An AI readiness assessment is good at scoring the parts of your company that were probably never the problem. Almost every version on the market rates six or seven dimensions, usually some mix of strategy, data, governance, technology, infrastructure, people, and use cases, then hands back a number out of 100 and a ranked list of gaps. The data and governance sections are worth the time. What the report will not tell you is the one thing that decides whether AI enablement works at a company of 50 to 500 people. That is whether the employees holding the licenses can get usable work out of them on a Tuesday afternoon. Here is how to read one of these reports honestly: which sections to trust, which to discount, and what to measure yourself.
What an AI readiness assessment measures
The direct answer: it measures organizational capacity to support AI across a handful of named dimensions, and it outputs a baseline plus a gap list. The names vary by author; the set barely does. Microsoft’s self-serve assessment, published on Microsoft Learn, scores seven pillars: business strategy, AI governance and security, data foundations, AI strategy and experience, organization and culture, infrastructure for AI, and model management. Cisco’s AI Readiness Index, which surveyed more than 8,000 business leaders across 30 markets, groups the same ground into six: strategy, data, governance, human talent, culture, and infrastructure. Third-party rubrics shuffle the labels and add “use case and value.”
Read that list as a buyer and the shape is clear. Five of the six or seven dimensions ask about systems, records, policy, and plumbing. One asks about people.
Know the population-level results before you commission your own. Cisco has reported for three straight years that roughly 13% of organizations qualify as fully ready, a figure that has not moved. In its 2025 round, 54% said their infrastructure could not scale for AI workloads and only 32% had any process for judging whether an AI investment succeeded or failed. Assume in advance that you will land in the unprepared-to-developing range with nearly everyone else, and decide now what you would do differently knowing that.
The scoring math, and what the weights tell you
Most published rubrics work the same way: rate each dimension 1 to 5 against a written standard, multiply by a weight, sum, and scale to 0–100. The most transparent public example weights it this way.
| Dimension | Weight |
|---|---|
| Data foundations | 25% |
| Strategy and leadership | 20% |
| Governance and compliance | 20% |
| People, skills, and culture | 15% |
| Technology and infrastructure | 10% |
| Use case and value | 10% |
Scores then fall into bands, typically unprepared, planning, developing, implemented, and embedded.
Two things follow from that table. First, a composite score is nearly useless alone. A 58 built on strong data and weak people is a different company from a 58 built the other way around, and both get told they are “developing.” Ask for the pillar-level scores and ignore the headline number.
Second, look at where people sit. Human capability carries 15%, less than data and less than governance, even though the published research on why AI projects fail points at workflows and learning rather than at plumbing. That is not one author’s oversight. It is the structure of the whole category, and it is the gap the case against treating license rollout as an AI strategy was written about.
Self-reported comfort is not observed skill
Here is the common limit in the reports clients bring us: the people dimension is scored by asking leaders to rate their own team. A VP ticks “3 of 5, moderate capability” for a group of twelve people whose actual work with these tools they have never watched.
That estimate is reliably wrong, and the direction of the error is consistent. MIT’s Project NANDA study of enterprise AI, published in 2025 and drawing on 52 executive interviews, 153 leader surveys, and 300 public deployments, found that 95% of generative AI pilots produced no measurable impact on the profit and loss statement. The cause it identified was not model quality. It was a learning and workflow gap: tools that never entered the work they were bought to change.
BCG’s AI at Work 2025 survey of 10,635 employees across 11 countries puts numbers on the learning half. Only 36% think their training is enough, and 18% of people who already use these tools regularly report no training at all. The line between a regular user and a lapsed one sits near five hours of instruction: 79% above that line were regular users, against 67% below it. Only a quarter of frontline staff say they crossed it.
In our own enablement work the pattern repeats. A team arrives with a clean readiness report, confident the people box is fine, and in the first working session several participants cannot get a usable first draft out of a tool they have been licensed on since last year. Nobody lied on the questionnaire. The questionnaire never asked a question that could have surfaced it.
The work itself never gets inventoried
The second blind spot is narrower and easier to fix. No readiness framework asks you to list the actual recurring documents your team produces.
Scoring “use case and value” at 10% means asking whether use cases exist and whether anyone estimated a return. It does not mean naming the weekly client update, the scope memo, the RFP response, the board summary, or the renewal notice, which is the altitude where these tools either save time or do not. Cisco’s data shows how rarely companies get down there: only 32% had identified which human tasks they meant to hand to AI at all.
This matters commercially, because the task list is what a workshop gets built from. Training on a generic demo and training on your own renewal notice produce the same slide deck and nothing like the same Monday morning. A readiness score cannot tell the two apart, so it cannot tell you what your enablement spend should buy.
Who wrote the framework, and what they sell
Treat authorship as part of the data. Pillar selection and pillar weighting are commercial decisions, and you can usually read the author’s product line directly off the rubric.
The data-catalog vendor behind the most rigorous public rubric weights data foundations heaviest, at 25%. Microsoft’s self-serve assessment includes model management and infrastructure for AI, two pillars that only yield real answers if you are building on a cloud platform. Google Cloud’s AI Readiness program runs two to three weeks of surveys and joint workshops, and its stated deliverable is a roadmap routing toward Google tools, solutions, and partner services; pricing is quote-only. The free quiz on a consultancy’s site asks eight to fifteen questions and holds your result behind an email form, because the result is not the product. The meeting is.
None of this makes the assessments dishonest. The frameworks are real and the diagnostic questions are often good. It does mean a readiness report tells you what its author is equipped to sell, and that the dimension nobody sells, individual human fluency, is the dimension nobody weights. We carry the same bias in the other direction, which is why this page says to count things yourself before hiring anyone, including us.
The AI enablement questions a readiness score leaves open
Three questions decide whether license spend turns into capability. They are the three a corporate AI enablement program has to answer, and a standard assessment answers none of them.
Can each role produce an acceptable output on a real document? Not “is the team comfortable.” Can your operations coordinator draft the weekly client update to a standard you would send, using the actual last-quarter file, inside the time she has? That is a pass or fail you can observe in twenty minutes.
Which recurring tasks are in scope, and which are explicitly not? A named list of ten, with an owner per item. Without it, enablement work has no target and no before-and-after.
Who owns this after the readout? At a company of this size there is no Chief AI Officer. The work lands on a COO, a chief of staff, or a head of L&D as a second job. A roadmap handed to a person with no hours allocated to it is a document, not a plan. The route from license rollout to measurable team capability covers how to staff that ownership without creating a new role.
A one-week AI enablement diagnostic you can run first
Before paying for an outside score, run this. It costs about six hours and tells you more than most reports.
- Pull the license usage data. Your admin console shows weekly active users per seat. Compare it against who you pay for. This single number is more honest than any survey response about adoption.
- List ten recurring work products. Documents or outputs your team produces weekly or monthly, with a rough time cost on each.
- Watch three people work. Twenty minutes each, one real task from the list, their own file, no coaching. Note where they stop, what they retype by hand, and what they do not know the tool can do.
- Write one page of what good looks like. For two of the ten tasks, describe an acceptable AI-assisted output in plain terms. Most teams have never agreed on this, which is why five people invent five standards.
- Set a baseline. Minutes to first draft and number of revision rounds, measured today, on those two tasks. Everything you do later gets measured against these two numbers.
If step one shows half your seats idle and step three shows capable people stuck at the same point, you have the diagnosis without a maturity band. McKinsey’s 2025 state-of-AI survey found only 39% of organizations attribute any earnings impact to AI, and most of those put it below 5% of EBIT. Measuring two tasks well beats scoring seven pillars loosely.
When a paid assessment earns its fee
A third-party readiness assessment usually prices between roughly $2,000 and $35,000, with large-firm engagements well past that. It earns the money in four cases:
- You are in a regulated industry where a documented control review has independent value.
- The deployment spans several systems of record and someone has to map the integration risk.
- Your board or an investor wants an outside read rather than management’s own.
- You have already tried, failed, and do not know why.
Skip it if you know which workflow you want to improve and have budget to improve it. In that case the money is better spent on the thing the assessment would have recommended. Hands-on fluency workshops sit in a $2,000 to $15,000 market range and a custom automation build runs $25,000 to $150,000, so a mid-priced assessment can quietly eat a third of what the work itself costs. Before signing either, read what AI enablement services include and what they quietly exclude. The shortlist of providers working with mid-market companies shows which ones fold the assessment into delivery rather than billing it on its own.
One rule: never buy an assessment from a firm that cannot deliver the fix, and never buy one from a firm that can only deliver one kind of fix.
Frequently Asked Questions
What is an AI readiness assessment?
A structured evaluation of whether your organization’s data, systems, policies, people, and use cases can support AI, scored against a written rubric. The standard output is a baseline score, a maturity band, and a ranked gap list with recommendations.
What counts as a good AI readiness score?
The most detailed public rubric treats 66–85 as implemented and 86 or above as embedded, and most organizations land between 26 and 65. Treat the composite as noise and read the pillar scores. A 70 carried entirely by governance and data tells you nothing good about whether your team can work with these tools.
How much should an AI readiness assessment cost?
Third-party engagements commonly run from about $2,000 to $35,000 by scope and document-review depth; large-consultancy versions go higher. Vendor self-serve assessments are usually free. If a quote approaches the cost of the fix itself, renegotiate the scope.
How long does one take?
A free self-serve questionnaire takes under an hour; Microsoft’s is scoped at about 45 minutes. A paid engagement usually runs two to four weeks, with Google Cloud’s program stated at two to three weeks of surveys and one or two joint workshops. The internal version described above takes a week.
Are the free self-serve assessments any good?
Useful once, as a vocabulary exercise, if you have never thought about this in a structured way. They are self-scored, they verify nothing, and the ones hosted by consultancies exist to start a sales conversation. Use them to learn the dimensions, not to set a budget.
What is the difference between a readiness assessment and an AI audit?
A readiness assessment looks forward and asks whether you can support AI. An audit looks backward at systems already in production and asks whether they behave as documented, with evidence and control testing. Buying the first when your board asked for the second is a common, costly mistake.
Who should be in the room?
The budget owner, whoever administers the licenses, one person who owns the data, and two people who do the frontline work daily. Most assessments are filled in entirely by managers, which is how the people dimension ends up overstated.
Does any assessment test whether staff can use the tools?
None of the mainstream frameworks test observed capability. They ask leaders to rate team skill on a scale, which measures a manager’s impression. If you want the real answer, watch three employees complete a real task on a real file without help, and score what you see.
How often should we re-score?
Every six to twelve months is enough for the structural pillars, since data governance and infrastructure move slowly. Re-take the capability measures within 60 to 90 days of any training, against the same two baseline tasks, or you will never know whether the spend did anything.
Where to Start
A readiness assessment is a reasonable purchase for a specific set of reasons and a poor default. If what you need to know is whether the licenses you already pay for are producing work, the fastest path is the five-step diagnostic above, run on your own data this week, before any outside engagement puts a number on your maturity.
If you would rather talk it through, a free AI-readiness call is a working session on your own task list: which work products are worth training against, what an acceptable output looks like, and what you can measure in 60 days. No score, no maturity band. Book a free AI-readiness call, or start with the route from license rollout to measurable team capability if you want the framework first.
Arthur Wandzel
