The AI Adoption Gap: Why 85% of Your People Can Use AI and Only 25% Do

Sometime last year, a mid-sized company bought a few hundred AI licences, ran a launch webinar, and put a chatbot in front of every knowledge worker. For two weeks the usage dashboard looked wonderful. Then it sagged. By month three, most of those seats were dormant - assigned, paid for, and never opened. The company had not made a technology mistake. The model was excellent. The people were willing. And still, almost nothing changed.

This is the defining pattern of enterprise AI in 2026. IBM’s CEO study found that 85 percent of employees now have access to AI tools at work while only 25 percent use them regularly - a 60-point gap between capability and behaviour¹. The same study found that 83 percent of CEOs believe AI success depends more on people’s adoption than on the technology, and 86 percent believe their workforce already has the skills^1,2. The tools are ready. The people are ready. The value is missing anyway.

The usual explanation is that people need more training. That diagnosis is comforting and mostly wrong. The reason a generic chatbot gets tried and quietly abandoned is not that employees cannot prompt - it is that the tool does not know how their company works, so it gives generic answers they have to check and fix. This article explains why the gap opens, what unused seats really cost in euros, why training and rollout alone will not close it, and how to close it by inverting the problem: put AI where the work is, ground it in how your company actually operates, and measure the outcome instead of the logins.

The 60-Point Gap Hiding in Your AI Budget

Access to AI is now close to universal in large organisations, and use of it is not. That single mismatch is the most important number in enterprise AI, because a licence produces nothing until the work behind it changes. The gap is where budgets go to die quietly.

What the numbers actually say

• 85 percent have access - In IBM’s 2026 CEO study, roughly 85 percent of employees have AI tools available at work. Access is no longer the bottleneck¹.
• 25 percent use it regularly - Only about a quarter use those tools as a regular part of their job, leaving a 60-point gap between what is possible and what happens¹.
• 83 percent of CEOs blame adoption, not tech - A large majority of chief executives say AI success now depends more on people adopting it than on the models themselves¹.
• 86 percent think their people are ready - CEOs are confident the workforce has the skills, a confidence the usage data flatly contradicts^1,2.
• 64 percent of Copilot seats sit idle weekly - Where the gap is most measurable, in Microsoft 365 Copilot rollouts, weekly active use has run at roughly 20 to 30 percent of licensed seats^4,5.

Access curves and usage curves have separated

For most of the past two decades, rolling software out to people was the hard part; once they had it, they used it. AI has broken that link. Provisioning a seat takes an afternoon, but changing how someone does their job takes far more than access.

• Provisioning is instant - An admin can assign thousands of seats in a day, which is exactly why access reached 85 percent so fast.
• Habit change is not - Getting a person to route a real task through a new tool competes with a workflow they already know and trust.
• Generic value is easy to walk away from - When the tool helps a little but not reliably, dropping it costs nothing, so people do.
• The dashboard flatters you first - Launch-week usage spikes as people try it once, then decays - which is why month-three numbers are the honest ones.
• The gap compounds - Every quarter of low use makes the next rollout harder, because the workforce has already learned that AI tools do not stick.

A gap this consistent across studies and vendors is not a run of bad luck. It is structural, and structure is what the rest of this article is about.

Why Generic AI Tools Stall After the First Week

The quiet abandonment of AI tools has a specific mechanism, and it is not laziness. A general-purpose assistant knows the public internet and nothing about your company, so the first genuinely work-specific question it gets produces an answer that is plausible, generic, and slightly wrong. The employee corrects it, notices that took longer than doing it themselves, and files the tool under not-for-me.

The abandonment loop, step by step

1. Curiosity - A new tool arrives with a launch email. People try it on a real task from their actual job.
2. Generic output - The tool does not know the price list, the approval rules, the customer history, so the answer is right in general and wrong in the specifics that matter.
3. The correction tax - The employee has to verify and fix the output, which feels slower than the manual way they already trust.
4. Silent drop-off - No one files a complaint. They just stop opening it, and the seat goes quiet.
5. Shelfware - The licence renews on schedule, invisible to everyone except the finance line that pays it.

Why a horizontal assistant cannot own the work

Tools like ChatGPT and Copilot are horizontal by design: helpful across every task, responsible for none of them. That is a strength for open-ended help and a weakness for closing the adoption gap.

• It waits to be invoked - A chatbot only helps when a person remembers to open it and phrase a request. Usage depends entirely on human habit.
• It has no context - Without your systems and rules, it cannot answer the questions that actually recur in your business.
• It owns no outcome - It drafts and suggests, but it never finishes a job end to end, so no process actually gets faster on its own.
• It does not learn your corrections - The fixes an employee makes today are gone tomorrow, so the tool never gets more useful to that team.
• It cannot be measured on results - Because it owns nothing, the only thing to measure is usage - which is exactly the metric that keeps disappointing.

That is the whole game in one sentence. If AI only exists as a separate window someone has to remember to visit, adoption will always depend on willpower. If it operates inside the work, adoption stops being optional - which is the shift the second half of this article is built on.

The Real Cost of a Seat Nobody Opens

Unused AI licences look cheap because each seat is a small monthly line item. In aggregate they are not, and the subscription is only the part you can see on an invoice. The larger cost is the productivity gain the non-adopters never captured - a number that does not appear on any bill.

The visible cost: subscriptions for switched-off seats

Microsoft 365 Copilot lists at roughly 30 euros per user per month, or about 360 euros per seat per year. Apply the observed 64 percent idle rate and the waste is straightforward arithmetic.

Those figures are subscriptions alone, and they are not hypothetical - a Fortune 500 company with 50,000 Copilot seats has been estimated to waste over 13 million dollars a year on seats nobody uses⁶. The wider software picture is just as grim: enterprises waste an estimated 30 percent of software spend on idle licences, and AI premiums make each wasted seat pricier than before^7,8.

The invisible cost: the productivity that never happened

The subscription waste is the small number. The big one is opportunity cost, and it dwarfs the licence line.

• Documented uplift is large - McKinsey research puts the productivity gain from effective generative AI use on knowledge work at roughly 20 to 40 percent on suitable tasks¹².
• Non-adopters capture none of it - Every employee who quietly stopped is contributing zero of that potential gain, quarter after quarter.
• The base is your payroll, not your software bill - The value at stake is a slice of fully loaded salary cost, which is one to two orders of magnitude larger than the seat price.
• It compounds with volume - As work grows, the gap between what the team handles and what it could handle widens rather than holds.
• It is invisible by design - No invoice ever shows the report that took three hours instead of one, so the largest cost is the one nobody tracks.

Why Training and Rollout Alone Do Not Close the Gap

When usage disappoints, the reflex is to schedule more training, appoint champions, and run another enablement campaign. These help at the margins, but they treat a fit problem as a skills problem, so they buy a temporary bump and then the same decline. The evidence is in the CEO confidence itself.

Why the training reflex underdelivers

• The skills were rarely the blocker - IBM found 86 percent of CEOs believe their people already have the skills to work with AI, yet 75 percent still do not use it regularly. If skills were the constraint, those numbers would not coexist¹.
• Training cannot add context to a tool - No amount of prompt coaching makes a generic assistant know your approval matrix or your customer history. The gap between generic and specific is in the tool, not the user.
• Motivation decays without a payoff - A champion can push people to try again, but if the output still needs heavy correction, the second attempt fails the same way as the first.
• Rollout optimises the wrong metric - Enablement programmes chase logins and active users, which measure effort, not value. The moment the campaign ends, the metric falls.
• It puts the burden on the busiest people - Asking every employee to change habits taxes exactly the people with the least slack, which is why habit change loses to the status quo.

What the scaling data confirms

The pattern shows up one level higher too. Companies are excellent at starting AI and poor at scaling it, and the reason is the same: pilots prove capability, but only redesigned work produces value.

• Piloting is common, scaling is rare - BCG found roughly 64 percent of companies pursue AI pilots but only about 26 percent embed AI in a broader transformation¹⁰.
• Only a quarter see real value - Only around one in four executives report significant value from their AI initiatives, and they got it by changing core processes, not by handing out tools^9,10.
• The value gap is widening - BCG’s own framing is a widening gap between the few that scale and the many that stall⁹.
• Process change is the common factor - The winners scaled a small set of initiatives, redesigned the work around them, and measured returns - the opposite of a broad, shallow tool rollout¹⁰.

Inverting the Problem: Measure Outcomes, Not Logins

Every approach so far tries to raise a usage number: get more people to open the tool more often. Inverting the problem drops that goal entirely. Instead of pushing AI at people and hoping they adopt it, you put AI inside the work and measure what comes out the other end. Adoption stops being a behaviour you police and becomes a property of the process.

The two ways to think about closing the gap

What the inversion changes in practice

• The work comes to the AI, not the reverse - The AI sits on the inbox, the queue, the ticket, or the ERP transaction, so it acts when work arrives instead of waiting to be summoned.
• Value is legible - Because an AI employee owns a process, you can point at throughput and turnaround and say exactly what changed.
• One change, not thousands - You redesign a single workflow rather than trying to shift the daily habits of an entire workforce.
• Corrections make it better - Feedback on the process feeds back into the system, so accuracy compounds instead of resetting every day.
• Headcount is freed, not cut - The routine load moves to the AI employee, so people absorb more volume and spend time on judgement work rather than being replaced.

The Company Brain: AI That Knows How You Work

Inverting the problem only works if the AI in the workflow is actually good at your work, and that requires the one thing generic tools lack: context. A Company Brain is a living memory of how your business operates - your processes, rules, product and customer knowledge - grounded in your real systems and kept current as people come and go. It is the difference between an assistant that knows the internet and one that knows your company.

What a Company Brain holds that a chatbot does not

• Your processes - The actual sequence of steps a task takes across your systems, including the workarounds that keep it running.
• Your rules - Approval thresholds, exception handling, pricing logic, the account that always gets special treatment - the specifics that make output usable.
• Your product and customer context - Part numbers, configurations, contract terms, service history, so answers are grounded in what is true for you.
• Your tacit know-how - The judgement usually locked in a few experienced heads, captured so it does not walk out the door at retirement.
• Your live state - Connected to email, Teams, SharePoint, CRM and ERP, so it reflects the current state of the business, not a training snapshot.

How the Company Brain closes the gap at the root

A chatbot is a tool you have to remember to use. A Company Brain is an asset that makes every AI action in your business more accurate over time - and it is what an AI employee stands on to do real work.

AI Employees: Adoption by Default

If a Company Brain is the memory, an AI employee is what acts on it. Unlike a chatbot that waits to be invoked, an AI employee owns a process end to end - it sits on the work, does the routine parts, and escalates the judgement calls to a human. That is what makes adoption a default rather than a campaign: the value arrives whether or not anyone opens an app.

How an AI employee differs from a chatbot

• It owns an outcome - Not “help me draft this” but “handle this queue,” measured on throughput and turnaround rather than usage.
• It works where the work is - On the inbox, the ticketing system, the ERP, so it acts when work arrives instead of waiting for a prompt.
• It runs on your Company Brain - Every action is grounded in your processes and rules, so output is specific enough to ship without heavy correction.
• It knows its limits - Routine cases go through automatically; anything unusual is handed to a person with the context attached.
• It improves with feedback - Human corrections feed back into the Brain, so the same process gets more accurate the longer it runs.
• It scales without hiring - When volume doubles, the AI employee absorbs it, so the team handles more without a proportional headcount increase.

Where AI employees take over routine load

• Finance - Invoice intake, coding, reconciliation and approval routing, so month-end shrinks without adding accountants.
• Customer service - First-line replies grounded in your policies and order history, with edge cases escalated to a human.
• Sales operations - Quote preparation, CRM hygiene and follow-up chasing, freeing reps for actual selling.
• Back office and admin - Data entry across systems, status updates and document handling that no one enjoys and everyone does.
• Knowledge work - First drafts of recurring reports and responses, built from your templates and prior work rather than a blank page.

How Superkind Closes the Adoption Gap

Superkind does not sell more seats. We build a Company Brain that learns how your company works and the AI employees that run on it, so value shows up in the outcome instead of depending on whether each person remembers a chatbot. Most AI tools know the internet but not your company - we close exactly that gap, and we get one process live in a small number of weeks rather than months.

Core capabilities

• A Company Brain that learns your business - A living memory of your processes, rules and context, grounded in your real systems and kept current as people come and go.
• AI employees that own a process - Agents that take over routine work end to end - data entry, email, reconciliation, approvals - rather than waiting to be prompted.
• Deep integration, no rip-and-replace - Connects to email, Teams, SharePoint, CRM and ERP, so AI works inside the tools your team already uses.
• Outcome-based measurement - We track volume, cycle time and error rate on the process we automate, so you see value in the work, not a usage dashboard.
• Learning from daily feedback - Every human correction feeds back in, so the system gets more accurate the longer it runs.
• Human-in-the-loop by design - Routine cases run automatically; unusual ones are escalated to a person with full context attached.
• Fast time to first value - One high-pain process live in weeks, because you change one workflow rather than a whole workforce’s habits.
• DSGVO-ready deployment - EU-hosted or on-premise options, source-cited answers and audit logging, so personal data stays in your perimeter and answers stay traceable.

Superkind vs a generic AI rollout

A Practical Playbook to Close the Gap

Closing the adoption gap is a sequence, not a campaign. The order matters: pick a process where value is measurable, ground the AI in your context, automate end to end, and prove the outcome before you scale. This is the playbook we run, and it works because it never asks a whole workforce to change habits.

Step by step

1. Find your dormant seats - Pull the usage data on the AI tools you already pay for. The size of the idle gap is your starting business case.
2. Pick one high-pain, high-volume process - Choose a workflow that is repetitive, measurable, and eats hours every week - invoice handling, first-line service, quote prep. Not the whole company, one process.
3. Baseline the outcome - Measure today’s volume, cycle time, error rate and cost per transaction, so you can prove change later.
4. Capture the context - Build the Company Brain for that process: the rules, the exceptions, the systems it touches, the tacit judgement of the people who run it.
5. Automate end to end - Put an AI employee on the process, running the routine cases and escalating the unusual ones with context attached.
6. Keep a human in the loop - Start with review on the outputs, then widen autonomy as accuracy proves out on real work.
7. Measure against the baseline - Compare the outcome numbers, not usage stats. If throughput and turnaround moved, the gap closed.
8. Extend to the next process - Reuse the same Company Brain for the adjacent workflow. Each rollout is faster because the context compounds.

Seat Model vs Outcome Model: A Decision Framework

Both models can coexist - a general assistant for open-ended help alongside AI employees that own processes. The mistake is expecting the seat model to close the gap on its own. Use this framework to decide where each belongs.

When each model is the right call

Quick decision guide

Related Articles

• The Practical Guide to AI Adoption for the Mittelstand
• ChatGPT at Work: Where a Generic Chatbot Helps and Where It Stops
• AI Agents vs Copilot: What German SMEs Get Wrong
• Copilot vs a Company Brain: The Difference That Decides Adoption
• The AI Productivity Paradox: Why the Gains Do Not Show Up
• What Does a Company Brain Actually Cost?

Frequently Asked Questions

Henri Jung

Co-founder at Superkind

Co-founder of Superkind, where he helps SMEs and enterprises deploy custom AI agents that actually fit how their teams work. Henri is passionate about closing the gap between what AI can do and the value it creates in real companies. He believes the Mittelstand has everything it needs to lead in AI - it just needs the right approach.