AI readiness means something simple: your company's data is in a fit state to answer a specific question using artificial intelligence, and to do it in a way you can trust. It is not a certification and it is not about tools. It is a check, question by question, of six things. This guide explains them one by one, gives you a grid to assess yourself, and shows with an example how you get to a roadmap.
The six dimensions: why having the data is not enough
Almost every company says it has data. The problem is almost always something else: the data exists, but not in the shape that question needs. AI doesn't solve that problem, it amplifies it. A well-trained model working on inconsistent data produces answers that are tidy and wrong.
An example with made-up numbers. A company wants to know which customers are at risk of not buying again. Its ERP holds 4,200 customers and its online store 3,900. Matching them by email address, 3,100 line up, which is 74% of the ERP's customers (3,100 ÷ 4,200). The rest are duplicates, people with different emails or records with no email at all. Before any model, the question has no reliable answer, because one customer in four is either invisible or counted twice.
That is why the assessment runs on six dimensions, always tied to a concrete question:
- Access: can you reach the data without asking for favors or running manual exports?
- Quality: is the data complete, free of duplicates and consistent over time?
- Definitions: does everyone mean the same thing by "customer", "sale" and "order"?
- Integration: can the sources be linked through a common identifier?
- Ownership: does one person answer for each source and each definition?
- Goal: is the business question clear, with a decision that depends on it?
One check question and one warning sign for each dimension
| Dimension | Check question | Warning sign |
|---|---|---|
| Access | Who can get this data today, and how long does it take? | "We have to ask Marco to export it", or files travelling by email |
| Quality | If I pick 50 rows at random, how many have empty fields, duplicates or impossible values? | The same customer written three ways, missing dates, negative amounts with no explanation |
| Definitions | Do two people, asked separately, give the same definition of a "valid order"? | Two reports with the same name and different totals |
| Integration | Is there a field that appears in two systems and identifies the same thing (customer, product, order)? | Joining two sources means working by hand in a spreadsheet, every time |
| Ownership | If a number looks wrong, do I know who to call? | Nobody feels they own that data, or three people do |
| Goal | Which decision changes depending on the answer? | "Let's see what comes out", with no concrete choice behind it |
A self-assessment grid you can fill in
Take a single business question, for example "which customers will buy again in the next 90 days". For each dimension give a score from 0 to 2, referring to that question and that data.
| Dimension | 0 | 1 | 2 |
|---|---|---|---|
| Access | You have to ask someone every time | Repeatable manual exports | Direct or automatic access |
| Quality | Frequent, unmeasured errors | Known errors, partly corrected | Regular checks, rare errors |
| Definitions | Everyone has their own | Written down, but not used by all | Written, shared and applied |
| Integration | No common identifier | Partial or unstable identifier | Stable identifier across sources |
| Ownership | No owner | Informal owner | Named owner for each source |
| Goal | No linked decision | Vague decision | Decision, person and date |
Add up the six scores: the total runs from 0 to 12. Here is how to read it, bearing in mind that it is an orientation tool and not a scientific measure:
- 0-4: groundwork on the data is needed before any AI project.
- 5-8: you can start small, but fix the lowest dimensions first.
- 9-12: the question is in good shape to be tackled.
One rule matters more than the total: a dimension at 0 blocks the question, even if the others are at 2. Missing integration can't be offset by an excellent goal.
A worked example: three questions, only one ready
A made-up company sells office furniture online and to resellers. It has three business questions and scores them with the grid.
| Question | Acc. | Qual. | Def. | Integr. | Own. | Goal | Total |
|---|---|---|---|---|---|---|---|
| Which products have the most returns? | 2 | 2 | 1 | 2 | 1 | 2 | 10 |
| Which customers will buy again within 90 days? | 2 | 1 | 1 | 0 | 1 | 1 | 6 |
| What does advertising return by channel? | 1 | 1 | 0 | 0 | 0 | 2 | 4 |
The first is ready: orders and returns sit in the same system, with stable product codes. It can start right away, with a dashboard or with AI. The second is not: customers are duplicated between the ERP and the store and there is no common identifier (the 74% from the example above). The third is furthest away: advertising spend lives in different spreadsheets, nobody has defined what a "campaign" is, and no field links a campaign to the orders it generated.
The roadmap that comes out of it, in order:
- Now: answer the returns question.
- Next: build a single customer record with a shared identifier, and name who owns it. Once that is done, the repeat-purchase question moves from 6 to a much higher score.
- Last: unify the spend sources, write down the definitions and link campaigns to orders.
The useful point: nobody had to pick a model. The roadmap comes from the lowest scores.
Typical mistakes
- Starting from the tool. If you pick the technology first, you bend the question to fit what it can do, not the other way round.
- Assessing "the company's data" as a whole. AI readiness makes sense per question: the same data can be ready for one and unusable for another.
- Cleaning everything before starting. You clean what the chosen question needs. The rest can wait.
- Forgetting definitions. It is the least visible dimension and the most expensive: the AI will use a definition, and if it isn't the company's, the answers will look right and won't be.
- Leaving data without an owner. Without a name, nobody fixes the error when it surfaces.
A plan for the first week
- Monday: pick one business question and write down the decision that depends on it.
- Tuesday: list the sources needed to answer it and who owns each one.
- Wednesday: pull 50 random rows per source and count errors, blanks and duplicates.
- Thursday: check whether a common field exists across sources and write down the definitions of the terms in the question.
- Friday: fill in the grid, identify the two lowest dimensions and assign a person and a date to each.
What AI readiness is NOT
- It is not buying AI software. It is working out whether your data can support a question.
- It is not a grade for the company. It is a score for a question, and it changes with the question.
- It is not a years-long project. A first assessment can be done in a week, as above.
- It is not only a technical matter. Definitions, ownership and goal are business decisions, not IT ones.
When a ready-made tool is enough and when you need a custom path
If your question concerns a single system that is already supported, for example the sales of a Shopify store, a ready-made tool is usually the quickest choice. The integration and definitions dimensions are already handled inside the tool.
A custom path is needed when questions cut across several sources, or when the sources are unusual: internal ERPs, company databases, systems with no ready-made connectors. There, integration, definitions and data modeling are the real work, and they depend on how that particular company operates.
Where Opalca Lab fits
Opalca Lab is the part of Opalca devoted to custom projects: data source integration, API connectors and custom databases, custom dashboards, data visualization and data modeling. There is no standard price list: each project is assessed according to the company's data and questions.
One example is Fairflai, a multidisciplinary team of AI experts. It has a framework called AI Readiness Assessment and, with Opalca Lab, it put in place custom reporting and a tailored roadmap. More cases are on the Cases page.
Have a project in mind, or want to know where your data stands? Talk to Opalca Lab or write to us.
Keep reading
- Data integration: what you need before building a dashboard
- Ecommerce KPIs: the 10 metrics that matter and how to read them
Frequently asked questions
What does AI readiness mean?
It means a company's data is in a fit state to answer a specific question with AI, reliably. It is assessed on six dimensions: access, quality, definitions, integration, ownership and goal.
How do you assess the AI readiness of your data?
You pick a business question and score each of the six dimensions. The grid in this article is an orientation tool, not a scientific measure: it helps you see where to start.
How long does a first assessment take?
A first assessment of a single question can be done in a week by following the plan in this article. The fixing work that follows depends on the lowest scores.
Do I have to fix all my data before using AI?
No. You start from one question and fix the data that question needs. A ready question can start now, while the others wait.
When do I need a custom project?
When questions cut across several sources, or the sources have no ready-made connectors, such as internal ERPs and databases. For a single supported source, a ready-made tool may be enough.