An AI capability audit is a paid diagnostic that tells you which of your people can do which parts of their job with an AI assistant, and which cannot yet. That is the whole product. The rest of what a vendor packs around it is context.
Quotes for that product run from $1,500 to over a million dollars today, and the spread is not a quality signal. It is a scope signal, because the words carry no agreed meaning. Below is the line-item breakdown of what can sit inside one, the arithmetic for checking whether the number on your desk is fair, and the two items that decide whether you bought a measurement or a slide deck.
Why one name carries twelve different prices
Search the term and you get a dozen consultancy pages, each with a confident three-tier price table. No two tables agree.
OperatorZero publishes a flat $2,500 per workflow on a five-business-day turnaround, and says plainly that the audit does not review models or governance. Elevates puts a mid-market assessment at $5,000 to $15,000 against a six-week calendar, and enterprise work at $15,000 to $50,000 and up over six to twelve weeks. Aries Consulting Group headlines $2,000 to $30,000 and up. Pertama Partners, whose figures are in Singapore dollars and will mislead a US reader who skims, quotes SGD 40,000 to 85,000 for 100 to 1,000 staff, delivered in three to five weeks by two or three consultants. LaunchReady, the only one of the set treating “capability audit” as a term of art, puts specialist workforce firms at $19,500 to $96,000 per cycle and the Big Four at $250,000 to $1 million for six months.
Each of those is vendor list pricing, not market measurement. None of them is lying. They are selling unlike things under one name.
That matters because this is a buying decision made almost entirely out of sellers’ material. No analyst covers the category, no trade body defines the product, no public benchmark exists. The only defense left to a buyer is to stop comparing prices and start comparing scopes.
The six line items a scope can hold
Across the proposals we have read and the scopes posted above, an audit of this kind is built from six parts. A quote is a pick from this list, and the price follows the pick.
| Line item | What it produces | Typical analyst time |
|---|---|---|
| Scoping call and stakeholder map | Who is in scope, which functions, who signs off | Half a day to a day |
| Self-report survey | A maturity score, often 1 to 5, by person or function | A day to build, a day to chase |
| Stakeholder interviews | What leaders believe is happening | Two to three days for six to ten interviews |
| Task-based assessment | A score per person against a named proficiency framework | Two days to design, a day to run |
| Observation of real work | A record of what a person actually does with an assistant on a live task | Two days for a sample of ten to fifteen |
| Report, plan and readout | The document, the recommendations, the meeting | Three to four days |
LaunchReady’s own shape is a useful reference at the output end: an 8 to 15 page report in five sections, an assessment that runs under ten minutes per team member, one to two days of analyst time to write it up. That is a real description of a real deliverable, and a reminder of how small the measurement part can be.
Note what is missing from the list. A tool-stack inventory, a data-governance review and an infrastructure assessment all appear on pages labeled “capability audit,” and none of them measure a capability. They measure the building. Those belong in a readiness review, a separate purchase with its own buyer, and mixing the two is the most common reason a report lands on a desk and stays there.
Only two of the six measure capability
Items four and five. The task-based assessment and the observation of real work. The rest is scaffolding, opinion, or writing.
This is the substitution to watch for, because it is nearly universal. A vendor sells a capability audit, then delivers a maturity score built from a questionnaire that asked people to rate their own confidence with AI tools. A self-rating measures nerve, not skill, and the two part company in a known direction.
The Wharton Human-AI Research group and GBK Collective published The Adaptive Advantage on October 13, 2026, from a survey of roughly 1,400 full-time employees at US companies with 1,000 or more staff and over $50 million in revenue, fielded between July 9 and July 29. It included 409 entry- and mid-level employees alongside the leaders, which is what makes it useful here. On training adequacy, 78% of leaders said employees were adequately trained. Among employees, 56% agreed. Same companies, same question, a 22-point gap.
An audit built on self-report inherits that gap. An audit built on observation settles it, and observation is the line item that costs real money, which is why it is the first one quietly dropped when a vendor needs to hit a price.
So there is one test for any proposal in front of you. Does the report contain a per-person, per-task observation of someone doing their own work, or a score? If it is a score, you are buying a survey instrument with a consulting wrapper. That can be worth $5,000. It is not worth $40,000, and it will not tell you what to teach on Monday.
Stanford HAI’s 2026 AI Index put organizational AI use at 88%, up from 71% the year before, while agent deployment stayed in single digits across nearly every business function. Access rose. Applied skill did not keep pace. An audit that only confirms access has measured the easy half.
Price it from analyst days, not from headcount
Headcount bands are the wrong unit, because an audit’s cost is driven by sample size and depth, not by how many people are on the payroll. Price it from the bottom.
Posted rates across this set of pages agree: independents at $150 to $350 an hour, mid-tier firms at $300 to $500, Big Four and the strategy houses at $500 to $1,000 and up. Take a 150-person company auditing three functions, sixty people in scope.
A full scope using all six line items lands around twelve analyst days, which is roughly 96 billable hours. At independent rates that is $14,000 to $34,000. At mid-tier firm rates it is $29,000 to $48,000.
Now run the mid-market band backward. A $12,000 fixed fee at mid-tier rates buys 24 to 40 hours, which is three to five working days. Five days covers a kickoff, a survey, four interviews and a report. It does not cover watching anyone work, because observation alone is two of those days for a sample of twelve.
That arithmetic is the most useful thing a buyer can do with a quote, and it dissolves the apparent contradiction in the market pricing. The cheap end of the published range is not a discount on the same product. It is another product, and only one of the two measures capability.
The logic runs in reverse too. If a quote arrives at $85,000 for sixty people, ask what the extra 30 analyst days buy. Sometimes it is a defensible sample size. Sometimes it is a second partner on the call.
The cost that never reaches the invoice
Your own payroll. It appears on no vendor page in this category, and it is often what separates a quote that pencils from one that does not.
Count the hours the audit takes from your side of the table, same sixty-person scope: eight stakeholder interviews at an hour each, sixty assessments at twenty minutes, twelve observed sessions at forty-five minutes, fifteen hours for whoever runs it in-house, and ten people in a ninety-minute readout. About 67 hours.
The Bureau of Labor Statistics put employer compensation costs for private-industry workers at $46.89 per hour worked in June 2026, with wages and salaries at $32.82 and benefits at $14.07. For full-time workers the figure is $54.00. At the full-time rate, 67 hours is roughly $3,600, and that is a floor rather than an estimate, because the people an audit interviews are managers and senior staff whose compensation costs sit above the all-worker average.
On a $12,000 engagement that adds about 30% to true cost. On a $2,500 single-workflow audit, where client hours are smaller but far from zero, the share is larger still. Those hours come out of the same operating budget, and leaving them out is what makes a cheap audit look free. The same omission distorts training comparisons, which is why quotes belong in a common denominator before you rank them, as we work through in comparing AI training quotes on cost per hour saved.
Three line items to strike
Not every part of a standard scope earns its place at a company of 50 to 500 people. Three are worth cutting.
The tool-stack inventory. Your admin consoles already hold it. Seat counts, active users, last-login dates and license spend come out of your identity provider and vendor dashboards in an afternoon. Paying a consultant $2,400 a day to export reports you own is the commonest padding in these proposals. Hand over the export and buy a day of analysis instead.
The long roadmap. Pertama’s own list of outputs includes roadmaps running one to four years. Assistant features and pricing change every quarter. A four-year plan built around tools that will be replaced twice in that window has a shelf life shorter than its horizon. Buy ninety days of sequencing and plan to re-cut it.
Unnamed industry benchmarking. If a vendor offers to score you against their own sample, ask how many companies are in it, of what size, on what instrument, in what year. A sample they will not describe is not a benchmark.
Spend what you save on a larger observation sample and a measurement plan you can re-run yourself in six months. Capability only means something as a trend, which is the case the measurement playbook argues at length.
Ask whether the fee is credited against build work
OperatorZero states that its $2,500 audit fee is credited toward future builds or advisory work. That disclosure is more useful than it looks, and most of the field does not make it.
A credited fee means the audit is an entry point to a larger engagement. Nothing is wrong with that structure, and we use a version of it ourselves: a free AI-readiness call is a short diagnostic we expect, some of the time, to lead to work. What matters is what the structure does to the findings. An audit priced as a door-opener has a commercial interest in recommending whatever the vendor sells next.
Ask the question before the report arrives. Is this fee credited against anything? If yes, you have a scoping exercise with a deliverable attached, and should read its advice the way you read any proposal. If no, you bought an outside opinion, and should expect to pay outside-opinion rates.
Neither answer disqualifies a vendor. A refusal to answer does.
Frequently asked questions
Is an AI capability audit the same as an AI readiness assessment?
They are sold as if they were, and they are not. A capability audit measures whether named people can do named tasks with an assistant. A readiness assessment measures whether the organization has the data, tooling, policy and governance to support AI work. You can be fully ready and have no capability, which is the common case once licenses are bought and nothing changes.
How much should a 150-person company budget?
If you want observation in scope, budget $15,000 to $35,000 at independent consultant rates and 10 to 14 analyst days of effort, plus about 60 to 80 hours of your own people’s time. If a survey-based maturity read is enough, $5,000 to $15,000 is the published band and it is realistic for that narrower product. Quotes well outside those bands are not necessarily wrong, but they need explaining.
What should the deliverable actually contain?
A per-person or per-role record of observed task performance, the sample it came from, the gaps ranked by how much work they block, a 90-day plan naming who learns what, and a protocol you can re-run without the vendor. Without that protocol the audit cannot be repeated, and a one-off reading of capability is close to useless.
Can we run the audit ourselves?
Partly. You can capture a baseline of task times and tool usage without help, and you should, because it costs a few hours and no quote is readable without it. Observation is the hard part in-house, since people work differently in front of a colleague who will see them again than in front of an outside analyst. Mixed approaches work: internal baseline, external observation.
How long does it take?
Two to six weeks for fieldwork and report at this size, mostly depending on how fast your people make time for interviews. Elevates publishes six weeks for a mid-market engagement, which is a fair planning figure. LaunchReady’s 90-day cycle puts a six-week training program between two assessment points, which is a program with audits at each end, not an audit.
Why do vendors quote such different prices for the same words?
The words have no settled scope, so each vendor prices the version they sell. The spread tracks which of the six line items are in and how large the sample is. Two quotes $30,000 apart often describe unlike fieldwork, and the gap closes once both list their line items.
Does the audit tell us which assistant to standardize on?
Only in passing, and be wary of any vendor who leads with that promise. The audit measures what people can do, not which product is better. If one tool shows up in the observations as the one people reach for under time pressure, that is a finding. A tool pick offered without that evidence is a preference.
Should we audit before or after a training program?
Before, then again after. One reading gives you a snapshot with nothing to compare it to, and this is where most AI enablement spending loses its evidence trail. One baseline before any work starts and one read at 90 days is the cheapest version that proves anything, and the cost of the second read is lower than the first because the instrument already exists.
How does this fit into a budget case for the CFO?
The audit is the input, not the case. It hands you the gap and the sample; the case comes from converting that gap into hours or output and costing the work to close it. That conversion is the subject of our walkthrough on building an AI training ROI model your finance team cannot dismiss, and the peer spending figures sit in what companies actually spend per employee.
Where to start
Send every bidder the same six-row table above and ask them to mark which rows are in scope, how many analyst days each takes, and what sample their observation row covers. Add one question underneath: is the fee credited against future work.
The shape of those six answers tells you more than the total at the bottom, and a vendor who will not fill it in has told you what the report would have read like.
Then capture your own baseline in the week before anyone starts: current task times, which tools people open, what they give up on. A few hours of work, and the only thing that turns an audit from a document into evidence. Buying licenses is not the same as building capability, and an audit that skips the baseline repeats that mistake at a higher price.
If you are holding a quote and want a second read on whether its scope fits what you need, a free AI-readiness call is a direct way to get one. We run these sessions, so we will also say when the cheapest proposal on your desk is the right one.
Dirk Jan van Veen, PhD
