Messy data? Here is how you get a grip

Almost every organisation we meet has the same quiet worry: nobody fully trusts the numbers. A report says one thing, the spreadsheet says another, and someone in the meeting mutters that the figure looks off. The instinctive response is to launch a big project to clean everything up, and that project usually stalls, because cleaning all your data at once is enormous, thankless and never quite finished.

The good news is that you almost never need to. You can get a real grip on messy data in days, not months, if you resist the urge to fix it all and instead start from a single decision that actually matters. This article walks through why data gets messy in the first place, and a pragmatic sequence you can follow to make one important number trustworthy, then repeat.

Why data gets messy in the first place

It helps to know that mess is the default state, not a sign of failure. Data gets tangled for ordinary, structural reasons, and most of them have nothing to do with carelessness.

• Silos. Sales lives in the CRM, finance in the accounting package, operations in a separate tool. Each is fine on its own, but they never agree because nobody stitches them together.
• Manual entry. Every field a human types by hand is a field where a name gets misspelt, a date slips into the wrong format, or a step gets skipped when things are busy.
• Inconsistent formats. One system writes dates as day-month-year, another month-day-year. Amounts carry currency symbols in one place and not another. The same customer is 'Ltd' here and 'Limited' there.
• Duplicates. The same customer, product or invoice is entered twice under slightly different spellings, so counts and totals quietly drift.
• No single source of truth. When three systems each hold a version of the same fact, there is no agreed answer to 'which one is right?', so every report can be argued with.

None of this is unusual, and none of it needs to be solved everywhere before you can make progress. It just needs to be solved where it affects a decision you care about.

Do not boil the ocean

The single most useful principle here is also the most counter-intuitive: do not try to clean everything. A general clean-up has no finish line, no clear owner and no obvious payoff, which is exactly why these projects drag on and get abandoned.

Instead, anchor the work to one question that matters to the business. Not 'is our data clean?' but something concrete: which customers churned last quarter, what is our true margin per product, which channel actually brings in revenue. A specific question tells you precisely which data is in scope, and, just as importantly, which data you can safely ignore for now.

You do not need clean data. You need data that is clean enough to answer the question in front of you.

A pragmatic sequence

Once you have your question, the path is short and repeatable. Work through it in order, resisting the temptation to widen the scope as you go.

1. 1Pick one question that matters. Choose a single decision or metric where a trustworthy answer would genuinely change what you do. Write it down in one sentence so the scope stays fixed.
2. 2Find the data that feeds it. Trace which systems, tables and fields actually contribute to that answer. Usually it is fewer sources than you expect, often two or three.
3. 3Assess its quality. Look at the real data, not your assumptions about it. Where are the gaps, the duplicates, the odd formats? A quick eyeball of a few hundred rows tells you most of what you need.
4. 4Clean only what that question needs. Fix the fields that feed your answer and leave the rest alone. That column of notes nobody uses does not need your attention today.
5. 5Connect the sources. Join the cleaned data together on a shared key, a customer ID, an order number, so the question can be answered from one combined view instead of three arguing ones.
6. 6Document what you did. Note where the data came from, what you changed and why, and any assumptions you made. This turns a one-off fix into something repeatable and trustworthy.

That last step is the one people skip, and it is what separates a quick win from a lasting one. A short note today saves an afternoon of detective work in three months' time.

Common problems and quick fixes

When you look closely, most mess falls into a few familiar buckets, each with a well-understood remedy.

• Duplicates. Deduplicate on a reliable identifier where you have one, or on a combination of fields, name plus postcode, for example, where you do not. Decide upfront which record wins when two disagree.
• Inconsistent formats. Standardise to one convention: a single date format, one way of writing currency, consistent capitalisation. Do it once, in one place, so everything downstream matches.
• Missing values. Decide deliberately per field. Sometimes you can fill a gap from another source; sometimes the honest move is to flag it as missing rather than guess and quietly mislead yourself.
• New mess creeping back. Add validation at the point of entry, required fields, dropdowns instead of free text, format checks on dates and emails. It is far cheaper to stop bad data going in than to clean it later.

Notice the pattern: the first three fix what already exists, and the fourth stops the problem returning. Do only the first three and you will be back where you started within a year.

Good enough for the decision beats perfect

Perfect data is a fantasy, and chasing it is expensive. The right bar is 'good enough for the decision in front of you'. If you are deciding which product line to grow, margins accurate to the nearest percent are plenty, you do not need them to the cent. Match the effort to the stakes.

There is a point where investing in proper tooling, a data warehouse, automated pipelines, a real dashboard, genuinely pays off. That point arrives when you find yourself answering the same questions repeatedly, when manual clean-up eats hours every month, or when several teams need one shared, trusted view. Until then, a careful spreadsheet and a documented process will serve you well. Build the infrastructure when the recurring pain justifies it, not before.

Clean data is now what makes AI trustworthy too

There is a newer reason to care about this in 2026. As soon as you point AI at your data, a chatbot that answers from your documents, a retrieval-augmented setup, an analytics assistant, data quality stops being a back-office concern and becomes visible to everyone. Traditional reports can hide a bad field for years; a language model will confidently repeat it, or draw the wrong conclusion from a duplicate, straight to a customer or a colleague.

The mechanics are the same ones this article already describes. Deduplicated records, consistent formats, a clear single source of truth and a named owner are exactly what let AI retrieve the right answer instead of a plausible-sounding wrong one. That is why data quality has climbed back to the top of most organisations' priority lists: not because the techniques changed, but because messy data now fails loudly instead of quietly. The reassuring part is that the same one-question-at-a-time approach still applies, clean the data behind the decision you want AI to support, and you get a reliable assistant for that decision without a company-wide overhaul first.

Give every dataset an owner, and start small

One light-touch habit keeps data clean once you have fixed it: every important dataset needs a named owner. Not a committee, a person who is responsible for whether the customer list, the product catalogue or the sales figures can be trusted. When ownership is clear, problems get noticed and corrected. When it is nobody's job, mess quietly returns.

So do not wait for the perfect moment or the big project. Pick the one number you most wish you could trust, walk it through the sequence above, and give it an owner. You will end the week with one answer you can rely on, and a method you can point at the next question, and the one after that.