Key Takeaways
Data-driven decision making breaks down not because organisations lack data, but because that data isn’t being turned into something actionable before it reaches the person making the call.
- Dashboards and reporting are rear-view mirrors, accurate, but built to describe the past, not to inform what happens next.
- The gap shows up in demand forecasts built on last year’s patterns, pricing reviewed on a calendar rather than market signals, and risk flagged only after it’s already a loss.
- Machine learning doesn’t replace human judgement; it gives decision-makers a materially better starting point, including a model’s own measure of its uncertainty.
- The right place to start isn’t a data strategy overhaul. It’s identifying the one highest-impact decision currently being made on weak evidence.
- A model must be production-ready, explainable to the people relying on it, tested in fast cycles, and monitored for drift; otherwise, it won’t hold its value beyond launch.
A finance director opens the quarterly forecast three days before the board meeting. It’s built on last quarter’s spreadsheet, updated by hand, checked by one person who’s also updating four other tabs. The demand numbers assume the same seasonal pattern as last year, and nobody’s had time to check whether that still holds. The board approves the plan anyway, not because the numbers are trusted, but because whoever’s most senior in the room is comfortable with the direction. It’s a well-known dynamic in decision-making circles, sometimes shortened to HiPPO: the Highest Paid Person’s Opinion tends to carry more weight than the model does, whatever the model says.
Three months later, the actuals come in wrong, in the direction they’re usually wrong. Nobody’s surprised. Everyone moves on to the next forecast, built the same way.
This isn’t a data shortage. Organisations in this position are usually holding more data than they’ve ever had. The shortage is in what that data gets turned into before it reaches the person making the call, and in how much weight it carries once it gets there.
More data hasn’t meant faster decisions.
It’s tempting to assume that the businesses struggling with this have a technology gap: no dashboards, no data warehouse, nothing modern. In practice, the opposite is often true. They’ve invested in the infrastructure. The dashboards exist. The reports go out on schedule. The problem sits one layer deeper.
Fathom’s 2026 survey of financial professionals found that the large majority still consider manual spreadsheet processes a significant pain point in budgeting and forecasting, and that manual data entry regularly leads to critical errors, from broken formulas to version control problems that ripple into the decisions built on top of them. None of this is a failure of effort. Finance teams are working hard inside a process that was never designed for the volume or speed of decisions now being asked of it.
McKinsey’s 2025 Global Survey on AI points to a related pattern at the enterprise level. Nearly nine in ten organisations now report regular AI use somewhere in the business, yet only around a third have moved past piloting into scaling it, and just 39% attribute any measurable profit impact to AI at all, with most of those crediting less than 5%. Inaccuracy is the single most common negative consequence organisations report from using AI. Adoption has clearly broadened. Trustworthy, decision-grade output hasn’t kept pace with it.
Reporting tells you what happened. It doesn’t tell you what to do
Here’s the distinction worth naming plainly: a dashboard is a rear-view mirror. It’s accurate, it’s often beautifully designed, and it tells you, reliably, what already occurred. It doesn’t tell you what’s about to happen, what to price something at tomorrow, or which of ten thousand transactions is the one that needs a human look before it clears.
That’s not a criticism of reporting. Reporting does its job well. The mistake is treating it as if it were the same job as decision-making, when the two require fundamentally different tools. A report describes the past with precision. A decision needs a defensible view of the future, and reporting was never built to produce one.
This is where the gap lives, and it’s worth being specific about how it shows up: demand forecasts that assume next quarter looks like last quarter, pricing reviewed on a calendar rather than in response to what’s happening in the market, risk flagged only once it has already materialised into a loss. Forecast accuracy benchmarking research drawing on data from more than a thousand organisations puts median monthly demand forecast accuracy at around 85% at the category level, a number that can look healthy in aggregate while hiding serious errors at the individual product or location level underneath it. Averages reassure. They don’t run operations.
Machine learning doesn’t replace the judgement of the people currently making these calls. It gives them a materially better starting point. A model trained across thousands of historical outcomes can hold patterns no analyst could track by hand, and it can put a number on its own uncertainty in a way a spreadsheet simply can’t. The organisations pulling ahead in their sectors aren’t the ones with the most data. They’re the ones that have turned their data into a structural advantage rather than a monthly reporting exercise.
What closes the gap?
Closing this gap doesn’t start with a bigger data platform. It starts with a specific, uncomfortable question: which decision, right now, is being made with worse evidence than it should be, and what would it be worth to fix that one decision first?
That’s a deliberately narrow starting point, and it holds a few things in place that matter:
It’s built around a decision, not a dataset. A forecasting model, a pricing model, a risk-scoring model — each one is built to improve a specific call the business already has to make, not to explore data for its own sake.
It’s production-ready from the outset, not a notebook exercise. A model that performs well in a data scientist’s testing environment but can’t be maintained once it’s live isn’t a deliverable. It’s a proof of concept that never left the lab.
It’s explainable enough to be trusted by the people relying on it. A pricing team or a credit team that can’t see why a model reached a particular number won’t act on it, no matter how accurate it is on average.
It’s tested in fast cycles, not settled on the first attempt. The same discipline marketing teams have long used for A/B testing, running a small experiment, checking what moved the number, adjusting, and running the next one, applies just as well to a pricing model or a forecasting model. A hypothesis about what would improve a decision should be answerable in days, not committed to for six months before anyone finds out whether it holds. That speed is what makes the difference: an answer that takes half a year arrives after the conditions it was based on have already changed. An answer that takes a week can be implemented, checked again a week later, and adjusted while it’s still cheap to adjust — which is what lets a business actually pivot, rather than just react.
It’s monitored because the world doesn’t hold still. Conditions shift, customer behaviour shifts, and a model that was accurate on the day it launched can quietly drift out of step with reality if nobody’s watching for it. Ongoing monitoring and drift detection aren’t optional extras; they’re what keeps a model’s day-one accuracy from becoming a day-two liability.
Starting with the highest-value decision, not the whole dataset
The organisations that get this right don’t attempt to fix every reporting process at once. They identify the single highest-impact decision currently being made on incomplete or outdated evidence, whether that’s a demand forecast, a pricing call, or a risk assessment, and they build outward from there, starting with a fast, testable version of the model rather than a finished one. A concrete proposal for what a model would change about that one decision, what getting it right would be worth, and how quickly the first result could be tested is a far more useful starting point than a data strategy document that tries to cover everything and commits to nothing.
“More data is useless without the understanding of what it means and how to apply it. Running experiments to test hypotheses will give you the tracer bullets your decision making needs to really hit their mark.” – John O’Kennedy, Head of Bespoke Software & Solutions Architect, Warp Development
Ready to find the one decision in your business that a better model would change? Book a data diagnostic with Warp Development.
FAQs
Isn't this just business intelligence with a new name?
No. Business intelligence and dashboards summarise what has already happened. Machine learning models are built to inform what happens next, whether that’s a forecast, a price, or a risk score. The two are complementary, but confusing them is exactly what keeps organisations stuck with better-looking reports and no better decisions.
How do you stop a model from degrading after it's deployed?
Through ongoing monitoring and drift detection, checking the model’s predictions and inputs against what’s happening in the business on a continuing basis, not just at launch. A model is a living system, not a one-off delivery.
Where should we start if we're not sure the investment is worth it?
Start with a data diagnostic focused on one decision. Identify the single highest-impact call in the business that’s currently being made on weak evidence, and scope a concrete proposal for what a model would change about it, and what that change would be worth, before committing to anything wider.
