We get called in after something has already stalled. A business bought a tool, ran a pilot, and the results were vague enough that nobody could argue for expanding it or killing it. Almost every time, the fix is upstream of the tool, in data nobody thought of as a project.
What it looks like from the inside
The story is consistent enough that we can usually guess it before the call. Somebody connected an assistant to the CRM to summarise accounts. The summaries came back plausible and slightly wrong. Not wrong enough to throw out, not right enough to forward to a client. After a few weeks people quietly stopped opening it.
The diagnosis in the room is usually that the model is not good enough. It is almost never the model. It is that the same customer exists three times, under three spellings, with a different owner on each record, and the model dutifully summarised all three as separate relationships. It did exactly what it was told against the data it was given.
The four things we find every time
The same entity, several times
Acme Co, Acme Company, and ACME. A human glances at those and sees one customer. Software sees three, and every count you build on top inherits the error. This is the single most common thing we find and it is usually years deep.
Meaning that lives in someone's head
There is a status field with eight options and everybody uses four. One of those four means something different depending on who set it, and everyone in the office knows that, and none of it is written down. When you point automation at that field it takes the label at face value, because the label is all there is.
The real record is somewhere else
The CRM says one thing. The actual state of the job is in a text thread, or a whiteboard, or the scheduler's memory. Everyone routes around the system because the system was never quite right, and the workaround has been load bearing for so long that nobody remembers it is a workaround.
Free text doing structured work
A notes field carrying the delivery date, the discount, and whether they are difficult about invoices. Enormously useful to the person who wrote it. Nearly useless to anything trying to compute with it.
Why this is worse with AI than with people
A new hire hits the same problems and self corrects. They see Acme three times, ask, learn it is one account, and carry that correction forward without being told twice.
AI does not do that. It will not flag the duplicate, because it does not know it is a duplicate. It will produce a fluent, confident answer built on the broken version, and it will produce it faster than a person would, at volume, in a tone that reads as authoritative. The speed that makes it valuable is the same speed that propagates the error.
Automation does not fix a messy process. It runs the messy process faster and with more conviction.
What to do first
None of this requires a platform, and you should resist buying one to solve it.
- Pick one entity. Customers, usually. Not customers and jobs and invoices. One.
- Find the duplicates. Sort by name, sort by phone, sort by domain. An hour of this on a small business database is usually enough to find most of it.
- Decide what the fields mean and write it down. One page. What each status means and who is allowed to set it. This page is worth more than most software you could buy.
- Find the shadow system. Ask where people actually check the real answer. Whatever they name is your real system, and it either needs to become the record or feed it.
- Pull the structured bits out of free text. If a date matters, it needs a date field.
How you know it worked
Pick a question the business asks constantly. How many active customers do we have. How many jobs are open past their due date. Answer it from the system, then verify it by hand.
If the two numbers match, your foundation is sound and automation on top will hold. If they do not, you now know the exact size of the gap, which is a far better position than discovering it later inside an AI generated report a client has already seen.
This is unglamorous work. It also determines whether everything you build afterwards is worth anything. In our experience it is two to four weeks of attention for a business of ten to a hundred people, and it is the highest return work available to most of the companies we talk to.
