Skip to main content
How Can We Help?Contact

Is it the model, or the data underneath it?

A better model will not fix the data underneath it.

When an AI answer or a report cannot be trusted, the instinct is a bigger model or a new tool. Trace the answer first: if the source is stale, incomplete or defined three ways, no model will fix it. Then invest only where the evidence points.

Bring us one answer you don’t trust How the diagnostic works

Asked by CDOs, IT directors and executives who stopped trusting the numbers.

One AI answer, traced back to its source. Illustrative.

“How many orders did the Midwest region ship last quarter?”

The AI assistant38,800

Finance’s closed books42,000

Traced back, step by step

  1. The modelRead the warehouse correctlyRight
  2. The warehouseSound, no trail to the sourceSound
  3. The nightly loadFailed 2 of 30 nights, no alertBroken
  4. The source systemHolds all 42,000Right

The data, not the model.

It looks like a model problem. Check underneath first.

What leaders see, function by function, and what is usually underneath it.

  1. Finance

    “Three numbers for one figure, and a meeting to decide which is right.”

    UnderneathLoads that fail without an alert, and no line from the report back to the source.

  2. Sales and marketing

    “The assistant quotes last quarter’s prices.”

    UnderneathThe documents it reads are copies that nobody refreshes.

  3. Operations

    “The forecast drifts every time a source system changes.”

    UnderneathNo agreed format or schedule from each source, so changes arrive unannounced.

  4. Customer service

    “It cannot see the account the customer is calling about.”

    UnderneathThe data sits in a system nobody connected, or behind access rules nobody set up for it.

  5. People

    “HR and finance report different headcounts.”

    UnderneathTwo definitions of an employee, and nobody who owns the difference.

None of these is fixed by a better model. Left alone, each new AI use case finds the same gaps and pays to work around them again.

Once you know it is the data, there are three ways forward.

The choice depends on how many decisions rely on the same data, how contained the gap is, and how long the use case can wait. The three can be combined: narrow the use case now while a shared fix goes in behind it.

Fix what is shared
When the same data problems affect several important decisions or use cases.
What it takes Repair the common foundation once, keeping what already works, instead of paying to work around it again and again.
Work around a contained gap
When the gap is limited and the business case is strong enough not to wait.
What it takes A controlled workaround with a named owner and a date to fix the cause, so it does not become permanent.
Narrow the use case
When the data cannot carry the full ambition yet.
What it takes Deliver the valuable part the data supports, prove the outcome, and expand when the evidence justifies it.

Not deciding is the expensive option: the gaps stay, and every project pays for them separately.

Which of the three is yours? Bring one answer or number you do not trust. We review it with you and agree what it would take to confirm the cause.

Find out which of the three is yours

Or check your data maturity first: the free Data Maturity assessment, fifteen questions, about six minutes, no email required. It gives a first view of how ready your data is overall; it cannot tell you why one answer is wrong.

What changed when the data underneath was brought together.

Every figure is an operating result: a change in the work, as the client reported it.

A global asset manager, investment and distribution

The data was already there. Each team was reading it separately.

  • 25%fewer manual comparison stepsOperating result
  • 15%faster to find the portfolio that fitsOperating result

Portfolio, risk, mandate and performance data sat in one platform, but each team read its own view, so the best fit for a client still meant comparing by hand. The data was consolidated into one decision-ready view, without breaking its governance or lineage, and every team now works from the same ranked list.

Read the case study

15%better data quality at a health organization, after a data and AI literacy program for its staff. Not all of it is technology.Read the case study

When the problem is shared

Diagnose first, then fix only what is broken.

A diagnostic of your data estate, then a foundation built from what survives it: what works is kept, what is broken is fixed, what is missing is built. Often the estate needs repairs, not a rebuild.

The Data Platform Accelerator. Two to four weeks, on the platform you run or the one the diagnostic recommends. Fixed scope, fixed fee agreed in writing first.

Scope the Accelerator Explore the Data Platform Accelerator

What you get

  • The diagnostic: what you have, what it can support, and what is missing
  • A working foundation, with the access rules it needs from day one
  • The first data flows live, built the way the rest will be
  • A runbook owned by the team that will run it

Other ways in

  • Not sure it is the data?Take the free Data Maturity assessmentA first view of overall data readiness in about six minutes. No email required.
  • It may be bigger than the data.AI & Data Assessment SprintSix readiness dimensions scored against your systems in two weeks.
  • The use case can be narrowed to what the data supports.AI & Agent PilotOne use case on your data, measured, ending in a decision to scale, revise or stop.
  • It is a full platform program.Data Engineering & PlatformsThe platform built out across sources and teams, after the pattern is proven.

The practices behind this page: Data Engineering & Platforms, BI & Data Visualization, Strategy & Governance and AI & Machine Learning.

Questions data leaders ask first.

How do we tell whether it is the model or the data?

Trace a handful of answers you do not trust back to their source. If the source was wrong, stale or missing, it is the data. If the source was right, the cause can still sit in between: how the data was moved or reshaped, what the system was allowed to see, how a term is defined, or the model and its instructions. Tracing shows which, and it is where the diagnostic starts.

Do we need a new data platform?

Often not. The diagnostic decides, part by part, what to keep, what to fix and what to build. Often the estate needs repairs and one missing pattern, not a replacement.

Can we start on AI before the data is fixed?

Yes, if the use case is narrowed to what the data supports, or the gap is contained where the answer is made. Those are two of the three ways forward.

Who should own the data once it is fixed?

A named owner in the business for each source and each definition, not the data team alone. Ownership is one of the tests the diagnostic runs on every part of the estate, so a gap there is found before it undoes the fix.

Next: question 5 of 5, Adoption & WorkforceYou launched it. Are people working differently?All five questions

Bring us one answer you don’t trust. We’ll show you where to look.

Thirty minutes with the person who would lead the work: a senior AI and data strategist who has taken programs from strategy to systems running in production, in more than one industry. We review the discrepancy with you and agree what it would take to confirm the cause. If it turns out to be the model, we will say so.

Book 30 minutes on your data

Answers you can’t trust? Thirty minutes on one answer you don’t trust, with the person who would lead the work.

Trace one answer with us