Businesses that are disappointed by AI usually blame the tool. In our experience the tool is rarely the problem. The problem is that it was pointed at eleven years of accumulated files and asked to be clever about them.
An AI assistant answers from what it can see. If what it can see is three versions of the price list, a policy that was superseded in 2024 and a folder of drafts nobody archived, it will answer confidently from the wrong one. Here are five checks that tell you where you stand.
Check 1: can a human find the current version?
Ask somebody who joined in the last year to find your current standard terms, the latest price list and the most recent version of your health and safety policy. Time them.
If it takes more than two minutes each, an AI tool will not do better. It will do the same search, find the same four candidates and pick one without telling you it had a choice. The difference is that the new starter would have asked a colleague.
This is the single highest-value check because the fix is small. Identify the twenty or thirty documents that answer the questions people actually ask, put them in one clearly named place and archive everything they superseded.
Check 2: are permissions what you think they are?
This is the check that generates the most uncomfortable meetings, and it is the one to run before you switch anything on.
Microsoft 365 Copilot respects your existing permissions precisely. That sounds reassuring until you remember what those permissions actually are in a business that has been running SharePoint and OneDrive for six years. A folder shared with everyone in the organisation in 2021 for one afternoon. A link somebody sent externally that was never revoked. A departing employee's OneDrive still shared with half the company.
None of that was a live risk while finding it required knowing it existed. It becomes a live risk the moment anybody can ask what are our current salary bands in plain English and get an answer.
Copilot does not create an oversharing problem. It ends the accidental obscurity that was hiding one.
Before rollout, run a sharing report across SharePoint and OneDrive, deal with anything shared organisation-wide or with anonymous links and check what departed staff still have open. Our security solutions team does this as a standard pre-Copilot exercise and it usually takes a day or two.
Check 3: how much duplication is there?
Duplication is what turns a good AI answer into an unreliable one. Six copies of a document with small differences between them means the tool has six candidate answers and no basis for choosing.
The usual sources are predictable: files emailed as attachments and saved locally, a shared drive that was migrated to SharePoint without cleaning, personal OneDrive copies made so somebody could work offline and folders named final, final v2 and final USE THIS ONE.
You do not need to deduplicate everything. Deduplicate the answer set — the documents that answer common questions — and accept that the archive stays messy.
Check 4: is the important data in documents or in systems?
This determines what AI can realistically do for you, and it is where expectations most often break.
Most AI assistants read unstructured data: documents, spreadsheets, email, meeting transcripts, PDFs. They do not read your accounting package, your job management system or your CRM database unless somebody has built a connection.
So if the questions your team wants answered are what did we quote this customer last time or which jobs are running over budget, no licence purchase will deliver that. Those answers live in systems, and getting AI to them is a system integration project. Worth doing, but a different project with a different budget.
Check 5: is there one source of truth per subject?
For each of the questions your business answers most often, there should be exactly one document that is definitively correct. Not one folder. One document.
Write the list out. Pricing. Standard terms. Delivery lead times. Warranty position. Holiday policy. Escalation contacts. For each one, name the file. If you cannot name a single file, that subject is not ready and any AI answer about it is a coin toss.
This exercise is worth doing even if you never buy an AI licence. It is the same list that makes a new starter productive in a fortnight instead of a quarter.
What to fix first
In order, because doing these out of sequence wastes effort:
- Permissions. Non-negotiable, and it must happen before rollout rather than after.
- The answer set. Twenty to thirty current documents in one obvious place.
- Archive the superseded. Moving old versions out beats deleting them and argues with nobody.
- One source of truth per subject. Written down as a list, so it can be maintained.
- Integrations. Last, because it is the expensive one and the first four often reveal that you need less of it than you thought.
None of this is glamorous and all of it pays back whether or not you go on to buy AI licences. If you would like a hand running the permissions review or reorganising the answer set, our managed IT team does both routinely for UK businesses. Ask us for a readiness review and we will tell you plainly what needs doing before you spend anything on licences.
