Here is the uncomfortable secret of business AI: most disappointing AI projects were not failed by the AI. They were failed by the data it was pointed at – outdated documents, five conflicting versions of the price list, policies that exist only in someone’s inbox, and spreadsheets nobody can interpret.
An AI assistant is a mirror of your information. Point it at chaos and you get confidently-delivered chaos.
The good news: getting data ‘AI-ready’ is not a year-long transformation programme. It is a focused cleanup with a clear checklist – this one.
Key Takeaways
- AI quality is capped by data quality: an assistant grounded in outdated or conflicting documents gives outdated, conflicting answers – confidently.
- AI-ready data passes three tests: findable (digital, in known locations), current (one authoritative version), and owned (someone maintains it).
- The most damaging data problem is not missing documents – it is OUTDATED ones the AI can’t distinguish from live truth.
- Permissions are part of readiness: decide who may see what BEFORE connecting an assistant, or the AI becomes a leak with a chat interface.
- Start narrow: clean the 20-50 documents behind your first use case, not the whole company archive.
- A data inventory + a deprecation habit + named owners = 80% of readiness, achievable in about 30 days.
Why data readiness decides your AI project
Modern business AI works by retrieval: when someone asks a question, the system finds the relevant passages in your documents and data, and answers from them (the RAG pattern behind every serious ‘chat with your company knowledge’ tool). That design has a direct consequence: the answer can only be as good as the best passage the system can find.
If the current price list is a PDF on someone’s desktop while the shared drive holds three older versions, the AI will retrieve one of the three – fluently. If the leave policy changed last year but the old one was never removed, both exist as ‘truth’. The AI has no way to know which document is alive; the organisation has to tell it, by cleaning up.
The three tests of AI-ready data
Test 1 – Findable. Is the knowledge digital, text-readable, and in known locations? Paper manuals, scanned images without OCR, and knowledge that lives only in a veteran employee’s head all fail this test. You cannot retrieve what was never written down.
Test 2 – Current. For each important topic, is there ONE authoritative version, and are superseded versions removed or clearly marked? This is the test most companies fail – not for lack of documents, but for surplus of dead ones.
Test 3 – Owned. Does a named person maintain each critical source? Ownerless documents rot silently; six months after your AI launch, un-owned sources are stale sources, and the assistant’s reputation rots with them.
The outdated-document problem (your #1 enemy)
It deserves its own section because it causes the most damage. Missing information makes an AI say ‘I don’t know’ – annoying but honest. OUTDATED information makes it give a wrong answer with full confidence: last year’s price, the cancelled procedure, the superseded policy.
The fix is a deprecation habit, not a technology: when a document is replaced, the old version is deleted from the AI’s sources or moved to a clearly-excluded archive – in the same motion as publishing the new one. One sentence of process, most of your accuracy protected. Grounding plus citations (see the hallucination guide) then lets users verify the rest.
Permissions: readiness’s forgotten half
Connecting an assistant to ‘all our documents’ has a second failure mode that has nothing to do with accuracy: the AI answers a junior employee’s question using the salary file, the acquisition memo, or a customer’s private data – because retrieval doesn’t respect boundaries nobody defined.
Before connecting anything, classify sources into tiers – public, internal, restricted – and decide which tiers the assistant may read, and for whom. Permission-aware retrieval (finance documents answer only finance staff) is a solvable engineering problem, but only if the tiers exist. This is the same discipline as the board-level data questions: where data may travel, and who may see it through the AI.
Structured data counts too
Documents are half the story; the other half is your systems – the ERP, the CRM, the spreadsheets. AI-readiness there means: fields actually filled (a CRM where ‘industry’ is blank on 60% of records can’t answer by industry), consistent codes (one product ID per product, not three eras of numbering), and clean joins (customer names spelled one way, not five).
You do not need perfection – you need the specific tables behind your first use case to be trustworthy. Ask: if a sharp new analyst queried this data on day one, would they get the right answer? That is exactly the situation your AI is in, permanently – the foundation for everything in plain-language ERP reporting.
The 30-day readiness sprint
Forget boiling the ocean. The path that works is narrow and fast:
- Week 1 – Pick the use case and inventory its sources. One assistant, one audience (say: staff questions about HR policies, or support answers about products). List every document and table it would need – typically 20-50 items, not 5,000.
- Week 2 – Clean that set. Identify the authoritative version of each; delete or archive the dead ones; fill the two or three genuine gaps (the policy that was never written down).
- Week 3 – Assign owners and tiers. Every source gets a named maintainer and a permission tier. Write the one-line deprecation rule into the publishing process.
- Week 4 – Connect, test with real questions, measure. Fifty realistic questions; check answers and citations; fix the sources (not the AI) where answers are wrong.
The narrow start compounds: the next use case reuses the habits and half the sources, and each cleanup makes the underlying business documentation better even before AI touches it.
Signs you are NOT ready (and that’s fine – now you know)
- Nobody can say where the current version of a core policy lives.
- The answer to ‘who owns this document?’ is a shrug.
- Critical knowledge lives in chat threads and inboxes only.
- The same customer or product exists under multiple names/IDs across systems.
- Access has never been classified – everyone can technically open everything.
None of these are AI problems, which is exactly the point: they are the pre-existing conditions AI will amplify. The 30-day sprint above turns each into a checkbox – and a business whose knowledge is findable, current, and owned wins even on the days the AI is switched off.
Who should own data readiness?
The trap is assuming IT owns it. IT can build the retrieval, the tiers, and the connections – but the CONTENT decisions are business decisions: which version of the pricing policy is authoritative, whether the old procedure is truly dead, what a junior may see. The pattern that works is a pairing: one business owner per knowledge domain (HR owns HR sources, sales ops owns pricing) with IT owning the plumbing.
Make the ownership visible in the tool itself – each answer’s citation shows the source and its owner. When users spot a wrong answer, they know exactly who fixes the document, and the feedback loop that keeps AI accurate becomes part of normal work rather than a special project.
The hidden bonus of the cleanup
Here is what surprises teams after the 30-day sprint: the business improves before the AI even launches. New employees find the current policy on the first try.
Support stops quoting dead prices. The two departments that maintained rival versions of the same procedure finally reconcile them.
That is because ‘AI readiness’ is really just information hygiene with a deadline attached. Companies postponed it for years because nothing forced the issue – the AI project is simply the forcing function. Whatever your assistant ultimately delivers, findable-current-owned knowledge pays for the sprint on its own, which makes data readiness one of the rare preparations with zero regret: useful even if the AI plan changes completely.
The four levels of data readiness
It helps to know where you sit, because the right next step differs by level.
Level 1 – Scattered: knowledge lives in inboxes, chats, and people heads; there is no single source for anything. Level 2 – Stored but messy: documents are on shared drives, but with duplicates, dead versions, and no owners. Level 3 – Organised: each topic has one authoritative, owned source in a known location, and permissions are defined. Level 4 – AI-ready: level 3 plus clean structured data and a deprecation habit that keeps everything current.
Most businesses discover they are at level 1 or 2 – which is normal, not shameful. The value of naming the levels is that it turns a vague sense of not being ready into a concrete next action: get to the next level for the specific documents behind your first use case, not the whole company at once.
Answering the objections you will hear
This will take forever. Not if you scope it. The 30-day sprint targets only the 20-50 sources behind one use case, not the entire archive – and that narrow win funds the next.
It is IT job. IT builds the plumbing, but content decisions – which version is authoritative, whether the old procedure is dead – are business decisions. The pattern that works pairs a business owner per knowledge domain with IT on the technology.
Cannot the AI just figure it out? No – the AI reads what you give it and cannot tell live truth from a dead draft. That is precisely why readiness is your job, not the model.
We will clean up later, ship now. Shipping AI on messy data produces confident wrong answers that erode trust fast – and rebuilding trust after a bad launch costs far more than the sprint would have. Readiness first is the cheaper path, not the slower one.
Start this week: three questions
You do not need a strategy offsite to begin – you need three answers. First: what is the one use case where AI would help us most?
Name it concretely (staff questions about policies, support answers about products), because everything else scopes from it. Second: what documents and data would that use case need, and where do they live right now?
List them – it will be twenty to fifty items, and the gaps will be obvious immediately. Third: for each of those sources, who owns it and is it current?
The shrugs you get are your to-do list. Answer those three questions this week and you have converted a vague AI ambition into a concrete, finite readiness project – the difference between companies that talk about AI for a year and companies that quietly ship something useful in a month.
Readiness is not glamorous, but it is the whole game.
The businesses that win with AI are rarely the ones with the fanciest models – they are the ones whose information is findable, current, and owned. That advantage compounds quietly: every cleaned-up source makes the next AI use case cheaper, and makes the business better even on the days the AI is switched off.
Readiness is the unglamorous work that decides everything downstream. Start it now, start it narrow, and let it grow.
Frequently Asked Questions
What does ‘AI-ready data’ mean?
Data that passes three tests: findable (digital, text-readable, in known locations), current (one authoritative version per topic, with superseded versions removed or marked), and owned (a named person maintains each critical source). AI assistants retrieve and answer from your data as-is – readiness decides answer quality.
Why do outdated documents matter so much for AI?
Because missing information makes an AI honestly say ‘I don’t know’, but outdated information makes it answer WRONG with full confidence – last year’s prices, cancelled procedures, superseded policies. A one-line deprecation habit (retire the old version when publishing the new) prevents most of it.
Do we need to clean all our company data before using AI?
No – that instinct kills projects. Clean the 20-50 documents and the specific tables behind your FIRST use case in a focused 30-day sprint: inventory, deduplicate to authoritative versions, assign owners and permission tiers, then test with real questions. Each next use case reuses the habits and much of the work.
What about data permissions and AI?
Decide them before connecting anything. Classify sources into tiers (public, internal, restricted) and configure which tiers the assistant may read and for whom – otherwise retrieval happily answers a junior’s question from the salary file. Permission-aware retrieval is solvable engineering, but only if the tiers exist.
Our systems data is messy – can AI still help?
Yes, if you scope honestly. AI needs the specific fields and tables behind your use case to be trustworthy – filled fields, consistent IDs, clean joins – not company-wide perfection. Ask whether a sharp new analyst could get right answers from that data on day one; that is the AI’s permanent situation.
Want a data-readiness check before your AI project?
We assess your documents, systems and permissions against the use case – then build grounded AI on infrastructure you control, once the foundation is real.