McKinsey has a number that ought to stop every CFO in their tracks.
In their April 2026 analysis of where AI is creating value, McKinsey reported that 94 percent of organisations using AI have not seen “significant” value from those investments.
Ninety-four percent. They frame it as a modern Solow Paradox. You can see the AI revolution everywhere except in the financial performance numbers. That’s not a soft pilot failure rate. That’s the vast majority of organisations pouring money into AI and getting nothing measurable back.
It tracks with what I see in the finance teams I work with. The AI rollouts that stall aren’t usually because the AI doesn’t work. They stall because something underneath has been quietly ignored for ten years and finally got exposed.
This is the longer version of an argument I made in my guide to using AI in finance. That piece touched on the data problem briefly and pointed here. This is where I unpack it.
Because in 2026, with agentic AI moving from concept into deployment across enterprise finance functions, the data problem in finance isn’t just inconvenient anymore. It’s expensive. And the CFOs who are quietly slowing their rollouts in the second half of this year are doing so for one reason. They’ve looked underneath and realised what they were about to automate.
Why AI Projects in Finance Fail
The pattern in most finance teams looks the same.
There’s enthusiasm. A vendor presentation that demoed beautifully on someone else’s clean data. A pilot that went well in a controlled environment. A scaled rollout that started producing outputs which were technically impressive but kept being corrected by humans before they got sent anywhere. And then a slow, polite winding down that nobody really wants to call a failure.
The pattern is showing up clearly in finance functions specifically. McKinsey’s November 2025 study of 102 CFOs found that 44 percent of finance leaders are now using generative AI for more than five use cases (up from just 7 percent the year before) and 65 percent will increase their gen AI investment in 2025. So enthusiasm is real, and the budget is flowing.
But the same study notes that nearly two-thirds of respondents say their organisations have not yet begun scaling AI across the enterprise. What stops the pilot from turning into scaled production is, almost always, the same thing. The data underneath.
I tell a story in my FBP training that I’ll borrow here, because it makes the point better than any framework. It comes from Matt Bevan’s podcast Iran’s Global Network of Goons.
A bad actor in Iran wanted to set fire to a Jewish bakery in Bondi, on Curlewis Street. He hired two local goons, gave them the address, and they typed it into Google Maps. Only problem: they got the wrong address. They didn’t set fire to the bakery. They torched Curly Lewis Brewery instead.
The point? Garbage in, garbage out.
AI doesn’t care whether your master data is right. It acts on what you give it. If your cost centre mappings are wrong, AI will produce variance commentary on the wrong cost centres faster than you ever could.
If your customer master has duplicates, AI will reconcile against the duplicates.
If your product hierarchy was set up by someone three system migrations ago and nobody has touched it since, AI will build forecasts off whatever’s in there.
And it will do it confidently. With well-formatted output. Often quite quickly. And it won’t flag that anything is wrong.
What “AI-Ready Data” Means in a Finance Context
You’re going to hear AI-ready data a lot in 2026. Every vendor will sell you a version of it. Most finance leaders I speak to aren’t quite sure what it means, beyond a vague sense that their data isn’t it.
Let me translate.
AI-ready data has roughly four dimensions: quality, structure, context, and governance. In a finance context, that translates to:
Master data that’s clean and consistent.
Your customers, products, cost centres, geographies, and account hierarchies. Each entity exists once, mapped consistently, with the same naming convention. Not three versions of the same customer because of an acquisition fifteen years ago. Not five cost centres that no longer have anyone in them. Not products coded under two different taxonomies, depending on which system you pulled from.
Hierarchies that hold.
When a cost centre rolls up to a department, which rolls up to a function, which rolls up to a region, those rollups need to be stable, documented, and consistent across the systems AI will touch. If your GL hierarchy doesn’t match your reporting hierarchy, or doesn’t match your planning hierarchy, AI is going to produce three different answers and pick whichever one happens to come first.
Granularity that matches the question.
If you’re asking AI to do customer-level profitability analysis but your costs are only allocated at the product line level, the answer is going to be a confident fabrication. The granularity of the data has to match the granularity of the question.
Context and metadata.
This is the newer concept, and where most finance teams haven’t even started. Metadata is the data about your data. When was it created, by whom, from which source, and on what basis? Without it, AI can’t tell whether the number it just pulled is yesterday’s actual, last week’s re-forecast, or a three-month-old budget. The investment in metadata is the unglamorous work that determines whether your AI is right or hallucinating.
Governance and lineage.
When AI gets a number wrong, can you trace back to where the number came from, what data fed it, and which rules were applied? If you can’t, you can’t audit, you can’t fix, and you can’t trust.
None of this is new work for finance. We’ve been talking about master data quality for thirty years. What’s new is that AI raises the cost of not doing it.
Agentic AI Just Raised the Stakes
Here is the shift that has changed the conversation in 2026.
Until recently, AI in most finance functions meant a chatbot. You typed a question, it gave you a text answer, and a human read the answer before doing anything with it. If the underlying data was wrong, the AI’s response was wrong, but you spotted it, fixed it, and moved on. The cost of bad data was a wasted hour and slightly less trust in the tool.
In 2026, AI in finance increasingly means agents. Autonomous systems that don’t just respond, they execute. They post journals. They process invoices. They run forecasts and push the outputs into the planning system. They flag exceptions to other agents, which then act on them.
Because when AI executes on bad data, you don’t get wrong text. You get wrong actions. This is why you need an FBP alongside AI usage.
An agent running customer profitability analysis on a cost allocation hierarchy that hasn’t been refreshed in two years can flag a key account as deeply unprofitable, leading the business partner to walk into the next commercial review recommending a customer exit on numbers that were never right.
An agent generating variance commentary on a dataset that doesn’t include credit holds or shipment delays will confidently attribute a major revenue miss to “competitive pricing pressure”. The business partner runs that up to the executive team, and operations chase corrective actions for a driver that doesn’t exist.
An agent running price sensitivity analysis on cost data that pre-dates the last manufacturing change can recommend a five percent price reduction that looks margin-neutral but isn’t. The business partner walks the recommendation into the sales team (who execute it) and next quarter’s gross margin collapses.
Because agents now hand off to other agents in multi-step workflows, the errors compound. A bad first step becomes the trusted input for the second step. By the time the business partner reviews the final output, they’re signing off on a chain of confidently wrong inputs.
This is why I have CFOs ringing me in 2026, not asking how do we accelerate AI, but how do we slow it down without looking like we’re falling behind. They’ve looked underneath and realised the master data work they put off for the last few years is now sitting in the critical path of every agent they’re about to deploy.
The data problem hasn’t changed. The consequences have.
The Honest Test for Your Finance Data
You don’t need a consultant to tell you whether your data is ready for AI. Three questions will get you most of the way there.
Would you let a new graduate run a forecast manually from your current master data without senior review?
If the answer is no, because I don’t trust the underlying customer or product hierarchies, then you should not be letting AI do it either. Speed doesn’t fix the underlying problem. It amplifies it.
Could you reproduce, today, any number that came out of an AI-generated report from last month?
Trace it back to the source data, the transformations applied, the rules used. If you can’t, you don’t have data lineage. Which means when something goes wrong (and something always does), you cannot audit, fix, or defend.
Does your master data have a clear owner inside the finance function?
Not “IT looks after the systems.” Not “everyone owns it.” A named person whose job description includes maintaining the integrity of the customer, product, cost centre, and account hierarchies. If the answer is not really, that’s your first hire or first reallocation. AI without a data owner is governance theatre.
If you answered confidently to all three, you’re in better shape than most finance functions I work with. If you flinched at any of them, the AI question for your team is not which tool to buy. It’s which data needs cleaning first.
How to Fix Your Data (Without a Three-Year Project)
The trap most finance leaders fall into when they realise the data problem is real is the boil-the-ocean response. A two-year, organisation-wide master data programme. A committee. A consulting engagement. By the time it’s halfway done, the team that sponsored it has been restructured, and the project is dead in the water.
There’s a better way.
Prioritise by use case, not by completeness.
You don’t need all of your data to be AI-ready. You need the data that the next three high-value AI use cases need to be ready. If you’re starting with month-end variance commentary, the GL hierarchy and cost centre mapping have to be right. If you’re starting with customer profitability analysis, the customer master and cost allocation rules have to be right. If you’re starting with pricing sensitivity work, the product cost master and discount data have to be right. Fix what’s in the line of fire of the next use case, not everything everywhere.
Make finance own it.
Master data is not IT’s problem. It’s finance’s lifeblood, and finance professionals are the only people who actually use it daily and feel the pain when it’s wrong. If your function has outsourced this to a “data team” that doesn’t sit in finance, that’s the first thing to undo. The people responsible for the integrity of the numbers must be the people responsible for the integrity of the data underneath them.
Build hygiene, not heroics.
One-off cleanups don’t last. Six months after the project finishes, master data drifts again because no ongoing discipline exists. The teams that get this right embed maintenance into the monthly close. A thirty-minute check on what new master data records were created in the month, who created them, and whether they conform to the taxonomy. Boring, repetitive, and exactly the kind of work AI agents are now genuinely good at.
Use AI to expose the problem before you use it to solve the problem.
This is a small but useful inversion. Before deploying AI to do the production work, deploy it to validate the data underneath. Run an LLM over your customer master and ask it to flag suspected duplicates. Run one over your cost centre hierarchy and ask it to flag inconsistencies. Use the technology against the data first. It will surface, in an afternoon, problems that may otherwise sit unresolved for years.
The same critical-thinking habits that protect you from AI errors at the output stage can be turned around to expose data problems at the input stage.
The Blindspot Has Changed
When I first wrote about the AI finance blindspot in 2025, the argument was that finance teams were too excited about the tools and not paying enough attention to what was underneath them.
That argument has aged into something more urgent. The tools have become more powerful, more autonomous, and more integrated into the workflows that actually move money and book transactions. Which means the cost of the data problem has gone from embarrassing to expensive.
The teams that are succeeding with AI in finance in 2026 are not the ones with the most sophisticated AI strategy. They’re the ones that did the work of cleaning their master data, naming a data owner inside finance, and tying every AI use case to a specific dataset they could vouch for.
If you’re a CFO or head of finance reading this, and you’re feeling pressure to accelerate AI deployment, this is the moment to look underneath instead. The pressure to move faster is not going away. But the cost of moving fast on bad data has become the most important number in your AI business case.
Remember, once the robots arrive, the only thing left will be business partnering.
And the robots can only arrive properly if the data they sit on actually tells the truth.