Anna Totterdell
Projects Director
I want to show you what bad data actually looks like inside a mid-market business. Not in the abstract. In practice.
Your ERP holds customer records. So does your CRM. The two systems disagree on 15% of account names because one was entered by sales and the other by operations. Your finance platform has its own version of the same customer, with a slightly different address and a different contact. Nobody knows which one is correct. Nobody has checked.
Your product catalogue exists in three places. The ERP has SKU codes. The website has product names that do not match those codes. The sales team uses a spreadsheet with its own naming convention. When someone asks "how many units of X did we sell last quarter," the answer depends on who you ask and which system they check.
This is not an edge case. This is normal. And it is the reason AI will not work in your business until you fix it.
What happens when you put AI on bad data
AI does not interpret. It processes. Feed it inconsistent data and it will produce confident, articulate, completely wrong outputs. That is not a limitation of the technology. That is mathematics - which is why AI enablement without structured data is mostly performance, not capability.
When a vendor demonstrates AI summarising customer interactions, it is reading from a clean CRM where every record follows the same structure and every field is populated consistently. When you see AI forecasting demand, it is working from a normalised dataset with years of structured, validated history. When AI recommends an action, it is drawing from a single source of truth.
Now look at what you would actually feed it. Duplicate customer records. Conflicting revenue numbers across systems. Product data that does not reconcile. Dates in three different formats. Free-text fields where you needed structured categories.
The AI will still produce an output. It will look professional. It might even sound plausible. But the underlying logic is built on contradictions, and the decisions you make from it will be wrong in ways you cannot easily trace back. That is worse than having no AI at all.
What bad data actually costs you (before AI even enters the picture)
Forget AI for a moment. Bad data is already costing you real money, and most businesses have never measured it.
How long does your finance team spend each month reconciling numbers between systems? How many hours go into building a board report because the data has to be manually extracted, cleaned, and formatted before anyone can read it? How often has a decision been delayed because nobody trusted the numbers enough to act on them?
There is a compounding cost here that rarely appears on a balance sheet. Every time someone manually checks a figure because the system might be wrong, that is a cost. Every time two departments produce different answers to the same question, that is a cost. Every time a new starter spends their first month learning which data to trust and which to ignore, that is a cost.
In most mid-market businesses, this hidden tax runs into hundreds of hours per year across the organisation. It slows decisions, erodes confidence in reporting, and forces skilled people to do work that should not require human involvement.
Why this problem persists
Data quality is not a technology problem. It is a governance problem. It is also, quietly, where IT and process strategy fails first - nobody wants to own prioritisation until the numbers disagree.
Systems accumulate bad data gradually. A field gets left optional when it should be mandatory. A dropdown list does not cover every case, so people start typing into free-text fields. Two departments use the same system differently because nobody agreed on a standard. An integration breaks silently and nobody notices for three months.
Each of these is a small thing. But over years, they compound into a data estate that is fundamentally untrustworthy. And because the degradation is gradual, there is no single moment where someone raises the alarm. The business just slowly adapts - building workarounds, manual checks, and reconciliation processes to compensate for data it cannot rely on.
By the time someone suggests using AI, the gap between what the data looks like and what AI needs it to look like is enormous. And closing that gap is not a quick fix. It is the grinding work of data and systems integration: mapping every system, agreeing on standards, normalising fields, deduplicating records, and building integration layers that keep things consistent going forward.
What data readiness actually involves
This is not a data warehouse project. It is not a business intelligence initiative. It is practical, operational work.
First: map where data is created across your business. Every system, every spreadsheet, every form. Identify what is the source of truth for each entity - customers, products, orders, transactions - and where copies or conflicts exist.
Second: normalise the critical fields. Agree on naming conventions, categorisation standards, date formats, and mandatory fields. This sounds trivial. In practice, it requires cross-department agreement that most businesses have never attempted.
Third: connect the systems. Build integration layers that keep data consistent as it moves between platforms. When a customer record is updated in the CRM, the ERP should reflect that change without someone manually re-entering it. When an order is placed, every system downstream should receive the same structured data automatically.
Fourth: validate continuously. Data quality is not a one-time project. It degrades the moment you stop paying attention. Build checks, alerts, and governance processes that catch problems before they compound.
Once that foundation exists, AI becomes genuinely useful. Not in the abstract sense. In the specific, operational sense: AI that can summarise a customer's full history because it can actually access a complete, consistent version of it. AI that can flag supply chain exceptions because it can see across systems without contradictions. AI that can recommend pricing adjustments because it trusts the data behind the margins.
The uncomfortable question
If you are planning an AI initiative, ask your team one question before you start: can you produce a single, accurate list of your customers - with consistent names, addresses, and contact details - from your existing systems, without manual intervention?
If the answer is no, you are not ready for AI. You are ready for a data project. And that is where the real work - and the real value - begins.
