Your ERP has been collecting evidence for nine years

Two things are usually true at once in a Midlands manufacturing SME. The ERP or MRP system is older than some of the staff, and it contains a decade of information nobody has ever properly looked at.

When AI comes up, the reflex is that the system must be replaced first. That reflex is expensive and usually wrong.

What is already in there

Every job you have run. What you quoted and what it actually cost. Which customers pay late. Which suppliers miss dates, and by how much. Which parts get reworked. Which quotes you lost.

Most firms report on almost none of this, because the reporting module is unpleasant and the person who understood it has retired. The data is not missing. It is unread.

You do not need to migrate to read it

If the system runs on a database you can query, or can export to CSV on a schedule, you can extract a copy and work on that. Nightly is fine — almost nothing in this territory needs live data.

That separation matters for a practical reason: you are not touching production. The system that runs your business carries on untouched, and if the analysis is wrong or useless, nothing breaks.

The questions worth asking first

Start with questions where the answer changes a decision:

• Which jobs actually make money? Not which product lines — which jobs, by customer, quantity and complexity. Almost every firm finds a category it is losing money on and did not know.
• Where does the time really go? Quoted hours against booked hours, by job type, tells you where your estimating is systematically wrong.
• Which customers cost more to serve than they look? Late payment, change requests, expedites and returns rarely make it into the margin figure.
• What is your true on-time delivery? Measured against the original promised date, not the revised one.

None of these require AI in the fashionable sense. They require someone to read your data. AI lowers the cost of that reading dramatically, which is why it is now worth doing in a firm that could never justify an analyst.

Where AI adds something beyond reporting

Once the basics are visible, there are things a model does that a report does not:

Finding similar jobs. Matching a new enquiry to genuinely comparable past work, including from a drawing or description rather than a part number.

Reading the free text. Job notes, non-conformance reports and delivery comments are where the real reasons live, and they have never been analysable before.

Flagging the odd one. A job priced well outside the pattern for its type, a supplier whose lead times have quietly drifted.

The state of the data

It will be worse than you think, and that is normal. Part numbers entered three ways. Customers duplicated across records. A "misc" category absorbing 15% of everything.

Do not attempt to fix it all. Fix only what the first question needs. A cleanup programme with no question attached runs forever and delivers nothing, which is how firms end up concluding their data is unusable — it usually is not, it is just untidy in ways that only matter for some questions.

The honest sequence

Export, ask one question, get an answer you trust, act on it. Then ask the next. Each cycle takes weeks rather than months, and each one is independently worth doing — which is the opposite of a migration, where value arrives only at the end, if it arrives.