Is an AI readiness assessment worth it, or just a delay?

An AI readiness assessment is worth it if it takes minutes and ends with a number you can argue about. It is not worth it if it takes four weeks, costs five figures and ends with a slide deck that recommends a follow-on engagement. Same label, two entirely different objects. Judge them on duration and output, not on the word “assessment”.

So the honest answer to “is it a stalling tactic” is: sometimes, yes, and you can tell in advance. If the assessment has a project plan, a kickoff call and a steering group, it is a project. Projects consume the thing you are short of, which is not money. It is attention and political capital. A short self-assessment that returns a score and a band per pillar costs neither. You take it before a meeting, not instead of one.

The objection, stated properly

Charlene Li, co-author of Winning with AI, made this argument publicly and it deserves a straight response rather than a vendor shrug. She describes a scene most of us recognise: a leadership team decides to get serious about AI, and someone, usually a consultant, sometimes an internal champion, says “first we need to know where we stand, let’s do a readiness assessment.” Six weeks later there is a maturity matrix and no working software.

Her rebuttal of the “we need a baseline” justification is the sharpest part. She argues the most revealing baseline emerges about two weeks into actual implementation, because doing the work exposes realities no survey can capture.

That is correct. I have watched it happen for twenty-five years in forms that had nothing to do with AI. The CRM migration where the data quality problem only became visible when someone tried to load the records. The ERP rollout where nobody knew a particular department had been running a parallel spreadsheet process for six years until the parallel process broke. You find out what is actually true by touching the system. A survey asks what people believe is true, which is a different and usually more flattering thing.

So concede it. A self-reported questionnaire cannot tell you whether your customer records are clean. It can only tell you how confident your leadership is that they are clean, which is not the same data point, and the gap between those two numbers is often the most interesting thing in the business.

Where the argument breaks

It breaks on the word “instead”.

The case against assessments assumes assessment and implementation compete for the same slot. For a four-week engagement, they do. For seven minutes, they cannot. Nobody has ever failed to ship a pilot because a director spent seven minutes answering questions on a Tuesday.

It also assumes the alternative is starting work, and that starting work is self-evidently better. The evidence on that is thin. In Infosys research covering more than 1,000 senior executives at large US companies, 72% said fewer than 25% of their AI pilots had scaled to full production while delivering their intended ROI. These are not companies that failed to start. These are companies that started, repeatedly, and produced a graveyard.

PwC’s 29th Global CEO Survey, which covers 4,454 CEOs across 95 countries, found 56% achieved neither revenue growth nor cost reduction from AI in the previous twelve months. Only 12% got both. What separated the 12% was not that they moved faster. It was defined roadmaps, responsible AI processes and deployment at enterprise scale rather than in corners.

“Just start” has been tested at scale. The results are publicly available and they are poor. The problem with most AI programmes is not insufficient starting. It is that each start begins from an assumption about the business that nobody checked, and the assumption turns out to be wrong in the same place every time. Usually the data. Sometimes the process. Often the fact that nobody owns the outcome.

The test: duration and output

Here is the practical distinction, because “readiness assessment” now covers two things that share nothing but a name.

A consultancy diagnostic runs weeks. Accenture’s version examines strategy, workforce, workflows, risk and governance, data, engineering, operations and ecosystem. Deloitte assesses maturity across six pillars. These are real pieces of work and they have a place, usually in an organisation large enough that the map genuinely is the hard part. What they also have is a commercial logic. The diagnostic is the front end of an engagement. That is not a slur, it is how the model works, and you should price it accordingly.

A self-assessment runs minutes. It is thirty questions, you answer them yourself, and it ends. There is no kickoff, no steering group, no follow-on scope. Our free audit takes 7 minutes and returns a score out of 120, a band for each of six pillars, and a short synopsis. No sales call is booked at the end unless you ask for one.

If you want the test in one line: can a director complete it alone, without telling anyone, before a meeting? If yes, it cannot be a stalling tactic, because stalling requires something to point at while nothing happens. If no, ask what the deliverable is and when it arrives, and whether the answer is a document or a decision.

What a seven-minute assessment can actually tell you

Not everything. Let me be precise about the limits, because overclaiming here would prove Li’s point for her.

It does not audit your systems, your data or your code. Every answer is self-reported by the person taking it. It does not review what tools you have already bought or whether anyone still uses them. At the free tier it does not give you an action plan. It is not a call, a consultation or a demo.

What it does is force six specific questions into the open, and the value is in the specificity. The Data pillar opens with this one: if you needed a clean accurate list of your best 500 customers right now, how confident are you in what you would get? That question does three things at once. It is answerable in five seconds. It is embarrassing if the honest answer is “not very”. And it is verifiable within an afternoon, because you can ask for the list and time how long it takes to arrive. The gap between what you ticked and what shows up in your inbox is your real baseline, and it took a morning rather than two weeks of implementation to find.

The Process pillar asks how many of your core processes are documented well enough that a new starter could follow them without asking anyone. The People pillar asks how actively you personally are modelling AI use rather than delegating it. Governance asks who is responsible. These are not maturity-matrix abstractions. They are questions with a name or a number as the answer, and the uncomfortable ones are uncomfortable because you already know the answer and have been managing around it.

On ownership specifically, McKinsey found only 28% of organisations place direct AI governance ownership with the CEO, and 17% with the board. If you cannot name a single person, the assessment has just told you something that two weeks of implementation would also have told you, but later and with a vendor invoice attached. The question of who should own AI outcomes in a company does not get easier once a pilot is already running and three departments each believe a different person is accountable.

Self-reported data has a known failure mode, and it is useful

The obvious attack on any questionnaire is that people flatter themselves. True. But the direction and size of the flattery is itself a finding.

Schellman’s 2026 State of AI Governance Report found 74% of organisations believed they could pass an AI compliance audit today, while only 27% described their governance programmes as fully mature. Those two numbers come from the same respondents. People are confident about the test and unsure about the substance, which is a specific, diagnosable condition rather than general vagueness. It is also why the difference between AI governance and compliance matters more than most boards think.

There is a second reason the self-reported number is worth having. A survey of more than 500 technology leaders found 79% feel pressure to overstate AI progress, with nearly half saying that pressure comes from the C-suite or board. If the official story inside your business is optimistic by default, a private score that only you see is one of the few mechanisms that produces an unperformed answer. Nobody is watching you take it.

The actual sequence

Assess and act are not alternatives. The sequence that works looks like this.

Take the short assessment on a Tuesday morning. Seven minutes. Look at which pillar came back lowest. Then, the same week, run the physical test on that pillar: ask for the 500-customer list, or ask your three operations people to describe the same process, or ask who signs off an AI tool purchase. Compare the tick to the reality. Now start the pilot, with one known weakness accounted for rather than discovered in week three.

That is roughly the opposite of a stalling tactic. It is a way of making the first two weeks of implementation less surprising, which is the only honest argument for doing any of this. If you want the longer version of how to run that comparison, measuring AI readiness before you spend anything covers the physical tests in more detail.

The failure mode Li describes is real, and if your assessment has a Gantt chart she is describing you. Cancel it. But a score out of 120 that arrives before lunch does not postpone anything. It just tells you which conversation to have on Monday, and with whom.