The AI Readiness Assessment

The interesting number in 2026 is not how many businesses use AI. It is the gap between how many use it and how many can point to what it earned.

Adoption is effectively settled. 86% of middle-market organizations have partially or fully integrated AI into operations, and 97% report being satisfied with the investment. Then there is the other set of numbers. MIT’s NANDA initiative found that 95% of generative AI pilots produced no measurable impact on profit and loss. An analysis of enterprise deployments put the share that never reach production at 88%. RAND found more than 80% of AI projects fail, roughly twice the failure rate of conventional IT work. Gartner expects organizations to abandon 60% of AI projects not supported by AI-ready data, and separately predicts that more than 40% of agentic AI projects will be canceled by the end of 2027.

Both sets are true at once. Companies are satisfied and most projects are not paying. The honest reading is that satisfaction is being measured against expectations that were never tied to a number, which is itself one of the failure modes.

A word on the 95%. That figure comes from 52 executive interviews, 153 survey responses, and 300 public deployments. It is directional, not definitive, and it has been fairly criticized for it. We cite it because four independent studies using different methods all land between 60% and 95%. The exact number is arguable. The pattern is not.

If AI projects failed for unpredictable reasons, an assessment like this would be pointless. They do not. RAND’s five root causes are: the problem was misunderstood, the data was inadequate, the technology came before the use case, the infrastructure could not support it, and the problem was simply too hard. Four of those five are knowable in an afternoon. None are about model quality.

The middle-market survey ranks the barriers the same way. Data quality issues, 53%. Integration challenges, 47%. Unclear ROI, 33%. Security and compliance, 33%. Notice what is absent from every one of these lists. Nobody says the models are not good enough.

The most common reason a pilot fails to survive contact with production is that the pilot ran on a clean, curated dataset that does not exist in the real business. The demo worked. The demo was the problem.

Five dimensions. Four questions each. Score every question 0 to 3.

Answer for one specific task, not for your company in general. Pick the single piece of recurring work you would most like to stop paying a person to do, and hold it in mind for all twenty questions. Readiness is not a property of a business, it is a property of a business and a particular task together. The same company can be completely ready for one thing and hopeless at another.

If you cannot name that task, stop here. That is your answer, and Dimension One explains what to do about it.

Score honestly. The scoring exists to tell you what to fix. Inflating it produces a nicer number and a worse decision. Where you are unsure between two scores, take the lower one.

Root cause number one is a misunderstood problem. This dimension is the cheapest to fix and the most often skipped.

1. 1

   Can you describe the task in one sentence, without using the word "and"?

   • 0I have a general area, not a task
   • 1I can describe it, but it takes a paragraph
   • 2One sentence with one "and" in it
   • 3One clean sentence

   Compound tasks are two tasks. Split them and score the harder one.

2. 2

   Is it done the same way every time?

   • 0Every case is different
   • 1Mostly judgment, some pattern
   • 2Mostly pattern, some judgment calls
   • 3Same steps every time, exceptions are rare and recognizable
3. 3

   Is the procedure written down?

   • 0It lives entirely in one person's head
   • 1Someone could reconstruct it from examples
   • 2Partial documentation, out of date
   • 3Written, current, someone new could follow it
4. 4

   If the person doing it left tomorrow, how long to get a replacement to full speed?

   • 0Months, and we would lose things in the handover
   • 1Six to eight weeks
   • 2Two to four weeks
   • 3Under two weeks

   A proxy for how much undocumented judgment the task carries. A task that takes three months to hand to a human will not hand cleanly to a system either.

Dimension 1 subtotal/ 12

The single most cited barrier, at 53%, and the subject of Gartner's 60% abandonment forecast. 63% of data leaders say they do not have, or do not know whether they have, the right data management practices for AI.

1. 5

   Where does the information this task needs actually live?

   • 0Email threads, people's memory, paper
   • 1Spreadsheets on individual machines
   • 2Shared systems, but spread across several
   • 3In systems of record, queryable
2. 6

   If you needed the last twelve months of records for this task, how long to produce them?

   • 0We could not, reliably
   • 1Days of manual assembly
   • 2Hours, with someone who knows where to look
   • 3Minutes, self-serve
3. 7

   How often is that data wrong, stale, or duplicated?

   • 0Constantly, and we work around it
   • 1Often enough that we check everything twice
   • 2Occasionally, and we usually catch it
   • 3Rarely, and there is a process when it happens
4. 8

   When two systems disagree, is there an agreed answer to which one is right?

   • 0They disagree often and nobody has decided
   • 1It depends who you ask
   • 2Informally understood
   • 3Explicit, documented, enforced

   Questions 6 and 7 together are the best single predictor in this assessment. A pilot built on data you had to hand-assemble will not survive production, where nobody hand-assembles anything.

Dimension 2 subtotal/ 12

Integration is the second barrier at 47%. Legacy systems and talent gaps tie at 28% as enterprise-wide inhibitors.

1. 9

   Do the systems this task touches have an API or a documented export?

   • 0No, or we have no idea
   • 1One does, the important one does not
   • 2Most do
   • 3All of them, and we have the credentials
2. 10

   How many separate systems does the task touch?

   • 0Six or more
   • 1Four or five
   • 2Two or three
   • 3One

   More systems is not disqualifying. It is the difference between a two-week deployment and a six-week one, and you should know which you are buying.

3. 11

   Has anyone successfully connected two of your systems before?

   • 0No, and attempts have failed
   • 1No, but nobody has tried
   • 2Yes, one working integration
   • 3Yes, several, and someone here understands them
4. 12

   Who controls access, and how fast can they grant it?

   • 0A vendor who charges for it, or nobody knows
   • 1External IT, weeks of lead time
   • 2Internal, days
   • 3Someone in this building, today

   Question 12 delays more deployments than any technical problem. Credentials are the critical path more often than code is.

Dimension 3 subtotal/ 12

85% of middle-market leaders agree that executives are more enthusiastic about AI than their employees are. That gap is where deployments go to die.

1. 13

   Is there one named person who owns this outcome and can decide?

   • 0It is a committee, or it is nobody
   • 1Someone owns it but cannot approve changes
   • 2Clear owner, needs sign-off for anything material
   • 3One owner who can decide and commit
2. 14

   How would the people doing this work today react?

   • 0They would fight it, openly or quietly
   • 1They would be nervous about their jobs
   • 2Neutral, wait and see
   • 3They have asked for this

   A 0 here is not fatal, but it is expensive, and it needs to be handled before deployment rather than discovered during it.

3. 15

   Are people already using AI tools here without approval?

   • 0Almost certainly, and we have no visibility
   • 1Probably, we have not asked
   • 2Yes, and we know roughly who and what
   • 3Yes, sanctioned, with a policy

   Counter-intuitive scoring. Widespread quiet usage scores low because it is a governance exposure, but it is genuinely good news about appetite. It also tells you exactly which tasks people find most painful, for free.

4. 16

   Who checks the output before it matters, and do they have the time?

   • 0Nobody, or the person who would has no capacity
   • 1Someone would need to make time
   • 2Clear reviewer, workable
   • 3Review already exists as part of the process

Dimension 4 subtotal/ 12

Unclear ROI is cited by 33% as a barrier to scaling. Misaligned success metrics appear in every failure analysis we reviewed.

1. 17

   Do you know what this task costs today, in hours and in money?

   • 0No idea
   • 1A rough guess
   • 2A reasonable estimate we could defend
   • 3We have measured it
2. 18

   What happens when it is done wrong?

   • 0Regulatory, safety, or a lost customer, immediately
   • 1Real money, hard to reverse
   • 2Annoying, recoverable within a day
   • 3Caught downstream before it matters

   High blast radius does not mean do not automate. It means the system needs a human approving before anything commits, which is a design decision, not a blocker.

3. 19

   What single number would have to move for this to be obviously worth it?

   • 0I could not name one
   • 1Something vague like efficiency
   • 2A specific metric, roughly
   • 3A specific metric, with a target and a date
4. 20

   Can you measure that number today, before anything changes?

   • 0No
   • 1We could construct it with effort
   • 2Yes, approximately
   • 3Yes, and we have the history

   If 19 and 20 both score 0 or 1, you will not be able to prove the deployment worked. This is the most common reason a project that succeeded gets cancelled anyway.

Dimension 5 subtotal/ 12

Total/ 60

Also check your lowest single dimension. A total of 45 with one dimension at 4 out of 12 is not a 45. The weak dimension governs. Data at 4 out of 12 will sink a deployment that every other column says is ready, which is precisely the Gartner finding: it is not the average that abandons projects, it is the data.

We wrote this assessment because it is the first hour of every engagement we run, and because the failure statistics above describe a market where most of the spending is wasted on preconditions nobody checked.

If you scored under 24, we are not the right call yet. Fix the three things in the first tier and come back. We would rather tell you that now than take your money and discover it in week three.

If you scored 24 or above, the natural next step is a conversation about the specific task you had in mind.

Two things we do differently, both straight from the research above. We do not charge setup or build fees, because the first month covers deployment. And every deployment we run puts a human approval step in front of anything that commits, because that is the difference between an agent and a liability.

1. RSM US Middle Market AI Survey, 2026. 1,030 respondents (827 US, 203 Canada), fielded March 5 to 16, 2026, margin of error ±3.1 points at 95% confidence. Adoption, satisfaction, ROI, barriers, inhibitors, governance, leadership enthusiasm gap.
2. MIT NANDA, "The GenAI Divide: State of AI in Business," 2025. 52 executive interviews, 153 survey responses, 300 public deployments. The 95% figure, build versus buy, back office versus front office, use case focus. Directional rather than definitive.
3. Iris.ai enterprise analysis, 2026. 88% of pilots never reach production.
4. RAND Corporation. More than 80% of AI projects fail, approximately twice the rate of conventional IT projects. Five root causes.
5. Gartner, February 2025. 60% of AI projects unsupported by AI-ready data abandoned through 2026. Underlying survey of 248 data management leaders, Q3 2024: 63% lack or are unsure of appropriate practices.
6. Gartner, June 2025. More than 40% of agentic AI projects cancelled by end of 2027. Agent washing, approximately 130 genuine agentic vendors. Based on a January 2025 poll of 3,412 webinar attendees.
7. Shadow AI research, 2026. 54% of employees using unsanctioned AI tools; 665 distinct AI tools observed generating traffic inside corporate networks.

Figures current as of August 2026. Where studies disagree we have said so rather than picking the most alarming number.