Join AI Insiders
Back to Resources

Why does nobody use the AI software their company already paid for?

Why does nobody use the AI software their company already paid for?

Fewer than seven percent of Microsoft's 450 million Office seats carry a Copilot license, and the reason is not that the software is bad. This article answers the question underneath that number: why paid AI tools sit unused inside real businesses, and why that gap is the most durable service opportunity in AI right now.

The number that should have been a scandal

Microsoft just said the quiet part out loud. Fewer than 7% of its 450 million commercial Office seats carry a Microsoft 365 Copilot license. That disclosure, reported in the Neatprompts edition of September 30, 2026, landed in the same week Microsoft rebuilt Copilot around three tabs, Home, Code, and Autopilot, and started billing usage credits on top of the $30 per seat price.

Read that again. The most distributed software company on earth, with a direct line into every CFO in the Fortune 500, cannot get more than one in fourteen of its own customers to pick up its flagship AI product.

Here is the thing. That is not a product failure. That is a gap between capability and use, and the gap is where the work is.

So, the direct answer: nobody uses the AI software their company paid for because buying a license is a purchasing decision and using a tool is a behavior change, and almost nobody is paid to do the second one. The software arrives. The habit does not. The most valuable thing you can sell a business in 2026 is not another tool. It is the outcome the tool was supposed to produce, delivered.

Candidly, I spent two years on the wrong side of this. I bought tools. I collected logins. I had a browser full of tabs for platforms I had used exactly twice. I told myself I was investing in capability. I was buying permission to feel like I was moving. The tools were fine. I was the bottleneck, and no feature release was going to fix me.

Capability is not scarce anymore. Adoption is. And adoption is a service.

Key Takeaways

  • Fewer than 7% of Microsoft's 450 million commercial Office seats carry a Copilot license, which means the bottleneck is use, not capability.
  • Organizations use an average of 54% of the SaaS licenses they pay for, according to Zylo's 2026 SaaS Management Index.
  • MIT Media Lab's Project NANDA found 95% of enterprise generative AI pilots produced no measurable return, and blamed organizational factors rather than model quality.
  • McKinsey found executives badly underestimate how much their own employees already use AI.
  • The work that closes this gap is enablement, and enablement is sold as a service, not as software.

The problem is boring, which is why it stays unsolved

Everybody wants to build the tool. Almost nobody wants to sit with the bookkeeper at a 14 person plumbing company and watch her try to use it.

That is the problem in one sentence. The glamorous work is building. The valuable work is the hour after someone logs in for the first time and does not know what to type. One of those gets funding. The other gets ignored.

I know why it gets ignored, because I ignored it too. I would spend a weekend on an automation and then be privately annoyed that the people I built it for did not use it. I blamed them. That was wrong, and it took me longer than it should have to admit it. They did not need a better automation. They needed somebody to sit next to them while they used a worse one, and to keep sitting there until it stuck.

This is harder than it sounds, and I want to be honest rather than sell you a clean story. Changing how a person works is slow. It involves their ego, and the fear that the tool will expose that they did something inefficiently for eleven years, and the very real fact that most people are already at capacity.

There is also a status problem. Telling people you build AI tools sounds impressive. Telling people you help a dental office use the AI tools they already pay for sounds small. It is not small. It is the part everyone skipped.

Here is the reframe. Stop asking what you could build. Start asking what has already been bought and is not working. That question has a shorter path to a result you can point at, because the budget already cleared. Nobody has to be convinced to spend. Somebody has to be convinced the spend can finally pay off.

The evidence that the conventional story is incomplete

The conventional story is that AI adoption is slow because the technology is not ready. The data says something else.

Unused software is not an anomaly. It is the baseline. Zylo's 2026 SaaS Management Index, built on analysis of more than 40 million SaaS licenses and $75 billion in spend under management, found that the average organization uses just 54% of the licenses it pays for, and puts average annual waste from unused licenses at $19.8 million for large organizations. Microsoft's Copilot numbers are not an outlier. They are a documented pattern with a brand name attached.

The failures are organizational, not technical. MIT Media Lab's Project NANDA published "The GenAI Divide: State of AI in Business 2025," drawing on interviews with 150 executives, surveys of more than 350 employees, and analysis of 300 AI deployments. It found that 95% of enterprise generative AI pilots produced no measurable financial return, and attributed the failures to data readiness, workflow integration, and the absence of a defined outcome before building starts, not to model quality. The same report found deployments run with outside specialists succeeded roughly 67% of the time, against roughly a third of that for internal builds. Outside help worked better. Hold onto that.

Leaders do not know what is happening in their own buildings. McKinsey's January 2025 report "Superagency in the Workplace," which surveyed 3,613 employees and 238 C-level executives, found that C-suite leaders estimated only 4% of employees use generative AI for at least 30% of their daily work, while employees self-reported a figure roughly three times higher. In the same research, nearly half of employees said they receive minimal or no AI training.

The word "adoption" is doing a lot of work. The U.S. Chamber of Commerce surveyed 3,870 small businesses with fewer than 250 employees in June 2025 and found 58% using generative AI, up from 40% in 2024. The U.S. Census Bureau's Business Trends and Outlook Survey, asking whether AI is used to produce goods or services, put adoption at 17.8% in the same period. Both numbers are real. The space between them is the gap between a person trying ChatGPT and a business running on it. That space is the market.

Capability is not the constraint. Gemini 4 Argon took first place on Zapier's AutomationBench on September 30, 2026, at 51.3%, on a benchmark built around business process work. The models keep getting better at exactly the work small businesses need done. None of that solves the problem of a human being who has not opened the tab.

And then there is the counterexample. KOHLER, the 153 year old kitchen and bath manufacturer, reported reaching 98% Microsoft 365 Copilot adoption in a Microsoft customer story published August 10, 2026. Their goal was 95% in nine months. They hit it in 40 days. What did they do? Enterprise wide licensing, a champions network of more than 200 people, tiered training from beginner to advanced, and visible executive use. Their Chief Digital Officer, Eric Tabor, named what most companies miss: "The skilling that a lot of companies miss is the business skilling."

Same software everybody else has. Same $30 seat. 98% instead of 7%. The difference was people doing enablement work.

The shift: sell the outcome, not the access

The market is already telling you this. You just have to read what is selling.

On Product Hunt, Omnia Agent took 146 upvotes on October 1, 2026, and what it sells is not a tool for doing generative engine optimization. It sells 95% of your GEO work done. The day before, Pexo took the number one spot with 417 product page upvotes, with the maker framing video production as the bottleneck now that building is fast. Both sell a finished result. Neither sells a dashboard.

Alicia Lyttle posted on X on October 1, 2026, contrasting the old path, hire a developer and wait, with what she did instead that afternoon: run a client demo herself out of the ChatGPT model dropdown. The distance between an idea and a working demonstration collapsed. What did not collapse is the distance between a demonstration and a business that runs differently on Monday.

So here is the framework I use, and I use it on myself first.

Find the paid gap. Not a problem. A paid gap. Something a business already spends money on, every month, that is not producing a result. Microsoft 365 Copilot is the loudest example right now, but so is the CRM nobody logs into, the scheduling tool half the staff bypasses, the ChatGPT Teams account with four seats and one active user. Paid gaps are easy to confirm, because somebody can show you the invoice.

Sell the result, in their words. Not "AI training." Not "prompt engineering." Try this: "Your five person team pays for Copilot. In three weeks, every one of them will use it daily for the two tasks that eat their mornings, and I will show you the usage data." That is a sentence a business owner can say yes or no to without feeling stupid, which matters more than you think. Most owners will not admit they do not understand your offer. They will just not buy it.

Work with one person at a time, in their real work. This is where I got it wrong for two years. I taught concepts. Concepts do not transfer. The training transfer literature has said so since Baldwin and Ford's 1988 review in Personnel Psychology, which established that whether training sticks depends on the work environment and reinforcement, not just the quality of the instruction. Blume, Ford, Baldwin, and Huang's 2010 meta-analytic review in the Journal of Management confirmed that support in the actual work environment is among the strongest predictors of transfer. In plain language: a great workshop nobody reinforces on Tuesday is a cost, not an investment.

Build the champion before you leave. KOHLER had 200 champions. A nine person insurance agency needs one. Find the person on the team who likes this stuff, make them visibly good at it, and give them something to be proud of. They are why the habit survives after your invoice clears.

Measure what the owner cares about. Not prompts written. Not seats activated. Hours back, errors down, turnaround time cut. KOHLER reported 2 to 5 hours saved per associate per month and a 75% reduction in error defects on the engineering side. You will not have numbers like that on day one. You will have one process, one person, one before and after. Write it down. That is the only sales asset you actually need.

Candidly, this is the business I accidentally built. I did not set out to run AI coaching. I set out to build things. The coaching happened because people kept asking me the same question, and the question was never "what should I build." It was "I have the tool, now what."

What to do this week

1. Pick one business you already know. Not a cold lead. A client, a former client, a friend who runs something. The pitch for this first one is not money, it is access. One hour and honest answers about what they pay for.

2. Run a spend audit on their software. Ask for the monthly subscriptions and, for each one, how many people touched it last week. You will find the gap in under thirty minutes. Zylo's 54% benchmark means expect roughly half of it to be dead weight. You are not there to shame them. You are there to find the one line item worth reviving.

3. Pick exactly one workflow. The one that costs the most human hours per week and has the clearest before and after. Invoicing. Proposal drafting. Weekly reporting. Inbound email triage. Resist redesigning their whole operation, because a redesign never finishes and one workflow is a win you can bank in two weeks.

4. Write the current process down before you change anything. This is the step everyone skips and the one that makes you look like a professional. Time it. Count the steps. Note who touches it. It is your proof of value later, and it is what MIT's NANDA research found missing from failed deployments: a defined outcome before the build.

5. Use a prompt to design the intervention. Here is the one I use. Paste it into Claude or ChatGPT and fill in the brackets:

[The Job]
Help me design a two week enablement plan that gets one person at a small
business to actually use an AI tool they already pay for, daily, for one
specific workflow.

[The Background]
The business is [BUSINESS TYPE AND SIZE]. The person is [ROLE AND COMFORT
LEVEL WITH TECHNOLOGY]. The tool they already pay for is [TOOL NAME]. The
workflow I want to change is [WORKFLOW], which currently takes about
[CURRENT TIME] and involves these steps: [LIST THE STEPS]. What has already
failed or been tried: [WHAT HAS BEEN TRIED]. The owner cares most about
[OWNER'S PRIORITY].

[The Deliverable]
A day by day two week plan, written from the perspective of an experienced
change management practitioner who has rolled out software to non technical
staff. For each day give me the specific activity, who does it, how long it
takes, and what evidence tells me it worked. Flag the two most likely points
where this person quits, and what I do at each one. Plain language, no jargon,
no more than one page.

[The Questions]
Ask me any questions you have.

6. Deliver one visible win in the first week. Not at the end. In the first week. People stay with what is already working. One task that used to take forty minutes and now takes six buys you the patience you need for the harder parts.

7. Write down what happened. Two paragraphs. What it cost them before, what it costs now, who did the work. That is your case study, your proof, and the only marketing material that will not sound like everybody else's.

Frequently Asked Questions

Do I need to be technical to help a business adopt AI tools?

No. The bottleneck is behavior, not code. MIT Media Lab's Project NANDA found AI pilots fail on workflow integration and undefined outcomes rather than model quality. What you need is patience, the ability to map a process, and the willingness to sit with someone while they try. Those are facilitation skills, not engineering skills.

Why would a company pay someone outside to do this instead of handling it internally?

Because internal teams are already at capacity and nobody owns adoption. MIT's Project NANDA found deployments run with outside specialists succeeded roughly 67% of the time, against roughly a third of that rate for internal builds. An outsider also gets to ask the naive questions an employee cannot ask without looking uninformed.

Is this a short window that closes when the tools get easier to use?

The tools have been getting easier for three years and the gap has grown, not shrunk. Microsoft's own disclosure of fewer than 7% seat penetration on 450 million seats came alongside its most redesigned Copilot yet. Better products do not change habits by themselves. Humans change habits with other humans.

How is this different from selling AI training or workshops?

Training sells attendance. Enablement sells a changed process. The training transfer research, from Baldwin and Ford's 1988 review onward, consistently finds that reinforcement in the actual work environment predicts whether learning gets used. A workshop with no follow through is an expense. A workflow that still runs differently in six weeks is a result.

What kind of business is the best first client for this?

One that already pays for AI software and has between five and fifty employees. They are big enough to have bought tools and small enough that one changed workflow is visible to the owner. Zylo's 2026 benchmark of 54% average license utilization means almost any such business has an unused subscription worth reviving.

The gap is not a problem to complain about

Go back to that number. More than 418 million Office seats without a Copilot license. Every one belongs to a person sitting at a desk right now, doing something by hand that a tool on their own company's invoice could help with, and nobody has shown them.

That is not a statistic. That is a line of people waiting.

I am going to be blunt, because the people telling you the comfortable version are not helping you. You do not need another model release. You do not need a better tool. You do not need to wait until you feel qualified, and you never will feel qualified, because that feeling does not arrive, it gets replaced by the memory of having done the thing anyway. What you need is one business owner, one hour, and the nerve to ask what they pay for and never use.

Stop building in your bedroom for an audience you have not met. The hardest problem in AI is not technical and it never was. It is a person who will not open the tab. That person is not in the way of the future. That person is the work.

So this week, go sit with one of them. Watch them try. Do not fix it for them. Stay in the chair until it sticks. Then come tell me how it went, in the community, where I will be arguing with somebody about this same thing.

Everybody is racing to build what nobody is using. Go be the one who helps them use it.

About the author

Jonathan Mast is the founder of White Beard Strategies LLC, where he provides AI coaching and mentorship for entrepreneurs. He runs a Facebook community of more than 500,000 members and the AI Insiders membership, created the Perfect Prompt Framework, and speaks on practical AI adoption for small business. He is also a recovering tool collector who once paid for four AI subscriptions he had logged into a combined six times, which is how he learned that the gap between buying and using is where the real work lives. Come find him in the community and argue with him, he will take the other side on purpose.

Sources

Put This to Work With Us.

AI Insiders gives you practical training, direct coaching, and a community of entrepreneurs who are applying AI to real businesses.

Join AI Insiders