What to Fix Before You Buy an AI Assistant

Six things to put right, in order, each with a test you can run this week without buying anything.

Monday, 08:15. Four people round a table, coffee going cold.

Last week's demo was brilliant. Somebody typed "which orders are at risk today?" and the answer came back in two seconds, neat and certain. People actually said "wow". Out loud. In a meeting.

Now there's a quote on the table and a lot of energy in the room. The finance director wants a start date. The sales director wants it yesterday.

And then there's Sam, who runs operations, sitting slightly back from the table and saying nothing.

Sam isn't against it. Sam is the one who'll get the phone call at four o'clock when the AI promised a customer something the warehouse can't do. Sam is quietly running through every number that tool will read, and knows exactly which ones are wrong.

Nobody asks Sam. They ask when they can start.

It's the wrong question. Not because the tool is bad. Because it's the last item on a list nobody in the room has written yet.

This article is that list.

It's the final piece in a series that began with your AI is not guessing, it is reading bad data very confidently. That first piece set out four conditions an AI needs before it can answer an operational question properly, and each piece since has taken one of those conditions apart. Here they come together, in the order they need doing, so that the next time you're in that room, you're the person with the list.

The short version. Before you buy an AI assistant for your operation, fix six things, in this order: the questions you actually need answered, where each answer lives, whether the numbers are true, what each number means, who is allowed to see it, and only then the assistant itself. Each step depends on the one before it, and you can test the first five without buying any new software.

Why does the order matter?

Because each step rests on the one before it. Skip one, and the problem doesn't disappear. It moves further down the line, to a place where it's much harder to see.

Data problems don't shout. They wait.

In 2021, a team at Google Research interviewed 53 AI practitioners working in high-stakes fields such as health and conservation, in India, East and West Africa and the US. 92% of them had been caught by what the researchers called data cascades: a small problem with the data, early on, that compounds quietly and turns up later as a failure somewhere else entirely. The worst took two to three years to surface.

Two to three years. Long after everyone who made the decision has stopped thinking about it.

The paper's title is a line one of them said in an interview, and it will sound familiar to anyone who has ever been Sam: "Everyone wants to do the model work, not the data work."

Those were specialist teams building their own models, not operations businesses buying a tool. But the feeling carries straight across. Do things in the wrong order and the mistake doesn't announce itself. It sits there, patiently, until a customer finds it for you.

Gartner put a number on where that road ends. In February 2025 it predicted that through 2026, organisations will abandon 60% of AI projects that are not supported by AI-ready data. In the same release, 63% of organisations in a survey of data management leaders said they either didn't have the right data practices for AI, or weren't sure.

Not sure. That's the honest answer in most businesses, and there's no shame in it. The six steps below are how you find out, and the first one has nothing to do with data at all.

Step 1: What questions do you actually need answered?

Not "what could AI do for us?" That question leads to slide decks.

The real one is smaller: what would you ask it at 7:45 on a Tuesday, with a coffee in one hand and a problem in the other?

Write down ten. Real ones, in the words people actually use. Which orders will miss today's cut-off? Which customers are waiting on a delivery we haven't dispatched? Which vans are due a service this week?

Next to each one, write who asks it and what they do with the answer. A question nobody acts on is a dashboard, not a decision.

This is where a surprising number of projects go wrong before they've even started. When RAND interviewed 65 experienced data scientists and engineers in 2024, the first root cause of failure they found was that people misunderstand, or miscommunicate, the problem the AI is meant to solve. Their advice to industry was blunt: focus on the problem, not the technology.

The test: can you list ten questions, each with a named person who acts on the answer? If you run dry at four, don't worry. You're not ready to buy anything yet. You're ready for a better meeting.

Step 2: Where does each answer live?

Every question on your list needs a number to answer it. So the next job is to find out where each of those numbers actually lives.

Take each question and follow the number back to where it's kept. Not where it's reported. Where it's first written down.

This is usually the moment the room goes a bit quiet.

Because the number lives in three places, and they disagree. Or it lives in a spreadsheet that belongs to one person, who is on leave until Thursday. Or it doesn't live anywhere at all. It's worked out fresh every morning, in somebody's head, by somebody who has been doing it so long they've forgotten it's a skill.

Pick one home for each number and one person who owns it. Everywhere else becomes a copy.

The test: on the same morning, ask two departments for the same figure, separately. Stock of your best-selling line, say, or deliveries due today. If the numbers match, lovely, move on. If they don't, you've just found your step 2. And an AI will not settle that argument for you. It will simply pick a side, and sound very sure about it.

Step 3: Are the numbers true?

Giving every number one home is real progress. But it leads straight to the uncomfortable question: is the number in that home actually right? A number can live in exactly one place and still be wrong.

The first piece in this series covered research on how often stock records are wrong, and it isn't easy reading. The fix, though, is wonderfully old-fashioned. Pick twenty items. Go and count them. Compare your count with the system.

Then think about yesterday. Write down the three things that went wrong: a truck that wouldn't start, a short delivery, a customer who moved the date at the last minute. Now try to find each of them in your systems.

If they only exist in a WhatsApp group, or in the memory of the person who sorted them out, an AI will never know they happened. As far as it's concerned, yesterday went perfectly. What your dispatcher knows that your system does not is about exactly that gap, and the people quietly filling it.

The test: twenty items counted and matched, and yesterday's three problems found as records rather than messages. If either one fails, fix the habit before you buy anything that reads the records. It's a far cheaper lesson now than later.

Step 4: What does each number mean?

By now your numbers live in one place and match what's really there. The next risk is quieter. Two people can look at the same correct number and mean completely different things by it.

So write the definitions down. Late against what? Available to whom? Delivered when the driver taps the screen, or when the customer signs?

It sounds like admin. It feels like admin. Somebody will roll their eyes.

But one written definition per number, agreed by the person who owns it, is one of the biggest things standing between an AI and a right answer. AI is bad at reading databases. So stop asking it to. sets out the research behind that.

The test: take the three words your business uses most (late, available, done, whatever yours are) and ask three people to define each one in writing, without conferring. If you get nine different answers, don't panic. Everyone does. Now you know what this week is for.

Step 5: Who is allowed to see what?

Once everyone agrees what each number means, you can decide who should see it. The order matters here too: you can't sensibly protect a number until you know exactly what it is.

Go back to your ten questions. For each one, write down who may ask it, and who may not.

Margin is the obvious one. But there are quieter ones that matter just as much: one customer's prices, a driver's hours, an account that's been put on stop and doesn't know it yet.

If those are protected today only because people know not to look, that protection runs on good manners and trust. Your team has plenty of both. Software has neither.

Who is allowed to know that? explains why this has to be built into the data itself, not written into the assistant's instructions and hoped for.

The test: for every question on your list, can your systems already stop the wrong person seeing the answer? Not "would they look?" They wouldn't. Can the system actually stop them?

Step 6: Now buy the assistant. Then start small.

If steps 1 to 5 hold, something lovely happens. The AI becomes the quick part, because all the hard thinking is already done. It has real questions to answer, one home for every number, numbers that are true, clear meanings and sensible limits.

Here's how to bring it in without undoing all that work:

• Read-only first. Let it answer questions before it's allowed to change anything.
• Only from defined figures. It chooses from the numbers you defined in step 4. It doesn't work out its own.
• Test it on questions you already know the answer to. Ten questions, ten known answers. Score it the way you'd score a new hire in their first week.
• Actions last, with a person confirming. Re-dispatching a vehicle or escalating an order should always need a human yes.

None of this is exotic. It's exactly how you'd welcome a new person into the team: show them around, check their work, then trust them a little more each week.

And it isn't only common sense. The wider research points the same way. In McKinsey's March 2025 survey of 1,491 participants, redesigning workflows had the biggest effect on whether an organisation saw any bottom-line impact from generative AI. Yet only 21% of those using it had fundamentally redesigned even some of their workflows.

The businesses getting results didn't just buy a tool. They changed how the work gets done.

The test: could you hand the assistant's first week of answers to the person who owns each number, and would they happily sign them off?

What if you fail the early steps?

Not every business will reach step 6 on the first try. If you fail one of the early steps, then honestly, congratulations. You've found the real piece of work before it found you, and it's usually cheaper than the quote on the table.

Most businesses that fail step 2 or step 3 aren't short of data. They have too much of it, in too many copies, kept by systems that were never built to agree with each other. That's not anybody's fault. It's how most operations grow: one sensible decision at a time, over years.

But it's far better to discover it in a Monday meeting than six months after the contract is signed. And it raises a fair question: how do you stop those copies piling up in the first place?

How this runs on one platform

That's the problem Illuminate was built around, which is why steps 2 and 3 are where the platform does most of the heavy lifting. Illuminate is a unified business operations platform. It doesn't ask you to throw out what you already run. Your ERP, your accounting software, your carriers and your devices connect into it in real time, alongside its own products, and together they become one screen for running the operation: orders, stock, deliveries and assets, all in one place.

Inside that one place, each record is written once, by the product doing the work. A delivery confirmed in Cargo TMS updates the order in Flow OMS and the stock in Depot WMS at the same moment, and the update can flow straight back to your ERP, with nobody re-keying a thing. For the operational numbers on your list, there's one place each can live.

Exceptions are recorded as they happen, too. A vehicle off the road sits in Tagz AMS with its current status, and the Cargo dispatch board counts exceptions right next to ready, dispatched and en route. Yesterday's three problems are already records, not messages.

Steps 4 and 5 are part of setting up. Illuminate maps each client's data, business context and reporting needs, and builds the figures the business runs on, each one scoped by role. The finance team's figures show cost, the operations team's show how busy each vehicle is, and a regional manager's stop at the edge of their region.

And it doesn't have to happen all at once. A business can start with the product that holds its most painful question and add others as it goes, each one connecting to those already in place. Fraser Health Authority went live on six modules across 110 locations in under ten days.

None of that is an AI assistant. It's the five steps that have to be true before one is worth buying.

The short version

Questions first. Then where the answers live, whether they're true, what they mean and who may see them. Then the assistant, started on questions you can already answer.

Done in that order, the AI arrives in a business that can check its work, and Sam finally gets to enjoy the demo.

Done the other way round, you find out what you skipped from a customer.

Talk to us about where your numbers live

Frequently asked questions

How do I know if my business is ready for an AI assistant?

Run five checks before buying. List ten real questions with a named owner for each; confirm each answer lives in one place; count a sample to confirm the numbers are true; write down what each key number means; and confirm your systems can stop the wrong person seeing each answer. If all five hold, the business is ready for an assistant. If not, the failed check is the work to do first.

What should you fix before implementing AI in operations?

Fix the data underneath it, in order: the questions that matter, a single home and owner for each number, the accuracy of those numbers, a written definition of each one, and who is allowed to see it. Each depends on the one before it. An assistant added on top of a business that has not done these will answer confidently from numbers nobody can vouch for.

Why does the order of AI preparation matter?

Because data problems compound downstream and show up late. In a 2021 study of 53 AI practitioners in high-stakes fields, 92% had experienced these "data cascades", and the most severe took two to three years to surface. Defining a number before agreeing where it lives, or restricting access before knowing what the number means, builds each step on one that is not yet true.

Do you need new software to get ready for AI?

Not to find out where you stand. Every test in the first five steps can be done with existing systems, a count and a meeting. Fixing what the tests uncover may need changes to how systems are connected, particularly where the same figure is kept in several places that disagree.

Should an AI assistant be allowed to take actions in operational systems?

Eventually, and carefully. Start read-only, answering from figures the business has already defined, and test it on questions with known answers. Actions such as re-dispatching a vehicle or escalating an order should come later and always require a person to confirm them.

What is AI-ready data?

Data that an AI can rely on without a person correcting it first: kept in one place, accurate, clearly defined and protected by role. Gartner predicted in February 2025 that through 2026, organisations will abandon 60% of AI projects not supported by AI-ready data, and 63% of organisations it surveyed were unsure or lacked the right data practices for AI.

References

Sambasivan, N., Kapania, S., Highfill, H., Akrong, D., Paritosh, P. and Aroyo, L. M. (2021). "Everyone wants to do the model work, not the data work": Data Cascades in High-Stakes AI. Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. ACM.

Gartner (2025). Lack of AI-Ready Data Puts AI Projects at Risk. Press release, 26 February 2025.

Ryseff, J., De Bruhl, B. F. and Newberry, S. J. (2024). The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed. RAND Corporation.

McKinsey & Company (2025). The state of AI: How organizations are rewiring to capture value. Survey of 1,491 participants, July 2024, published 12 March 2025.

More in this series

This piece closes a series. The overview, with the research on why AI answers go wrong on operational data, is your AI is not guessing, it is reading bad data very confidently.

The deep dives: what your dispatcher knows that your system does not, who is allowed to know that?, and AI is bad at reading databases. So stop asking it to.

For media inquiries, please contact:
contact@illuminate.ae