Your Data Will Never Be Ready. Start Anyway.

The advice is everywhere: “Don’t start your AI journey until your data is sorted.” It sounds prudent. It feels responsible.

And it guarantees you’ll never move.

Here’s what the data-first advocates rarely mention: 77% of organisations rate their data as average or worse for AI readiness. Only 12% report data of sufficient quality and accessibility for effective AI implementation. If everyone followed the “wait until ready” advice, nearly nine in ten organisations would remain permanently on the sidelines.

Some already have. S&P Global’s 2025 survey found that 42% of companies abandoned most of their AI initiatives last year, up from just 17% in 2024.

The Perception Gap

RSM’s 2025 AI Survey found that 92% of companies using generative AI encountered challenges during rollout.

Data quality was cited by 41% as a top issue.

These weren’t organisations that ignored their data. They were organisations that thought they were ready.

The gap between perceived readiness and actual readiness isn’t a measurement error. It’s a feature of how data quality works. You don’t know what “good enough” looks like until you try to use the data for something specific.

This is why the “sort your data first” advice fails. It asks organisations to solve a problem they can’t yet see clearly, using criteria they can’t yet define.

Scope Management, Not Perfection

ISG’s 2025 State of Enterprise AI Adoption report explicitly warns against what they call “boil the ocean” approaches: multi-year data transformation programmes promising to “get our data right” before tackling AI. The organisations making progress take a different path: target specific datasets tied to specific use cases.

The difference shows up in scope.

The Head of Commercial at a mid-sized publisher doesn’t need clean data across every system. She needs clean metadata for the 500 titles she’s testing in a recommendation engine. That’s a three-week data hygiene project, not a three-year transformation programme.

The Registrar at a university doesn’t need a unified data warehouse before piloting AI-assisted student enquiry triage. She needs accurate course information and admissions data for one faculty. That’s probably achievable before the next term starts.

What “Good Enough” Actually Means

Every AI use case has different data tolerances. Understanding this distinction determines whether your pilot succeeds or stalls.

High accuracy required: Anything involving financial decisions, regulatory compliance, or customer-facing recommendations. The data powering these workflows needs active governance. Errors here create liability.

Moderate accuracy acceptable: Internal productivity tools, content summarisation, first-draft generation. These use cases tolerate noise. A 90% accurate document summary still saves time; a 90% accurate credit decision creates problems.

Low accuracy sufficient (initially): Exploratory analysis, pattern identification, hypothesis generation. The AI is showing you where to look, not making the decision. You’ll validate what matters.

The Transformation Director working on process automation doesn’t need every field in the CRM to be pristine. She needs to know which fields her pilot touches and what accuracy threshold that workflow requires.

Start there. Fix that. Learn what the next project will need from the lessons of this one.

The Hidden Cost of Waiting

Informatica’s CDO Insights 2025 survey identifies data quality and readiness as a top obstacle to AI success, cited by 43% of respondents alongside technical maturity. This is often cited as evidence for the “data first” position.

Read it differently.

The organisations abandoning projects are the ones who started without understanding their data requirements for that specific use case. They didn’t fail because they started too early. They failed because they started without scope discipline.

Meanwhile, companies with strong data integration achieve 10.3x ROI from AI initiatives, compared with 3.7x for those with poor connectivity. That’s a nearly threefold difference. And connectivity doesn’t mean “perfect data everywhere.”

It means the right data, accessible, for the workflows that matter.

The cost of waiting isn’t just competitive positioning. It’s institutional learning. Every month spent perfecting data in the abstract is a month not spent discovering what your AI use cases actually require.

The Practical Path

The organisations gaining ground treat data quality as a use-case-specific investment, not an enterprise prerequisite.

Start with one workflow. Not “AI strategy” in the abstract — and if your board is still finding its feet, that’s fine; start iterating while the board learns. One process, one department, one bounded problem. The Commercial Director testing recommendation logic. The Operations Manager piloting document classification. Specificity creates clarity.

Map the data dependencies. What inputs does this workflow require? Which systems hold them? What’s the current state of those specific fields? This is a week’s work, not a programme.

Define your accuracy threshold. What quality level does this use case demand? A customer-facing tool and an internal triage assistant have different requirements. Build governance proportionate to risk.

Fix what needs fixing, for this use case, now. Targeted data hygiene for a bounded scope. Not a governance framework for every eventuality. The framework emerges from accumulated experience, not from planning workshops.

Expand deliberately. Each successful pilot teaches you what the next one will need. Your second AI project benefits from the data work done for your first. Capability compounds.

The Craft of Starting

There’s a reason the “sort your data first” advice persists. It feels safe. It positions the speaker as prudent. And it requires no immediate action, which is comfortable for everyone except the organisation’s future competitive position.

The alternative isn’t recklessness. It’s disciplined pragmatism.

Perfect data is a destination you’ll never reach. Good enough data, for a specific purpose, is how you start moving.

Worth considering: while your organisation debates data readiness, the competitor who picked one use case and fixed one dataset is already learning what works.

Related reading: I Built My Family a Chief of Staff. It Runs on a Raspberry Pi. - a practical example of starting with imperfect data sources and letting AI reason over the connections.