Back to Guides
Enterprise Software 14 min read

The Capability Tax: what a company pays for every untrained seat

The Capability Tax: what a company pays for every untrained seat

The license is the cheapest thing about an untrained seat. A business assistant subscription runs roughly $20 to $30 per person per month, so $240 to $360 a year. The best available estimate of what untaught AI use costs the people around that person is about $2,200 a year — six to nine times the invoice, paid by colleagues, recorded nowhere.

That gap is the real cost of untrained staff using AI, and the reason AI enablement exists as a budget line at all. This article does the arithmetic, names every assumption so you can argue with it specifically, and ends with the three cases where the honest answer is to cut seats rather than buy training.

What the capability tax is

The capability tax is the recurring cost a company carries when it gives someone a tool and not a method. It is not a penalty for buying AI. It is what you pay for the distance between access and skill, and unlike the subscription, it does not appear on a statement.

Four components make it up. One is the license, which you can see. Three are time (rework, duplicated effort, and the fraction of a paid seat that goes unused), which you cannot.

The invoice scales with seat count. The time costs scale with how much work moves between people, which is why the tax is worst in the middle of a company rather than at its edges.

The framing matters. Most companies treat this as an adoption problem, which implies the fix is encouragement. It is a capability problem, and AI enablement names the work that closes it: teaching people to get reliable output from a general-purpose assistant on their own documents, and agreeing as a team what gets checked before it is sent.

Component one: the invoice, the only cost with an owner

Published business pricing sits in a narrow band. The enterprise add-on for Microsoft’s assistant is widely listed at $30 per user per month on an annual commitment, and ChatGPT’s business tier sits around $20 on annual billing after a 2026 price cut. Call it $240 to $360 per seat per year, and check your own contract: sub-300-seat tiers carry promotional rates that move.

For a 150-person company that bought seats for everyone, that is $36,000 to $54,000 a year. It is a real number, it has a named owner, and it renews on a date someone has in a calendar.

It is also the smallest of the four components, which is the whole problem: the component with an owner is the small one, so the owner optimizes the small one. Renewal conversations are about discounts and seat counts, never about whether the seats produce anything.

Component two: rework, charged to someone else’s calendar

The largest measurable component is work that arrives unfinished and gets finished by its recipient.

BetterUp Labs and Stanford’s Social Media Lab surveyed 1,150 US workers for Harvard Business Review in September 2025. They found that 41% had been handed AI-generated filler in the previous month: content that looks complete and lacks the substance to move the task forward. Each instance cost the recipient an average of 1 hour 56 minutes, which the researchers priced against respondents’ own salaries at roughly $186 per worker per month.

Annualize that and you get about $2,200 per worker per year, against a license of $240 to $360.

Three caveats. It is per worker, not per untrained seat: the cost lands on whoever receives the document, who may not hold a seat at all. It is built from self-reported time, so read it as an order of magnitude, not a line item.

And it is not evidence that AI produces bad work. It is evidence that untaught AI use pushes work downstream, invisibly, because the sender records a finished task.

Watch a team work and the mechanism is obvious. Someone asks an assistant to summarize six client calls, reads the output, finds it plausible, and forwards it. The recipient opens all six transcripts to find out which claims are real.

One person saved forty minutes; one person lost two. The dashboard shows a prompt sent and a document created.

This is the same measurement failure that makes license usage dashboards overstate real capability. An admin console counts whether the product was opened. It cannot count whether what came out of it was worth sending.

Component three: the seat paid in full and used at a fraction

The third component is simpler and harder to price: you are paying the full seat price for a narrow slice of what the seat does.

Section’s AI Proficiency Report tested 5,026 US knowledge workers on hands-on tasks rather than asking how they felt about AI. It found 5.5% at practitioner or expert level, 73.5% stuck as experimenters, and 20.9% as novices. A workforce of permanent experimenters is exactly what a license rollout with no teaching produces: everyone has the tool, almost nobody has a method, and the use case stays at search-and-summarize.

The Census Bureau’s Business Trends and Outlook Survey shows how common that is. In its November 2025 to February 2026 AI supplement, 64% of businesses using AI reported changing nothing about the business in order to use it, and only 15% had trained existing staff. Adoption on the broadened question runs around 18% of businesses. The modal American company using AI has bought access and changed no process, no role, and no review step.

If three-quarters of your seat holders use an assistant to search and summarize, you are paying a premium tier for a commodity capability. Teach the rest of the tool or downgrade the plan. Both are fine. Doing neither is the expensive option, and it is the default.

Component four: risk, and why you should not budget against the headline

Risk is the component most written about and the one most companies in the 50-to-500 band should discount heavily.

IBM’s Cost of a Data Breach Report for 2026, built on Ponemon Institute interviews at more than 600 organizations breached between March 2025 and February 2026, put the global average breach cost at $4.99 million, a record. Unapproved AI tool use was involved in 43% of security incidents, more than double the prior year. Roughly seven in ten of those organizations had no policy for managing AI or detecting unapproved use.

The 43% figure is the one to take seriously. It describes the behavior of people who were given no sanctioned method: they improvise with whatever tool is open in another tab. That is a training and policy gap expressing itself as a security event.

The $4.99 million is the one to leave alone. It averages across large, often regulated enterprises, and a 120-person services firm that budgets against it will either freeze or overspend. Your real exposure is narrower: which documents your team pastes into consumer accounts, what client contracts say about third-party processing, and whether the paste rule is written down. That costs a meeting to find out.

Why the AI enablement cost stays invisible

Three structural reasons, and none of them is carelessness.

The invoice has an owner and the time does not. Software spend has a budget line, an approver, and a renewal date. The two hours a manager spends repairing a summary have no cost center. It gets absorbed into a day that was going to be busy anyway.

The losses arrive below the escalation threshold. Nobody escalates two hours. Two hours is a bad afternoon, not an incident. The same loss arriving once as a $2,200 charge would trigger a review; arriving in two-hour pieces across eleven people’s calendars, it never becomes a number anyone reports. Most AI enablement programs fail for related reasons — the thing that would have justified the program was never counted in the first place.

Spending patterns reflect the visibility, not the cost. Writing in Fortune in March 2026, SHRM Foundation president Wendi Safstrom noted that AI spending is projected to rise 44% in 2026 while training budgets grow 5%. Nearly three-quarters of knowledge workers now use AI at work, and 60% report no formal training. That is not a judgment about priorities. It is what happens when one side of the ledger is itemized and the other is not.

The payoff for fixing it is measurable. Cognizant’s June 2026 survey of 1,100 senior leaders and 4,400 employees found 64% of trained workers reporting productivity gains of 20% or more, against 36% of untrained ones — a 28-point gap on the same tools. The tool is not the variable. The method is.

Three cases where cutting seats beats buying AI enablement

Almost every page on this subject concludes that you should buy training. Sometimes you should not, and knowing which is which is most of the value here.

The role has no recurring job for the tool. Company-wide rollouts hand seats to people whose week contains almost no document work: field technicians, warehouse staff, some clinical and production roles. No amount of teaching creates a use case that the job does not contain. Cut those seats and put the money and the teaching where people write and read for a living.

No one will protect the first eight weeks. Capability is built by people practicing on live work with a named owner checking in. If nobody will hold two recurring hours against delivery pressure, a workshop produces a good day and no change. Fix the ownership question first; it is free, and buying around it is not. The case for treating this as an operating discipline rather than an event is the same case.

The confidentiality question is unanswered. If your team handles client material under contract terms nobody has read against third-party processing, write the paste rule before you expand access. A one-page rule and a sanctioned account beat a training day you cannot safely act on.

In all three cases, cutting seats and reporting the saving beats buying a program that will not land. A vendor who cannot tell you which case you are in is selling a catalog, not corporate AI enablement.

The one number worth tracking

Divide the number of people who can show you one task they now do differently by the number of seats you pay for. That ratio is capability per dollar, and you can measure it this week without a platform.

Pick ten seat holders at random. Ask each to open their screen and walk you through one task the assistant changed, on real work. Count how many can.

If it is two, you are paying for ten seats and have two. If it is eight, your spend is working and the question becomes depth, not teaching.

Run it again eight weeks after any training, same ten people, same question. It is the cheapest honest measurement here and the hardest to game, which is why license counts and real capability come apart so reliably in companies that only track the console.

If you want a second opinion on your own numbers before a renewal, we run a free AI-readiness call: bring your seat count, your renewal date, and the two functions you think are furthest behind. Book it here. For the full sequence from license rollout to measurable capability, start with our AI enablement playbook. For the argument that access and capability were never the same purchase, read why buying licenses is not the same as building capability.

Frequently asked questions

What does an untrained AI seat actually cost per year?

Roughly $240 to $360 in subscription cost, plus an estimated $2,200 per worker per year in time spent fixing AI output that arrived unfinished, from the Harvard Business Review study of 1,150 US workers. The second figure is an estimate, not an invoice, and it lands on colleagues rather than the seat holder.

Is there a study that prices untrained AI use directly?

No. No credible study isolates the cost of an untrained user. What exists is research on adjacent costs (rework time, proficiency levels, breach involvement, trained-versus-untrained output), which you combine with your own seat count and loaded hourly rate. Anyone quoting a single per-employee figure is extrapolating, usually without saying so.

How do I calculate this for my own company?

Take your seat count times your annual per-seat price for the visible cost. For the hidden cost, multiply your headcount in document-heavy functions by $2,200 as a starting estimate, then replace that with your own figure: incidents per person per month, times two hours, times your loaded hourly rate. The Bureau of Labor Statistics put total employer compensation cost for private industry workers at $46.89 per hour worked in June 2026, which is a floor rather than a knowledge-worker rate.

Why does nobody notice this cost?

Because it arrives in two-hour pieces on calendars that have no cost center, and two hours is below the threshold at which anything gets escalated. The subscription, by contrast, has an owner, an approver, and a renewal date. Companies optimize the component they can see.

Is this an argument against buying AI licenses?

No. It is an argument against buying them alone. The seat is a precondition for capability and not a substitute for it, and the cheapest version of this mistake is buying fewer seats with teaching attached rather than more seats without.

What does AI enablement training cost in the market?

Most AI enablement services quote hands-on fluency workshops at roughly $2,000 to $15,000 depending on group size, session count, and how far the material is built around your own documents. Custom automation work is a separate purchase at roughly $25,000 to $150,000. Compare quotes on whether the session uses your live work or a demo file.

How long before training shows up in the numbers?

Expect four to eight weeks before behavior changes visibly, and only if someone owns the follow-through. A single session with no scheduled practice reliably produces a good day and no measurable change, which is the most common way training budgets get wasted.

Should we train everyone or start with one team?

Start with one team that writes and reads for a living, run by a manager who will protect the practice time. Small groups let people watch a colleague with the same job solve a problem they both recognize, which is where most usable learning comes from. Company-wide rollouts lose that.

Does unapproved AI use create a real security cost?

For some companies, yes. IBM’s 2026 breach report found unapproved AI tool use involved in 43% of security incidents at the breached organizations it studied, and roughly seven in ten of those organizations had no policy for managing AI or detecting that use. Treat the mechanism as real and the dollar averages as inapplicable to a mid-sized firm; price your own exposure from your contracts instead.

Key takeaways

  • An untrained AI seat costs roughly $240 to $360 a year on the invoice and an estimated $2,200 per worker per year in rework pushed onto colleagues — six to nine times the visible cost.
  • The rework estimate comes from BetterUp Labs and Stanford’s Social Media Lab for Harvard Business Review: 41% of 1,150 US workers got AI-generated filler in a month, at 1 hour 56 minutes each.
  • Only 15% of US businesses using AI have trained existing staff, and 64% changed nothing about the business to accommodate it, per the Census Bureau’s 2025–2026 AI supplement.
  • Cognizant found a 28-point gap in reported productivity gains between trained and untrained workers on identical tools, which is the clearest evidence that method rather than tool is the variable.
  • Discount the headline breach averages and price your own exposure, but take the 43%-of-incidents finding seriously as a training and policy signal.
  • Cut seats rather than buy training when the role has no recurring use, when nobody will protect the practice time, or when the confidentiality question is still open.
  • Track one ratio: people who can demonstrate a changed task, divided by seats paid for.

Last Updated: Nov 18, 2026

AW

Arthur Wandzel

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles