Your Data Is the Hard Part — And It's Messier Than You Think

The demo always works. That's the thing about demos.

Someone shows you an AI answering customer questions flawlessly, and it does, because the person building the demo fed it ten clean, tidy documents that they picked. Then it meets your business, where the answer to "what's our returns policy?" exists in four places, two of which contradict each other, one of which is a PDF from 2021, and the real answer is whatever Deb in customer service has been saying for the last three years.

That gap — between the demo's data and your data — is where most AI projects quietly die.

The problem isn't quantity, it's authority

Small and mid-sized businesses often assume they don't have "enough data" for AI. Almost none of them are right. You have plenty. What you usually don't have is an agreed answer to a simpler question: for any given thing, which source is the one that's true?

That's authority, and it's a completely different problem from volume. It's also one you can fix without a data warehouse or a six-figure project.

Contradiction is the first symptom. The website says 14 days for returns. The T&Cs say 28. The team have been doing 30 for good customers because it saves arguments. All three are "real." An AI employee will confidently pick one, and it will be the wrong one about a third of the time.

Staleness is the second. Price lists, opening hours, staff names, supplier terms, the process for booking a service. Documents rarely get deleted when they're superseded; they get a new version saved next to the old one, forever. pricing_2024_final.xlsx is still sitting there looking authoritative.

Tribal knowledge is the third, and the biggest. The most accurate information in most businesses isn't written down anywhere. It's the twenty things everyone on the team just knows. No system can use it, which means no automation can either — including the automation you were hoping would free those people up.

Where the truth actually lives in a typical business

Have a look at your own operation and you'll probably recognise most of these.

The inbox. Years of real answers to real questions, in your actual tone of voice. This is genuinely the richest source most businesses have, and almost nobody thinks of it as data. It's also full of one-off exceptions that were never policy, so it needs a filter, not a firehose.

The shared drive. Documents, some current, most not. Useful once someone spends a day marking which folder is authoritative and archiving the rest. That day is the project.

The system of record. Your accounting package, booking system, CRM, job management tool. Structured, reliable, and usually the easiest thing to connect. The catch is that it holds transactions, not explanations — it knows the invoice was raised, not why the discount was given.

People's heads. The largest store, the least accessible. Getting it out doesn't require anything technical: it requires sitting with two or three experienced people for an hour each and writing down what they say.

Nowhere at all. Sometimes the honest answer is that the business has never decided. Nobody has ever set the returns policy; it's just been whatever seemed reasonable each time. This isn't a data problem — it's a decision that's been deferred for years, and AI is simply the thing that finally forces it.

What to do about it, concretely

None of the following needs a technology budget. All of it makes any future AI project dramatically more likely to work — and most of it is worth doing even if you never automate anything.

Write a one-page source-of-truth list. For each of the ten things customers or staff ask about most, name the single place where the correct answer lives. One place. If you can't name one, you've found something to fix. This page is the single most valuable artefact in any AI project, and it takes an afternoon.

Resolve the top five contradictions. Not all of them — the top five. Pick the ones that come up daily, decide what's true, update the authoritative source, and delete or clearly archive the rest. You'll feel the benefit within a week regardless of AI, because your team will stop guessing too.

Interview your two most experienced people. Ask them what they get asked, what they answer, and where they know the written answer is wrong. Record it, write it up, put it in the authoritative place. This is the tribal-knowledge extraction, and it's the highest-value hour you'll spend.

Set a freshness rule and a name against it. Every authoritative source gets an owner and a review date. Without this, you'll have rebuilt the same mess in eighteen months. With it, the AI stays right without anyone doing anything heroic.

Start narrow so the data job stays small. If you scope the first project to after-hours enquiries, you only need the data behind after-hours enquiries to be clean. Scope it to "everything customers ask," and you've signed up to clean the whole business first. This is the single most common reason projects go from six weeks to nine months.

What changes when you get this right

Businesses that do the source-of-truth work first tend to find their AI projects run boringly. The thing goes in, it answers correctly, people stop checking it after a fortnight, and the conversation moves on to what else it could do.

Businesses that skip it tend to end up in a loop: the AI gives a wrong answer, someone loses confidence, a human starts double-checking everything, and within two months the system is being ignored — not because it couldn't work, but because it was asked to be authoritative about a business that hadn't decided what was true.

The good news is that this is one of the few parts of an AI project where the work is entirely within your control, requires no specialist skills, and pays off whether or not you ever build anything.

Our free AI-readiness audit is largely a structured version of the questions above — it'll tell you in a few minutes how close your data is to ready. If you'd rather have someone do the source-of-truth pass with you, that's part of how we work.

At BuildPulse, we spend the first fortnight on your data, because that's where the project is actually won.