are understood
- Measure by the work removed, not by the number of questions asked
The part nobody mentions
An assistant is only as current as what it reads. Most companies discover their documents are in worse order than they thought.
- Three versions of the leave policy in three folders
- Circulars that contradict the handbook
- Prices in a spreadsheet nobody owns
- Documents scanned as images with no text in them
Sorting this out is real work and it belongs to you, not to the model. It is also useful whether or not the AI project goes ahead.
When this is the right choice.
- Several departments, each with records the others should not read
- Policies and contracts spread across Drive, email and a shared folder
- A team answering the same internal question every week
- An existing ERP or CRM holding data people cannot get answers out of
- An auditor or a client who will ask how a decision was reached
When it is not.
- A small office where everything sits in one folder and a search box would do
- Documents that are out of date, because the assistant will quote them with total confidence
- A decision that has to be defensible, where a person must remain the one who made it
- A process that is broken, since the same wrong answers will simply arrive faster
- An organisation where nobody will own which document is the current one
Questions we are asked about it.
Will our documents be used to train somebody else’s model?
Not under the business terms of the main providers, which exclude your content from training. We put the provider, the terms and what they cover in writing at the start, because the answer depends on which service and which plan you are on.
Can it run on our own servers?
Yes, with open models on hardware you control. The trade is real: you take on the infrastructure and the answers are usually a little weaker than the largest hosted models. Worth it when the data genuinely cannot leave.
How do we stop it showing someone what they should not see?
By filtering before the model reads anything. The search runs as the person asking, so a document outside their access is never retrieved and never reaches the model. Instructing the model to keep a secret is not a control.
What happens when it gets something wrong?
It will, so the design assumes it. Answers cite their source, so a person can check. A reviewer sits between the system and the customer while trust is being earned, and wrong answers are collected and used to fix retrieval.
Where should we start?
With one team and one question they are tired of answering. Something with a right answer that somebody in the room can confirm. Starting with a company-wide assistant is how these projects end up quietly unused.
Do we need to move to the cloud first?
Not necessarily. What matters is whether your documents and records can be read programmatically, which is a different question from where they are hosted. A readiness assessment answers it before anyone commits to a platform.
Services that use it.
What it sits with.
Not sure Enterprise AI is the right choice?
Tell us what the software has to do and who opens it. If something else fits better, we will say so, and say why.
- No obligation
- We reply the same working day
- Your details stay private