Starting on AI with the data you already have

The data that has to be ready before you start on AI is the data behind the first job you give it. Most companies tidy far more than that first. Last year the UK government used AI to sort more than 50,000 written responses to a review of the water industry into themes. That was free text, written responses to a public call for views, and by the government's own assessment the tool was at least as accurate as human reviewers. If text like that was good enough, the emails, documents and forms in your business may well be good enough for a first job too. Forecasting and personal data are different: there, the data work comes first, but only for that one job.

Why it feels as if the data has to come first

Most of what you read says company data is not ready for AI. One survey, by the data software firm Precisely with Drexel University, found that only 12% of people who manage company data said it was good enough for AI. Surveys like this ask about all of a company's data at once, and they are often run by firms that sell the fix. That matters because no first AI job touches all of your data, so the answer to that question tells you little about where to start.

The worry is fair where information is spread across systems that do not agree with each other. That is a reason to keep the first job narrow, so that the mess you sort out is only the mess that job touches.

Ready means ready for a particular job

Gartner, the research firm, puts it simply: "The readiness of data for AI depends on how the data will be used." An inbox of loosely worded customer enquiries can be good enough for sorting those enquiries, even though it would be no use for a forecast. So you cannot say whether your data is ready until you have chosen the job.

A company-wide data clean-up with no job attached also tends to stall. Gartner's Saul Judah says data rules and tidying that are not tied to a business goal fail: a programme "that does not enable prioritized business outcomes fails". Choosing the job first gives the data work a purpose, which is what keeps it alive.

Emails and documents are often ready now

Jobs built on text are where a first project can usually start. Made Smarter, the government-backed programme for manufacturers, advises firms to "start using what they already have to build confidence and improve visibility, then strengthen data quality as adoption progresses". You can start before the data is perfect.

One of its case studies is Arden Dies, a die maker in Stockport. Many of its customer orders arrived by email in no fixed format, and staff retyped them into the business systems by hand. It tested AI tools that read those emails and pick out the customer, product, quantity and delivery requirements. The tools were tried out separately rather than run live, so there are no live results yet. The point for you is that the data this job needed was already there, in the inbox.

Forecasts and personal data need the work first

Some jobs do need the data sorted before anything else. Made Smarter says jobs like predicting when a machine will fail, or planning production, need better data before anyone can rely on them. A forecast learns from your history, so if the history is patchy or inconsistent, the forecast will be too.

Personal data adds a legal step. The ICO, the data protection regulator, says most uses of AI on personal data count as high risk, so the law requires a data protection impact assessment first. That assessment is made for a particular use of data, so it can be done for the one job you start with.

The ICO also says you cannot collect personal data just because it might be useful one day. So cleaning everything first, in case AI needs it later, is slow and hard to justify.

What to do

Start now, with one job. Pick it before you spend anything on data, and if you can, pick one that runs on emails, documents or forms. Look at the data that job touches, and fix only what gets in its way. If the job forecasts or uses personal data, do that data work and the impact assessment first, for that job alone.

Each job you do this way leaves some data in better shape, so the second job starts further ahead than the first.