Skills Over Seats: Your AI Licenses Are 15% Used. Fix It Before Renewal.
| Advanced |
Seats aren't skills. A seat is a suit with nobody in it: buy one for everyone and by Month 6 you're at ~15% usage with a renewal 40% higher and nothing on the P&L to point at. One documented Copilot rollout: 4,000 seats, $1.4M, 12 people still using it after three months. The fix isn't more seats or a 45-minute webinar nobody watches. Put five operators in the suit, prove one workflow in 90 days, count the hours and the dollars, then expand from the win. Skills over seats.
I stop buying seats and start building skills ... and I don't mean skill.md files
Your AI rollout has a structural flaw, and the window to fix it is closing.
You bought seats for everyone. Six months in, 15% of them get used and the renewal just came in 40% higher. You've seen this movie: same script as every "transformation" before it, new tool. The difference now is the bill is bigger and the competitor who did the boring foundation work isn't waiting.
The Decision Point: Month 6
| Timeline | Current Trajectory | Skills-First Alternative |
|---|---|---|
| Month 6 |
|
|
| Month 12 |
|
|
That's the gap. Same budget, different plan.
Case Study: The $1.4M Adoption Gap
In December 2025 a tech exec laid out a scenario I've watched play out more than once: Microsoft Copilot to 4,000 employees at $30/seat/month (still list price in September 2026, but now there's bigger/better). $1.4M a year. Board approval in eleven minutes under "digital transformation."
Three months post-launch: 47 employees had opened the tool. 12 were still using it.
The response? A request for 5,000 more seats.
The stated rationale: "Adoption means mandatory training. Training means a 45-minute webinar no one watches. But completion will be tracked. Metrics go in dashboards. Dashboards go in board presentations." Great, your using spending as strategy, metrics as theater. Say one thing, do another. Behaviour speak louder than words. Ask me how I know.
Separately, J.P. Morgan's math says [AI investments need $650 billion in annual revenue](https://x.com/mweinbach/status/1987912908567916693) to return 10% on current buildout costs. That's roughly $35 per iPhone user or $180 per Netflix subscriber, forever.
The disconnect: big checks, nobody in the suit, nobody measuring.
Organizations winning this transition aren't spending more. They're spending differently.
I outlined this in [The Blank Slate](https://rileytech.net/blank-slate) series - the seat license buys you the [enthusiast machine](https://rileytech.net/posts/some-people-do) in every chair, all the stats, none of your context, telling the ticket-number guy "great idea" all the way down. Seats hand out [suits](https://rileytech.net/posts/the-suit). Skills put someone in them.The Core Issue: Scale Without Foundation
The default play: buy everyone a seat and assume adoption follows. "We're AI Enabled!". It doesn't. You get 15%, and 85% of very expensive suits hanging in a closet. The other play: five people who can explain, predict, and defend their own work prove it on one real workflow, and everyone else starts asking how.
Capital allocation: $50K to develop five experts with measurable outcomes vs. $5M in distributed licenses with uncertain adoption. One generates ROI data in under 90 days. The other becomes a sunk cost you're defending in next year's planning.
A Fortune 500 logistics operator ran this in Q1 2025. Five supply chain analysts, one problem: route optimization.
Month 3: 40% faster planning, $1.2M annualized fuel savings.
Month 6: second cohort training.
Month 9: the CFO asking why the other divisions weren't keeping up.
Market Performance (2025–2026 Data)
The 2025–2026 numbers say the same thing:
- MIT 2025: 95% of enterprise pilots generate zero measurable P&L impact - S&P Global 2025: 42% of firms terminated the majority of their AI initiatives in 2025 (vs. 17% in 2024) - Gartner 2025: Organizations prioritizing skills development show 2x+ likelihood of reaching mature implementation with sustained ROI - Gartner, April 2026: 1 in 5 AI projects in IT infrastructure and operations fails outright; 57% of I&O managers report at least one failure behind them - Microsoft, July 2026: 30M paid Copilot seats against ~450M commercial M365 seats. About 6.6% of the vendor's own base, two and a half years after launch
Same pattern every time: loud launch, nobody gets good at it, quiet shutdown. Then it gets blamed on "culture," same way Agile got blamed on "the framework," while your best people leave for the competitor that did the foundation work.
Six Strategic Adjustments (Plus One)
1. Target High-Impact Applications First
Pick a real bottleneck. Define success in numbers: hours saved, revenue generated, error rates down. Be honest about whether the team can execute inside 90 days. Your high performers need to understand both what these models are good at (fast iteration) and how they fail (confidently wrong). Skip that and you're funding frustration and flailing of your top people.
2. Develop Prompt Engineering as Organizational Capability
The gap between a lazy prompt and a structured one is the gap between search engine 2.0 and an operator in the suit.
Basic: "Summarize this data"
Structured: "Analyze Q3 enterprise software sales. Summarize: (1) deal velocity by region, (2) discount patterns above $100K, (3) win/loss themes. Format as executive brief—3 key insights with metrics, flag data gaps." For a deeper example, [go here](https://www.linkedin.com/posts/jeriley_so-heres-what-ive-been-saying-its-coming-activity-7369715991582289925-0awr?utm_source=share&rcm=ACoAAAGEPucBo42BkbKGL2I2dhyEoNnoUsvyOQ8).
Teams that treat this as a learnable skill (context + task + constraints + format + examples) get 5–10x, and more. Teams that don't call AI "not production-ready." It's not the model. Everything you didn't say, the machine invented.
### 3. Diversify Model Selection StrategicallyOne vendor is a leash. As of late 2025: Claude was the strongest at complex reasoning for articles and writing. Gemini was the pick for code and massive contexts. Grok was good at the latest and greatest ideas and concepts. Capabilities vary by use case and by month. Test aggressively across vendors on your workflows. What you get out of it depends on how you use it. New model comes out? Test it against what you've already built.
4. Create Knowledge-Sharing Infrastructure
When DevOps writes a prompt that saves three hours a week, Product should know within days. When development starts shipping 5x the stories, DevOps and the business should know within days. Set up the channels: a dedicated Slack space, demo sessions, searchable docs ...ORRR accept that every team rediscovers the same solution on its own. It's value on the shelf, but internal and it multiplies, fast.
5. Conduct Vendor Due Diligence
A lot of "AI-powered" products are a foundation model, a system prompt, and a markup - the good one's have vector stores bolted on. The costume is billable. The idea is free. Look at who owns the technology, how the pricing moves, and what it costs to leave. The question that matters: what's our exposure if this vendor quadruples pricing or gets acquired? What happens if they get straight up cut off? (it happens)
6. Plan for Continuous Capability Development
Model capabilities move weekly. Six months without a refresh and your team is working off a model that doesn't exist anymore. Budget quarterly refreshers, not just launch training.
Bonus: It's not just for software
Yes, the SDLC side of the house goes first. That doesn't mean the patterns, and the idiosyncratic knowledge your org already has, don't apply everywhere else. Same patterns, different words - aka fundamentals.
Implementation Framework
Identify Change Agents – Who's already maxing their AI quotas? That's your first cohort.
Launch Focused Pilot – Fund them, point them at one high-pain problem, surface blockers within weeks.
Enforce Rapid Win/Kill – Month one: the metric moved or the project is dead. No zombie initiatives.
Establish Metric Accountability – Demand 50%+ utilization. Documented improvements within 90 days. Quarterly reviews, public scorecards.
Assess Organizational Prerequisites – Orgs with clear goal-setting and real problem-solving adopt fast. If yours is political and risk-averse, AI won't fix that. It'll read it back to you, louder.
Do this and payback shows up in weeks. Miss the window and you're documenting failure for next year's planning.
Scale from demonstrated wins, not strategic assumptions. Then, and only then, go full send.
FAQ
Why do AI pilots fail?
Because they weren't pilots. They were purchases. Thousands of seats, no target workflow, no owner, no kill date, so nobody can tell you what "working" means. MIT put it at 95% of pilots with zero P&L impact in 2025. A real pilot is five people, one painful problem, and a Month 1 decision: it moved or it's dead.
Why is Copilot / AI seat adoption stuck around 15%?
Seats don't force anything. That documented Copilot rollout: 4,000 seats, three months, 47 logins, 12 people still using it. The gap isn't access, it's skill. Someone who can't write a structured prompt hires the machine as search engine 2.0, gets confident vagueness back, and decides the tool isn't ready. 15% is what you get when you buy licenses and skip the part where people get good.
How should we measure AI ROI before renewal?
Name the metric before the pilot, not after. Hours saved per person per week, revenue, error rate. Demand 50%+ utilization from the pilot cohort and documented wins inside 90 days. The logistics team hit 40% faster planning and $1.2M in annualized fuel savings by Month 3. That's a number you can hand a CFO at renewal. Can't name your number? Don't sign.
What does a 90-day AI proof look like?
Month 1: find the people already maxing their AI quotas, fund them, point them at one bottleneck, surface the blockers fast. End of Month 1: the metric moved or the project ends. Months 2–3: document the workflow, spread the wins across teams, hit 3+ hours saved per person per week. Roughly $50K for five experts. Not $5M for seats nobody opens.
Seats vs skills — which should we buy first?
Skills with a few seats. Five people with real outcomes give you ROI data in under 90 days. $5M in licenses gives you a sunk cost to defend next planning cycle. The first cohort trains the next 50 and the seat demand shows up on its own, with a business case attached. Buy seats when people are asking for them, not to make them ask. Seats hand out suits. Skills put someone in them.