Skip to content
AI10 min

Before AI, your data: what has to be ready for a model to be useful to you

By · co-founder and Chief Executive Officer of Hábil ·

How to prepare your data for artificial intelligence: five questions on source, quality, freshness and use, before you choose the tool.

In 1 minute

An AI connected to your company answers with the data it finds. If sales say one thing in the system and another in Finance's spreadsheet, the AI does not resolve the difference: it picks one, and states it with the same confidence it would state the right one. Before buying the tool, a simple test pays off: pick three questions leadership wants to ask the AI and answer them by hand. If two departments arrive at different answers, the work still missing is not AI work, it is data work. This article gives five questions to check it.

Imagine a company that sells through its online store and its branches. It is fictional, but it looks like many. Leadership has already seen what an AI does with documents and wants the same for its operation: to ask in plain language "how much did we sell yesterday in the north?", "which customer stopped buying?" or "how much does this customer owe us?", and get the answer in seconds.

The vendor runs a demo with sample data and everything works. Then they connect the AI to the real systems and the first question returns a number nobody recognizes. Sales says the north sold more. Finance says that figure includes returns. The systems team says the online store is loaded the next day. All three are right, and the AI had nothing to tell which definition applied.

This article is for whoever is going to approve an AI project on the company's data. It makes one point: the model inherits the mess in the data; before choosing the AI, each business question needs an agreed source for its data, with its date and its history. You do not need to know about models to check it; you need to know where each number comes from.

Why AI does not fix the data: it uses it

A language AI is very good at understanding a question and writing an answer. What it does not resolve on its own is which of two numbers is the right one, if nobody gave it the rules or told it which source prevails. When it is connected to the company's systems, it answers with what it finds there, and if what it finds is incomplete, duplicated or out of date, so is the answer, well written.

The AI risk management framework from NIST, the United States standards institute, lists it among the risks AI brings compared with traditional software: the data used to build a system may not properly represent the context in which it will be used, and data quality issues can affect how trustworthy the system is [1].

The same happens when the AI checks documents before answering, the technique known as RAG. The paper that proposed it in 2020 presents it precisely as a way to give a model access to an external memory, and names as open problems of models on their own giving the source of their answers and keeping what they know up to date [2]. RAG helps with that, because the AI checks external information before answering. But what it finds depends on what is there: if the current manual and the one from two years ago sit in the same folder, the search may bring back the old one or leave unclear which one applies. That is why it pays to mark the current version, organize the documents and check the important answers against their source.

So the useful question before an AI project is not which model to use. It is whether the data it will use can answer well, today, the questions it will be asked.

The three-question test

Before reviewing systems, pick three questions leadership wants to ask the AI. Make them concrete and everyday, like the ones in the example company. Then ask two different departments to answer them by hand, with the data they use today, separately.

If the answers match and each department can say where it got its number, the data for that question is on the right track. If they do not match, you have already found the missing work, and you found it before paying for the tool.

The team that knows the data today can run the test, within a short window that does not interrupt operations. And it leaves something useful for what comes next: the list of the data each question needs.

Five questions your data has to answer

For each piece of data on that list, check five things. The first four are quality questions that bring together the six dimensions of the UK government's data quality framework, taken from the professional association DAMA [3]. The fifth is different: if there is personal data, whether the intended use is allowed by Mexican law [4].

1. Is there a single source for each piece of data?

If "daily sales" exists in the online store's system, in the branches' system and in a spreadsheet someone puts together every Monday, the AI does not know which one prevails. The UK framework calls consistency the degree to which values that represent the same thing do not contradict each other [3].

What has to be decided is not technical: which is the official source of each piece of data and who answers for it. If that agreement does not exist, it pays to start there; that is the subject of another article in this series, Before you move a piece of data, name its owner.

2. Is it complete and free of duplicates?

A customer who appears three times with three different emails looks, to the AI, like three customers. One with no region assigned does not show up in "the north". The UK framework separates the two: completeness, that all the records are there along with what matters inside them, and uniqueness, that there is a single record for each thing they represent [3].

A quick way to start: measure how many customer records have empty the field the question needs, and look for possible duplicates with agreed rules —matches on name, email, tax ID and address— before confirming them, because a shared email or tax ID is not always a duplicate.

3. Is it up to date for the question it will be asked?

"How much did we sell yesterday?" needs yesterday's data. If the online store is loaded the next day, the morning's answer comes out incomplete even if every piece of data is correct. The UK framework calls timeliness the degree to which the data reflects the period it claims to represent and is up to date [3].

Not every question needs the same. Yesterday's sales call for yesterday's data; the quarter's margin can live with a weekly load. What matters is that each answer says what date its data is from.

4. Is it what it claims to be?

A delivery date earlier than the order date, a negative amount on a normal sale, a postal code with four digits. The UK framework calls validity the degree to which the data is in the expected range and format, and accuracy the degree to which it matches reality [3]. Validity is checked with rules. Accuracy is checked, when needed, against evidence of what happened in the operation; comparing against the source system helps find differences and trace where the problem started, but it does not prove that the source is right.

This is where something the same framework recommends helps: fix problems at the source, not in the copy [3]. If the date is wrong in the order system, fixing it only in the report leaves the error waiting for the next load.

5. Can it be used for this?

That a piece of data exists does not mean it can be used for anything. Mexico's Federal Law on the Protection of Personal Data Held by Private Parties, published in the Official Gazette on 20 March 2025, requires the data controller to observe, among others, the principles of purpose, proportionality and quality in the processing of personal data [4].

In practice, before connecting the AI to customer data it pays to ask: does the purpose for which it was collected cover this use? Does the AI need the full data or is part of it enough, such as the region instead of the address? Who can see the answer? Your legal team gives the answers, not the tool vendor; the project has to be able to comply with them.

A sheet for each dataset

The five answers are more useful if they are written down. In machine learning there is a well-known proposal for that: to accompany each dataset with a sheet, like the datasheet of an electronic component, stating what it was put together for, what it contains, how it was collected and what it is recommended for [5].

You do not need the full format to start. Who answers for each piece of data and under what rules is decided in data governance work; if that is not clear yet, Before you move a piece of data, name its owner explains how to agree on it. To prepare a question for the AI, one page per piece of data it will use is enough, with four fields:

A sheet for each dataset
FieldExample (daily sales)
Agreed sourceOrder system; the online store comes in through its own load
Up to date as ofClose of the previous day, loaded at 6:00
What it coversBranches and online store; returns go separately, marked as such
Intended useInternal reports and answers to leadership; no customer data

Invented data to illustrate the format.

With that page, when the AI answers, someone can say where the number came from and what date it is from. Without it, the meeting's argument repeats itself, now with the AI in the middle.

The signs that your data is not ready yet

  • Two departments come to the meeting with different numbers for the same question.
  • Nobody can say, without digging, which system a number on the dashboard comes from.
  • The important report is put together by hand and only one person knows how.
  • Errors are fixed in the spreadsheet and not in the system where they started.
  • Nothing is written down about which customer data can be used and for what.

These signs do not rule out using AI on their own: they show where to start.

Five questions for your next board meeting

  1. What are the three questions we want to ask the AI, and who answered them by hand?
  2. What is the official source of each piece of data they need, and who answers for it?
  3. What date is the data it would answer with, and is that enough for the question?
  4. What customer data would it use, for what purpose, and who reviewed it?
  5. If the AI gives a number nobody recognizes, how do we find out where it came from?

Where to start this week

Pick the three questions and ask two departments to answer them separately. Write down, for each number, which system it came from and what date it is from. If they match, you have a good first test for an AI pilot; if not, you have the list of what needs to be put in order first.

When it is time to test the AI, the article The AI pilot went well in the demo explains how to measure whether it works; and if the doubt is the technique, RAG or fine-tuning? explains which one fits each case.

What Hábil does

At Hábil we put the company's data in order and connect the AI to it, in that order.

If your numbers do not match across departments, the proposal is in Hábil Lake, your company's data platform. It brings together in one place the data from your systems, with its full history and already organized. On every load it reconciles it —it squares it— against the system it came from, so Leadership, Finance and Sales arrive at the meeting with the same number and anyone can see where it came from.

It stores the data in open formats and runs on your infrastructure or ours, with Hábil in charge of operations.

When the data already answers well, Hábil AI connects the AI to it: it answers with your real data, with permissions and a log, and checks your documents before answering. They are contracted separately: you can start with whichever solves today's problem.

The conversation can start with your three questions and what each department answered, even if they do not match. With that, we review together which data is missing, which needs to be put in order and whether it makes sense to start with the data or with the AI. Write to us on WhatsApp; to protect your information, do not share data about your customers or your systems through that channel: we will agree in the conversation on the channel for the case.

Request your first conversation, at no cost.

Consulted on 18 November 2026. Each reference carries its confidence: high (primary source read).

References

  1. NIST, "Artificial Intelligence Risk Management Framework (AI RMF 1.0)" (NIST AI 100-1, January 2023), appendix B, among the new or increased risks of AI compared with traditional software: "The data used for building an AI system may not be a true or appropriate representation of the context or intended use of the AI system" and "harmful bias and other data quality issues can affect AI system trustworthiness". It is for voluntary use. Opened live on 18-Nov-2026. Confidence: high. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
  2. Lewis, P. et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks" (2020), abstract: "providing provenance for their decisions and updating their world knowledge remain open research problems", and the proposal to combine the model with an external memory ("explicit non-parametric memory"). Opened live on 18-Nov-2026. Confidence: high. https://arxiv.org/abs/2005.11401
  3. Government Data Quality Hub (United Kingdom), "The Government Data Quality Framework", section on the six quality dimensions "as defined by DAMA UK": "Accuracy describes the degree to which data matches reality"; "Completeness describes the degree to which records are present"; "Uniqueness describes the degree to which there is no duplication in records"; "Consistency describes the degree to which values in a data set do not contradict other values representing the same entity"; "Timeliness describes the degree to which the data is an accurate reflection of the period that they represent, and that the data and its values are up to date"; "Validity describes the degree to which the data is in the range and format expected". It recommends addressing problems "at source". Opened live on 18-Nov-2026. Confidence: high. https://www.gov.uk/government/publications/the-government-data-quality-framework/the-government-data-quality-framework
  4. Chamber of Deputies (Mexico), "Ley Federal de Protección de Datos Personales en Posesión de los Particulares", "Nueva Ley publicada en el Diario Oficial de la Federación el 20 de marzo de 2025"; latest amendment published in the Official Gazette on 14-11-2025. Article 5: "El responsable deberá observar los principios de licitud, finalidad, lealtad, consentimiento, calidad, proporcionalidad, información y responsabilidad en el tratamiento de datos personales" (the controller must observe the principles of lawfulness, purpose, loyalty, consent, quality, proportionality, information and accountability). Opened live on 18-Nov-2026. Confidence: high. https://www.diputados.gob.mx/LeyesBiblio/pdf/LFPDPPP.pdf
  5. Gebru, T. et al., "Datasheets for Datasets" (2018), abstract: "we propose that every dataset be accompanied with a datasheet that documents its motivation, composition, collection process, recommended uses, and so on", by analogy with the datasheets of electronic components. Opened live on 18-Nov-2026. Confidence: high. https://arxiv.org/abs/1803.09010