You're reading The AI Directive — practical AI intelligence for business leaders. Subscribe here if someone forwarded this.
THE SIGNAL
The Filing Cabinet Problem
There is a very old computing phrase that still does useful work:
Garbage in, garbage out.
It is not sophisticated. It is not fashionable. It is also still true.
But it is not the whole answer.
If the message is simply "AI is useless unless your data is perfect", most businesses may as well pack up now. No organisation has perfect data. Product information sits in one system, supplier terms in another, customer commitments in email, decisions in meeting notes, exceptions in Teams, and the real process in someone's head.
That is normal.
The leadership question is not whether the filing cabinet is messy.
It is whether you know it is messy, and whether the AI workflow has been designed accordingly.
That distinction matters.
AI will not magically turn poor source material into reliable judgement just because the output sounds confident. If you point it at stale documents, contradictory records and old assumptions, it can produce a very neat version of the wrong answer.
That is the familiar risk.
But there is a better opportunity as well.
Used deliberately, AI can help improve the filing cabinet.
It can compare product records across systems. It can spot duplicates. It can highlight conflicting dates, prices, owners or status fields. It can check whether the latest action log contradicts an old project note. It can classify messy records, suggest standard wording, flag missing fields and show where the underlying data is not good enough to support a decision.
That is not magic.
It is structured data-quality work, accelerated.
The difference is design.
If an AI assistant is asked, "What is the current position on this supplier?", and it is allowed to rummage through years of mixed-quality material with no hierarchy of sources, no date logic, no confidence flag and no route back to a system owner, the answer may be fluent but weak.
If the same assistant is told, "Use the latest supplier master record as the base, compare it with the latest order log and contract folder, flag conflicts, treat emails as supporting evidence only, and show anything that needs human confirmation", that is a very different proposition.
Same messy business.
Better operating model.
This is where the simplistic version of the argument fails.
The point is not that AI needs pristine data before it can add value.
The point is that AI needs to know what kind of data problem it is dealing with.
There is a world of difference between:
"Here is a pile of information. Tell me the answer."
and:
"Here are five imperfect sources. Reconcile them, show the conflicts, explain which source you trusted, and tell me what needs cleaning."
Most organisations already have the second problem.
Product data is spread across ERP, PIM, ecommerce, supplier spreadsheets and marketplace feeds. Customer data is duplicated across CRM, finance and service systems. HR data is partly structured and partly buried in documents. Finance processes depend on invoices, purchase orders, approvals and notes that do not always agree.
AI can be genuinely useful in that environment.
It can help build the map.
But it should not be allowed to pretend the map is the territory.
That is the executive point. Data quality still matters. Master data still matters. Ownership still matters. Source hierarchy still matters. Review and exception handling still matter.
AI does not remove those fundamentals.
It gives you a more powerful way to work through them, provided you are honest about the mess at the start.
That is why AI memory needs care.
Memory sounds obviously useful. If an assistant remembers the business, the decisions, the priorities and the way work gets done, surely it becomes more valuable.
Yes, if what it remembers is curated.
No, if what it remembers is just accumulated history.
Think about a filing cabinet.
If it is organised, labelled and reviewed, it is an asset. You can find the current contract, the approved policy, the signed-off process and the latest decision note.
If it is neglected, it becomes storage. Old policies, duplicate versions, superseded decisions, forgotten exceptions, draft files and notes from conversations that never became actions.
That is not useful memory.
It is context debt.
AI can help reduce that debt. It can profile the cabinet, identify weak spots and propose clean-up work. It can say: these two records appear to refer to the same customer; these product descriptions conflict; this supplier term is newer than the one in the old onboarding pack; this project decision appears to have been superseded by last week's action log.
That is valuable.
But it only becomes reliable when the workflow makes uncertainty visible.
The worst version of AI memory hides the uncertainty. It turns mixed-quality source material into confident prose.
The better version exposes the uncertainty. It tells you what it used, what conflicted, what it ignored, and what a person still needs to confirm.
That is the standard leaders should be asking for.
Not "can the AI remember everything?"
Not "is our data perfect?"
The better question is:
Can the AI distinguish between trusted facts, stale context, conflicting evidence and open questions?
That is a much more useful test.
It also turns data quality from a blocker into a design requirement.
For low-risk drafting, rough context may be enough.
For team knowledge work, there should be source ownership, review dates and a clear hierarchy of what counts as current.
For customer, supplier, finance, HR or compliance workflows, the bar is higher again. The AI should be able to reconcile sources, flag exceptions, preserve an audit trail and stop when the evidence is not good enough.
That is not bureaucracy.
It is how you stop a clever tool from becoming a very efficient rumour mill.
The old phrase still applies.
Garbage in, garbage out.
But the better version for AI is:
If you know the data is messy, design the work so the mess is visible.
Then use AI to help sort it.
FIELD NOTES
The Memory Review I Would Actually Run
Pick one AI workflow where continuity matters.
Not the whole company. One workflow.
Then ask five questions:
What should this AI treat as the current source of truth?
Which sources are useful but less reliable?
What conflicts should it flag rather than resolve silently?
What data should expire?
Who is allowed to correct the memory?
That review changes the conversation.
It stops memory being a vague feature and turns it into an operating control.
For a leadership assistant, the answer may be fairly simple: current priorities, live commitments, preferred formats and standing constraints.
For customer service, finance, HR, supplier management or product data, it needs more structure.
Not because AI cannot handle messy information.
Because it can handle messy information so confidently that the organisation needs to decide where confidence is justified.
THE SHORTLIST
1. Poor data quality does not make AI useless. It means the AI workflow needs to be designed to detect, reconcile and flag uncertainty.
2. AI memory is not the same as stored history. Useful memory is curated context: current decisions, approved sources, rules, risks, owners and review dates.
3. The real test is whether AI can show its working: what it trusted, what conflicted, what is stale and what needs human confirmation.
ASK ME ANYTHING
"Should we fix our data before using AI?"
— Not as a blanket rule
If the data problem is known, AI can be part of the fix.
It can help profile the mess, identify duplicates, standardise records, compare sources and highlight where the evidence is weak.
But that needs a structured process.
If you simply point AI at mixed-quality information and ask for the answer, you are relying on confidence rather than control.
The sensible approach is to use AI with source hierarchy, review points, exception handling and human confirmation where the risk justifies it.
That is slower than a demo.
It is also how the work survives contact with the real business.
ONE THING
Do not ask whether your data is perfect. Ask whether your AI workflow knows which data to trust, which data to challenge and when to stop.
FROM THE EDITOR
If you only do one thing this week, pick one AI use case and write down the source hierarchy.
What is trusted? What is supporting evidence? What is stale unless confirmed? What must be checked by a person?
That small exercise will tell you whether AI is improving the filing cabinet or just reading from it faster.
See you Tuesday.
— Toby