How to organise your data before using AI
What needs tidying in company data before applying AI, and why skipping that step produces answers that are confident and wrong.
The sentence you hear most after a disappointing AI pilot is that the tool gave wrong answers. Most of the time it gave answers consistent with the data it received, and it was the data that failed to describe the company's reality. That distinction matters, because it leads somewhere different: the fix is not changing tools, it is tidying what the tool reads.
This piece lists what needs tidying, in the order worth tidying it.
Why the error here is worse than usual
An ordinary system fed bad data breaks, or shows an odd number somebody notices. A language model fed bad data produces a fluent, plausible answer. The tool does not flag uncertainty the way a report would, and the error slips through.
Bad data in a report becomes a strange number and somebody questions it. Bad data in an AI becomes a convincing sentence and nobody questions it.
The five fixes, in order of return
1. Single identity
The same customer, product or supplier needs one identifier. If it exists as "Silva Trading", "Silva Trading Ltd" and "Trading Silva", any analysis splits what should be together, and any assistant answers with half the picture.
It is the most tedious fix and the one that unlocks the most.
2. Agreed meaning
What counts as a sale, which date applies, what is in and what is out. Without that agreement, two departments ask the same question and get different answers, both correct within their own logic.
Write it on one page. It does not need more.
3. A declared source of truth
For each important piece of information, one place that is authoritative. If a customer address exists in the ERP and in the CRM with different values, decide which wins. Without that, AI will cite whichever it finds first.
4. History that records what did not happen
The most forgotten one. A sale lost for lack of stock, a rejected proposal, a customer who asked for something the company does not do. Those events appear in no system, and their absence distorts any analysis: the history shows what the company managed to do, not what the market asked for.
5. Accessible text
Contracts, procedures, policies, proposals. If they sit in scattered folders, in email and on somebody's laptop, they do not exist for practical purposes. Gathering them somewhere organised is usually the step that turns into results fastest.
What does not have to be perfect
Worth saying, because a demand for perfection paralyses: you do not need a corporate data quality programme before starting. You need one clean slice.
The practical path is the opposite of the big project:
- Pick one concrete use case.
- Work out which data it uses. Usually very little.
- Tidy only that.
- Run it, measure, learn.
- Repeat with the next case.
Companies that try to tidy everything before using anything finish neither.
How to know when it is good enough
A simple test: ask the AI five questions whose answers you already know for certain. If it gets all five right, the data in that slice will serve. If it gets one wrong, check whether the error belongs to the model or to the data, and it is almost always the data.
Repeat that test every time you widen the scope. It is cheap and it is the only honest way to know where you stand.
Frequently asked questions
Do I need a data lake or something like it?
For a mid-sized company, almost never. The gain is in single identity, agreed meaning and organised text, and none of those requires new infrastructure.
What about personal data?
Before putting any base containing personal data into a tool, define the legal basis, the purpose and the retention period, as Brazil's LGPD requires. That does not prevent use, it organises use.
How long does it take?
For one slice, days. For the whole company, it is a project that never ends, which is exactly why you should not wait for it.
Where this sits in the method
Organising data is the same discipline as organising process: deciding what each thing means, who owns it and where it lives. It is not a technical stage, it is a management stage, and that is why it cannot be delegated entirely to a vendor.
If you want to map which slice to tidy first in your operation, that is where a strategic diagnostic starts.
Sources
Apply
Want to apply this in your business?
A complimentary conversation to understand your context and decide, together, whether it makes sense to move forward.
IA no suporte interno: o help desk da empresa
Como usar IA para responder as dúvidas internas que consomem o tempo de quem sabe, e por que o ganho aparece antes no RH que na TI.
IA no estoque: previsão e reposição
Onde a IA ajuda na gestão de estoque, por que previsão sem cadastro limpo não funciona e o que resolver antes de tentar prever demanda.
Agente de IA próprio ou plataforma pronta
Quando vale usar uma plataforma pronta de agente de IA e quando vale construir o seu, com os critérios que decidem a escolha.