Is our data ready for AI?
Probably ready enough to start, if the data for one job can be found, someone is allowed to use it, and a person can tell a right answer from a wrong one. It does not need to be clean, complete or in one system first: most of the tidying an AI project needs is done during the project, for that one job only.
Evidence
The messiest data I have made usable is the archive behind Home A Roam, a travel video channel: years of raw footage in folders, with no transcripts, no tags and no list of who appears where. Nothing was cleaned up before the project. HARchive organised it as it indexed it, on one laptop.
- 440 hours of raw, untagged video made searchable during the project
- 12.7 TB indexed on one laptop, with nothing uploaded
- 79% of visual test queries found the right moment first
Read the full story in the HARchive case study.
What "ready" means in practice
Data is the most common reason AI pilots stall. Among middle-market companies whose AI pilots had only moderate or limited success, 53% named data quality as a barrier to scaling and 47% named integration with their existing systems (RSM Middle Market AI Survey 2026). Neither means the data has to be perfect. It means it has to be good enough for the one job, and reachable from where the work is done.
- It can be found. You know which system, folder or inbox holds the data for the job, and it can be exported or read by a program.
- You may use it. Contracts, privacy rules and your own policies allow it to be used for this purpose, and you know who decides.
- A person can check the answer. Someone who does the job today can tell when the AI gets it right and when it gets it wrong.
- There is enough of it. A few dozen good examples of the job done well is often enough to start; thousands are not needed for most uses.
A quick self-check
Pick the one job you want AI to help with, and answer these for that job alone:
- Can you name the system or folder where its data lives?
- Could someone send you an export or a sample this week?
- Do you know who is allowed to say yes to using it?
- Could a person on the team mark ten AI answers right or wrong in an hour?
- Does the result need to go back into a system, and does that system have a way in?
Yes to the first four means you can start a pilot. A no to the last one is the integration question, and it is cheaper to answer before the build than after.
What can be fixed during the project
Most data problems are fixed as part of the work, not before it, and only for the job in hand:
- Inconsistent formats and duplicates. A program can normalise them as it reads them.
- Missing labels and tags. Models can add them, as HARchive added transcripts, captions and places, with a person confirming the ones that matter.
- Data spread across systems. The project pulls only what the job needs into one place.
What cannot be fixed by the project: data you do not have, data you are not allowed to use, and a job nobody can say what a good result looks like for. Those need a decision first, and finding them is part of an AI Audit.
Next step
Tell me the job you want AI to take on and where its data lives today, and I will tell you plainly whether it is ready to start, what would have to be fixed along the way, and what cannot wait.