Your AI Is Only as Good as the Mess You Point It At

Here is a conversation I have had more times than I can count this year. An owner tells me they tried AI properly. Bought the licences, ran a pilot, got the team on board. And the results were fine. Underwhelming. Not worth the noise. So they quietly wound it back and now they are a bit cynical about the whole thing.

Then I ask what the AI was reading when it gave those underwhelming answers. And there is usually a pause.

Because in almost every case, nothing had gone wrong with the model. The model did exactly what it was asked. It was pointed at a pile of half-finished documents, an inbox, a shared drive with four versions of the same price list, and a business whose actual operating knowledge lives inside three people who have been there a decade. It produced a plausible answer from a mess. That is what these systems do when you give them a mess.

Sixty Per Cent Of AI Projects Are Getting Quietly Shelved

Gartner put a number on this that should stop any owner mid-scroll. Roughly 60% of AI projects are expected to be abandoned because the organisation could not supply data the AI could actually use. Not because the technology underperformed. Because the business turned up to the job without the raw material.

And the sequencing is the painful part. Something like seven in ten organisations only discover their information is not fit for purpose after they have committed to the initiative. The money is spent, the expectations are set, the team has been told this is the year of AI, and only then does someone open the folder and realise the last time that process was documented properly was 2021.

The reality is that AI has made a problem visible that was always there. Your business has been running on undocumented knowledge for years and getting away with it, because humans are brilliant at filling gaps. Sarah knows. Dave remembers. Someone will ask in the group chat. A machine cannot do that. It has no instinct for what is missing, so it fills the gap with a guess and hands it to you in a confident tone.

What “AI-Ready” Means When You Are Not A Bank

The phrase “AI-ready data” has been captured by the enterprise software world, and they have made it sound like something that requires a platform, a migration and a two-year roadmap. For a business with fifteen or fifty people, it is much smaller than that.

Three questions. Is the information written down somewhere a machine can reach? Is it current? Is there one version of it? That is the test. Nothing about warehouses or pipelines. If the answer to all three is yes for a given piece of knowledge, your AI can work with it. If any one is no, you will get output that looks right and is subtly, expensively wrong.

Most businesses fail on the third question rather than the first. There is not an absence of information. There is too much of it, scattered, contradicting itself. Four price lists, two of them stale. A process documented in an old induction deck and again in a Slack thread and again in a spreadsheet, all slightly different. Point an AI at that and it will confidently average your contradictions into something nobody in the business actually believes.

Where Your Business Really Keeps Its Knowledge

Go looking and you will find it lives in four places, roughly in this order.

• In people's heads. The biggest store by far, and the one nobody counts as data. How your best salesperson reads a room. Why you always quote that job type higher. The client you would never take on again and the reason.
• In inboxes. Years of decisions, negotiations and reasoning, sitting in individual accounts, invisible to everyone else and legally awkward to mine.
• In systems that do not talk. The accounting package, the job management tool, the CRM someone set up in 2019 and half-populated. Each holds a slice, none holds the picture.
• In documents nobody has opened in two years. The shared drive. Genuinely useful material buried under versions, drafts and files called final_v3_ACTUAL.

Notice that only the last two look like data. The first two are where the value actually sits, and they are the hardest to get at. I wrote a while back about whether you really know how your best people do their job, and this is the same problem wearing different clothes. The knowledge that makes your business worth something has never been written down, and now you have a technology that can only work with what is written down.

Do Not Start A Data Project. Start With One Job.

The instinct at this point is to launch a cleanup. Get the drive sorted, document the processes, fix the CRM. I would push back hard on that. A general data cleanup has no finish line, no visible return until the very end, and it dies around week six when something urgent lands. I have watched it happen.

Do the opposite. Pick one job you genuinely want AI doing well. Just one. Answering the twenty questions customers ask before they buy. Drafting the first version of a quote. Getting a new hire productive in two weeks instead of two months. Then work backwards and ask what small slice of information that specific job depends on. Usually it is startlingly little. A current price list. Six well-written process notes. One honest FAQ.

Get that slice right and only that slice. You will have something working in a fortnight instead of a year, and the business will see a result while it still remembers why you started. Then take the next job. This is the same argument I make about building capability rather than buying tools. Narrow and finished beats broad and ongoing, every single time.

And use AI to do the work. This is the bit people miss. Getting your knowledge into shape used to mean weeks of someone typing up documentation, which is precisely why it never happened. Now you can sit with your most experienced person for forty minutes, record the conversation, and have AI turn it into structured process notes you edit rather than write. The expert talks. The machine does the transcribing and organising. Someone who knows the subject checks the output before it becomes the source of truth, because human in the loop is not optional when you are creating the thing everything else will rely on.

Once You Organise It, Be Careful Where You Put It

There is a consequence to all this that deserves saying out loud. The moment you gather your pricing logic, your client history, your hard-won process knowledge and your competitive reasoning into one organised, machine-readable body, you have created something genuinely valuable. Arguably the most valuable asset your business owns, and it will not appear anywhere on your balance sheet.

So think about where it goes. Uploading all of it into whatever chatbot happens to be popular this quarter is a decision, even when it is made by not deciding. Open-source models have become crazy good and can run on hardware you control, which means keeping the sensitive core of your business knowledge in your own hands is a practical option now. I have written separately on renting versus owning your AI, and the data question is where that stops being philosophy and starts being an actual decision with consequences.

The Test To Run This Week

Forget strategy days. I would massively challenge you to do this one thing before Friday, and it takes about twenty minutes.

Write down five questions a new employee would ask in their first week. Real ones. How do we price this kind of job? What happens when a client wants to cancel? Who approves a discount and at what point? Then go and try to find the single authoritative answer to each, without asking a person.

You will land in one of three places. You find nothing written down, which means that knowledge lives in a head and is one resignation away from leaving. You find three answers that disagree, which means your AI will pick one at random and sound certain about it. Or you find one clear current answer, in which case that piece of your business is genuinely AI-ready today.

Count how many of the five land in the third bucket. That number tells you more about whether AI will work in your business than any vendor demo ever will. And the fix for the other four is not a platform. It is a couple of good conversations, written down properly, in one place.

Frequently Asked Questions

What does AI-ready data actually mean?

AI-ready data means the information your AI needs to do a job is written down somewhere a machine can reach, it is current, and there is one version of it rather than six. That is the whole test. It does not require a data warehouse or a data team. For most small and medium businesses it means taking knowledge that currently lives in someone's inbox, in a folder nobody else opens, or in an experienced employee's head, and putting it into a form your AI can actually read.

Why do so many AI projects fail because of data?

Gartner has projected that around 60% of AI projects will be abandoned because the organisation could not supply data the AI could use. The pattern is consistent. A business buys a tool, runs a pilot, gets mediocre output, and blames the model. What actually happened is that the AI was asked to answer questions using information that was contradictory, out of date, or simply never recorded. A model cannot retrieve knowledge that was never written down, and it will produce a confident guess instead of telling you it has nothing to work with.

Do I need to clean up all my data before starting with AI?

No, and trying to is how businesses waste a year. A full data cleanup is a project with no visible finish line and no return until the very end. The better approach is to pick one job you want AI to do well, work out the small slice of information that job depends on, and get only that slice right. Fix the data for one workflow, prove it works, then move to the next. Narrow and finished beats broad and ongoing every time.

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

Run a simple test. Pick a question a new employee would ask in their first week, something like how you price a particular kind of job, or what your policy is when a client wants to cancel. Then try to find the single authoritative answer without asking a person. If you find nothing, or you find three answers that disagree, your data is not ready for that task. If you find one clear current answer in a place a system could read, it is. Repeat this across five or six real questions and you will have a very honest picture.

Does using AI mean handing my business data to a big tech company?

It does not have to. This is exactly why the question of where your data goes matters as much as what your AI can do. Once you have organised the knowledge that makes your business valuable, that collection becomes one of your most important assets, and it is worth being deliberate about who holds it. Open-source and locally run models have become genuinely capable, which means running AI over sensitive business knowledge on infrastructure you control is now a realistic option rather than a theoretical one.