Most AI initiatives don't fail at the model. They fail three layers below it — in data nobody trusts, ownership nobody holds, and expectations nobody aligned. Here is how to find out before you spend.
An AI readiness assessment is a structured evaluation of whether your organisation can support AI initiatives that actually reach production. It scores you across five dimensions — data quality, infrastructure, talent, governance, and leadership alignment — identifies the specific gaps that would cause an AI project to stall, and produces a prioritised plan to close them. It is the difference between asking "which AI should we buy?" and asking the question that determines whether any of it works: "what has to be true about our data and our organisation before AI can deliver?"
We built our practice around that second question, because we kept meeting the same client: eighteen months into an AI programme, one impressive demo, nothing in production, and a data estate quietly disqualifying every use case on the list. The assessment exists so you can find those disqualifiers for the cost of a few weeks' work instead of a few quarters' burn. If you want the fast version first, our free AI Readiness Scorecard takes about fifteen minutes; this article explains what the full discipline looks like.
1. Data quality. The dimension that decides most outcomes. Not "do we have data?" — everyone has data — but: is it complete where it matters, consistent across systems, current enough for the decisions at hand, and documented well enough that a model's answer can be traced and defended? The practical tests are unglamorous: can two systems agree on how many active customers you have? Does "revenue" mean one thing or four? Is the field your use case depends on populated 98% of the time or 60%? AI amplifies whatever data quality you have — which is wonderful when it is high and catastrophic when it is not.
2. Infrastructure. Can data actually move to where AI can use it? This covers the state of your warehouse or lakehouse, whether pipelines are reliable or held together by one heroic engineer, whether the platforms you own (Fabric, Snowflake, Databricks, and their kin) are configured to serve governed data to models, and whether unstructured content — the documents a RAG system would read — is accessible and current rather than scattered across shared drives. The question is not whether you own modern tools; it is whether the path from source system to model is paved or improvised.
3. Talent. Readiness is not a headcount of data scientists. It is whether your team can build and maintain the plumbing (data engineering), define and defend the meaning (analytics engineering, semantic modelling), and operate what gets shipped (monitoring, retraining, incident response). Many organisations discover they are over-hired at the model layer and under-hired at the foundation layer — the ratio matters more than the total.
4. Governance. Who owns each critical data domain? Is there one governed definition per business metric, or does every dashboard freelance? Are access controls deliberate? Is there a policy for what data may be used to train or prompt a model, and who signs off? Governance is the dimension executives most want to skip and the one that most reliably kills production deployment — because a model no one is accountable for is a model legal will not let ship.
5. Leadership. The quiet one. Is there an agreed, written answer to why the organisation is pursuing AI and which business outcomes it serves? Are expectations calibrated — do sponsors understand that foundations come before pilots? Is there a single accountable executive, or is AI everyone's enthusiasm and no one's job? We have seen technically ready organisations stall entirely on this dimension, and technically mediocre ones succeed because leadership sequenced the work honestly. We explore a finer-grained version of this model — nine dimensions instead of five — in our companion piece on the nine dimensions of AI readiness.
Ticking two or more of these does not mean AI is off the table. It means the honest roadmap starts with foundations, and an assessment will tell you exactly which ones and in what order.
A formal assessment is evidence-based, not survey-based. The distinction matters: asking a team to self-rate their data quality produces optimism; profiling the actual tables produces facts. Ours runs in three passes.
Document and estate review. We read what exists — architecture diagrams, data dictionaries, governance policies, the backlog — and profile the data itself: completeness, consistency, and freshness on the domains that matter to your candidate use cases. This is where the gap between the org chart's view of the estate and the estate's actual condition becomes visible.
Structured interviews. Fifteen to twenty-five conversations across executives, data producers, and data consumers, asking the same core questions so the answers can be compared. The interviews surface what no document shows: the shadow spreadsheets, the workarounds, the metric disputes, and where leadership expectations sit relative to reality.
Scoring and gap analysis. Each dimension is scored against a defined maturity scale, with evidence attached to every score — a claim like "data quality: 2 of 5" arrives with the profiling results that justify it. The gaps are then ranked by two axes: how much they block the AI outcomes you care about, and how expensive they are to close. That ranking is what turns a diagnosis into a plan.
For a mid-sized organisation, the full exercise runs three to five weeks: roughly one week of document review and profiling, one to two weeks of interviews and technical inspection, and one to two weeks of scoring, analysis, and writing. Larger estates with many source systems take longer; a focused single-domain assessment can be faster.
You should insist on four deliverables. First, a scored readiness baseline — dimension by dimension, evidence attached, comparable when you re-run it in a year. Second, a gap analysis that names specifics: not "improve data quality" but "customer identity is inconsistent across CRM and billing; this blocks every customer-facing use case on your list." Third, a prioritised remediation roadmap, sequenced by dependency — which fixes unblock which use cases, in what order, at what rough cost. Fourth, a shortlist of one to three AI use cases you could credibly ship first, with the data prerequisites for each spelled out, so the AI programme starts from something real rather than something aspirational.
A useful quality test for any assessment you commission: if the final report could have been written without looking at your data, it was not an assessment — it was a brochure.
The assessment is the map, not the journey. What follows depends on what it finds. Organisations with strong foundations move directly into use-case delivery — the assessment simply de-risks the sequencing. Most organisations land in the middle: foundations need targeted work, and the roadmap interleaves it with visible delivery so the business sees progress while the plumbing improves. A minority discover foundational debt serious enough that the honest first move is an architecture and governance programme — a finding that stings for a week and saves a seven-figure disappointment.
Either way, the baseline becomes the management tool: re-score the dimensions every six to twelve months and readiness becomes a measured, improving quantity rather than a feeling. If you want to see where you stand before commissioning anything, start with the free AI Readiness Scorecard — fifteen minutes, scored against the same dimensions — and browse the rest of our free assessments for the data maturity and reporting-health equivalents.
A structured evaluation of whether an organisation can support AI initiatives that reach production. It scores the organisation across dimensions such as data quality, infrastructure, talent, governance, and leadership alignment, identifies the specific gaps that would cause an AI project to stall, and produces a prioritised remediation plan. It answers "what must be true before we invest?" rather than "which model should we buy?"
A self-serve scorecard takes about fifteen minutes and gives you a directional read. A formal assessment for a mid-sized organisation typically runs three to five weeks: one week of document review and data profiling, one to two weeks of stakeholder interviews and technical inspection, and one to two weeks of scoring, gap analysis, and roadmap writing. Larger estates with many source systems run longer.
Four things: a scored readiness baseline across each dimension with evidence attached; a gap analysis that names the specific blockers rather than offering generic advice; a prioritised remediation roadmap sequenced by dependency and cost; and a shortlist of one to three AI use cases the organisation could credibly ship first, with the data prerequisites for each spelled out.
You can, and many organisations do — which is why so many pilots stall. A pilot on ungoverned, inconsistent data usually produces a demo that cannot be trusted or scaled, and a failed pilot makes the second attempt politically harder. An assessment costs a fraction of a pilot and tells you whether the pilot will survive contact with your real data. If you are genuinely ready, the assessment confirms it quickly.
AI readiness is measurable: data quality, infrastructure, talent, governance, and leadership, each scored on evidence rather than optimism. If two of the warning signs above sound familiar, the cheapest next step is to measure before you spend. The free AI Readiness Scorecard gives you a scored baseline in fifteen minutes — start there, and let the gaps set the roadmap rather than the demo.