Why mapping your data estate before selecting an AI tool is the only way to traverse the minefield.

The year is 1993. Several suits sit around a long boardroom table, hatching a plan that will alter the course of their company’s history. The CEO puts pen to paper, signs on the dotted line, and lets out a confident chuckle. The next corporate black-ops mission is a go.

Code name: Delta III

The Mark: Operational bottlenecks choking FoxMeyer Drug, a $5 billion pharmaceutical distribution goliath.

Action Plan: A $100 million SAP ERP implementation to automate their warehouse and neutralise the competition. But in a brutal twist of fate, the drug company had unknowingly administered its own lethal injection. By choosing the SAP R/3 system, they had selected a system designed for manufacturers, not wholesalers. It could only process 10,000 orders a night; they needed 420,000. Implementors pushed to hit deadlines rather than validate the system worked. Employees perceived they were being automated out of a job. In response, warehouse staff damaged products and refused to fulfil orders.

By 1996, FoxMeyer had filed for bankruptcy. It was a masterclass in how not to adopt transformative technology. There was a fundamental failure to audit the operational and data foundations before committing to the tooling.

In 2026 the characters have evolved. The weapon of choice is Artificial Intelligence. But the story is all too familiar.

Business leaders are committing to AI subscriptions before running the diagnostics on the problem they are solving or assessing whether the data is fit for powering the solution. The poison is much slower than the FoxMeyer case, AI will confidently produce consistent-looking, plausible answers.  But by being rooted in a broken foundation, issues compound quickly as decisions are taken unknowingly on miscalculations.

The antidote is not to avoid AI. It is to start with an AI Opportunity Audit. This will map which opportunities will have the greatest business impact versus how ready your data estate is to deploy against. This article will outline two primary opportunity types and the framework to assess your position.

Recon Target 1: Using AI to Support Better Business Decisions

In most companies, there are critical junctions where operations grind to a halt because someone must ‘run the numbers’ manually. Forecasting inventory based on gut feel rather than data and relying on your resident ‘spreadsheet superman’ to manually flag anomalies.

This is where applied machine learning and operational intelligence can be used to find hidden stories within the data of the business. An off the shelf ERP can be used to track stock levels, though it’ll never question them. An AI-driven model can spot that future orders of your fast-moving stock are void of existence. Flags are thrown to not replenish the warehouse to avoid a dead weight stock burden. Likewise, AI can assess purchase price fluctuations as they happen and spot the trends that a procurement team may miss. Reporting suites tell you what happened, AI tells you what you should do about it.

The actions AI enables are where the real business impact is. Plus, by removing dullness, you free up capacity for your team to more meaningful work. For each use case, the audit establishes three things:

• The specific decision made and by who
• The metric the tool is attempting to improve
• The minimal viable dataset to run a credible pilot

Typically, this can be achieved within eight weeks if the data foundations are already in order.

Recon Target 2: Using AI to Answer Business Questions Faster

You probably have reports in some shape or form. Power BI dashboards with alluring colours or a vintage Excel style beast. But it leaves you feeling like something is missing, you can’t quite answer that question. It could be “How many active customers do we have?” or “What is the order status of order number 10132?”

The current process might be to raise a data request wait, chase and… wait. It might have entered the abyss of all the other requests. That’s not a dig at your team, you’re probably not the only one asking ad hoc questions and they have strategic work they need to do. This bottleneck is structural not personal.

AI natural language querying could be your new data ally. Given enough business context, it’ll convert your words into code and execute it against data and return the answer in seconds. You can now go action what you need to do, and your data team can crack on with the tasks at high value that’ll enhance your business.

The audit would uncover hidden risks here too. The literal nature of language models means they will guess when instructions are murky. Without a formally defined central semantic layer agreed across functions, AI will invent its own way for deciding if a customer is active. The numbers returned may look ‘about right’, but they’re simply incorrect. The speed and feasibility of the output make mistakes challenging to catch. The business impact of doing this right is clear. But the audit question will be, ‘do we have the governed, defined, trustworthy data that an AI interface needs to answer questions accurately?’

Your Data Estate: The Preconditions AI Tool Sellers Won’t Talk About

Both the opportunities discussed are completely within reach for SMEs without the need for a data science function or a FTSE 100’s data budget. But there’s a catch (there’s always a catch). Before you Delta III yourself, know that this stuff comes at a price. I don’t mean monetary; the cost of admission is the quality of your data estate.

Humour me briefly and consider this. System A records your customer as ‘VantaForge Tech’. Somewhere else in the same system, the same entity appears as ‘VantaForge Technologies.’ Without an explicit rule saying these are the same organisation, no AI model will assume they are. Your churn model will treat them as two separate customers. Your natural language query will return partial results. And the miscalculation will look entirely credible on the surface.

That’s assuming that the data is all in the same place. A more structural problem is having data across multiple unintegrated systems. Each hold fragments of the truth but require a very experienced human analyst to manually piece together the whole picture. A less experienced analyst would struggle, as would AI. The analyst will tell you when it doesn’t know, the AI doesn’t and fills the gaps with assumptions.

To deploy AI safely, these three infrastructure elements must be in place:

1. Stable automated ELT pipelines that integrate your systems into a single source of truth.
2. Strict schemas that enforce consistency at the point of data entry so ‘VantaForge Tech’ and ‘VantaForge Technologies’ cannot coexist as distinct entities.
3. A centrally defined semantic layer that translates raw database fields into agreed business logic.

Conducting the AI Audit: The AI Prioritisation Matrix

Sorry for being boring, but I can’t suggest buying software first. If you want to avoid being the next FoxMeyer and featured in a blog post by the 2030s equivalent of a data engineer the audit must be the starting point.

Start by mapping out areas where lots of the same manual tasks are repeated or predictive insights are hard to come by. Those are largely places where one feels, ‘there must be a better way of doing this’.

The objective then is to map these candidates on a 2×2 matrix. On the x-axis is the business impact (potential for margin expansion, cost reduction, or workflow acceleration) and on the y-axis is data readiness (maturity of defined business rules, cleansed and structured datasets, accessibility to the data).

This is plotted in more detail below:

Quadrant 1 (High Impact, High Data Readiness) – Deploy

This is where immediate, measurable ROI happens. Business processes are driven by and reported on centralised, structured datasets. There’s a semantic layer defined on rigorous business logic, delivered by robust automated ELT pipelines and a very clear picture of what the business case is.

Audit Questions:

1. Is this data governed by a single source of truth, or are there still manual overrides happening in local spreadsheets?
2. Are the definitions within this dataset strictly agreed upon across all departments (e.g., does Sales and Finance agree on what constitutes “Net Revenue”)?
3. Do we have the compute infrastructure to run predictive models against this data at the required frequency?

Quadrant 3 (Low Impact, High Data Readiness) – Defer

The datasets are curated in the same manner as Quadrant 1; they are mature, accessible, well-governed. The meaningful differentiator is the potential of the commercial return. Be wary of starting here. Even a technological success tends to see low adoption as no operational pain point is being addressed.

Audit Questions:

1. If we perfectly automate or predict this process, does it free up headcount, increase revenue, or fundamentally change how we operate?
2. Is this project being driven by a genuine business need, or by a technical team’s desire to experiment with new tooling?

Quadrant 3 (Low Impact, High Data Readiness) – Defer

The datasets are curated in the same manner as Quadrant 1; they are mature, accessible, well-governed. The meaningful differentiator is the potential of the commercial return. Be wary of starting here. Even a technological success tends to see low adoption as no operational pain point is being addressed.

Audit Questions:

1. If we perfectly automate or predict this process, does it free up headcount, increase revenue, or fundamentally change how we operate?
2. Is this project being driven by a genuine business need, or by a technical team’s desire to experiment with new tooling?

Quadrant 4 (Low Impact, Low Data Readiness) – Ignore

Generally, data exists in silos, it’s unstructured, difficult to access, there’s no documentation on derived values and it is maintained manually. Plus, the commercial justification of the opportunity here is low value which gives no motivation to fix such issues.

Audit Questions:

1. Why is the opportunity on our radar?
2. Are there fundamental issues we must address first with our data before we can consider further long-term projects?

Quadrant 2 (High Impact, Low Data Readiness) – Plan Carefully

This is the most important quadrant to get right!

The temptation in Quadrant 2 is understandable but dangerous. The opportunity is real, and with mounting pressure to show progress with AI is too. This is FoxMeyer dynamic again, but as already stated much slower and harder to catch due to the plausibility of the outputs.

The correct response is not to abandon the project. It is to recognise the need for data remediation. Use the audit to identify the steps to govern and structure the data to get this right. Define what the semantic layer is presenting the AI, establish the method of pipelining the data there and identify if external resource is needed to achieve that.

That exercise is the deliverable.

The organisations that will be rewarded the most aren’t the fastest acting. They are the ones that had the discipline to take the time to honestly audit their current position and make steps to move themselves up to Quadrant 1.