OPINIONS / #0005 · OPINION · THE FRAMEWORK · BUSINESS CASE · ADOPTION
The most expensive line in an AI business case is the one nobody wrote.
- AI business cases rarely carry an adoption number. The blank gets counted as every licensed user.
- Every party is paid before adoption is measured, except the buyer, who pays again at renewal.
- Surveys put weekly use of licensed AI assistant seats at 20–30%.
- Fund all three parts: measure use, name an owner per use case, and read the numbers monthly.
I've sat on the selling side of more than 50 AI business cases in the last five years. Almost none of them had an adoption number. They had a seat count, a value per seat, and a multiplication. Nobody lies about it. The adoption field gets left blank, and the spreadsheet treats a blank as everyone.
That blank has a name here: The 100% Assumption, the business-case line that assumes every licensed user will use the tool, every week.
Who is paid, and when
The deal closes at signature. The project closes at go-live. The renewal lands on the buyer, alone, months after anyone could have changed course. I call this the Shelfware Economy: a market where every party is paid before adoption is measured, except the one who pays for it.
What it costs
Surveys put weekly use of licensed AI assistant seats at 20–30%1. The tool works and the licences are live. Nobody tracked whether the people it was bought for ever opened it.
I'm not writing from outside this
I've built approval gates in my own CRM that made the data worse while making the dashboard look better. Rigour that demos beautifully can still poison the record. Some of these essays are about me.
Buy all three parts
Swapping tools won't fix it. Fund the other two-thirds too: the people whose week has to change, and the tracking that shows whether it did. Machine in the Loop is my name for buying all three parts, not only the one with an invoice.
Licences and models. They come with a price, so they get funded.
The people whose week has to change. No line item, so no budget.
Tracking that shows whether anyone uses it. Rarely costed, so rarely built.
Three things to put in the case before you sign
- 01Measure. Replace 100% with a measured number: weekly use, in the workflow the case depends on, at month six.
- 02Own. Name one owner per use case: the person whose week changes, not the sponsor.
- 03Track. Track use from day one and read it every month, not once at renewal.
- Collateral: one buyer archetype a fortnight, each acquitted, each costed.
- The ROI model: your business case, rerun with a measured adoption number.
- Pattern records: decisions and the usage data behind them, never a client's architecture.
- The field log: responses from people who measured, not people who guessed.
If your next renewal is within a year, start with the number.
Run the adoption-first ROI modelQuestions this answers
- What is The 100% Assumption?
- The business-case line that assumes every licensed user will use the tool, every week. It usually isn't written down: the adoption field is left blank, and the spreadsheet counts a blank as everyone.
- What is AI shelfware?
- Licensed AI tools that are paid for but not used in the workflow the business case depends on. Enterprise surveys reported by Fortune (29 May 2026) put weekly use of licensed AI assistant seats at 20–30%.
- What is Machine in the Loop?
- A name for buying all three parts of an AI system: the machine (licences and models), the humans whose week has to change, and the loop (tracking that shows whether anyone uses it). Humans drive. Machines accelerate.
- How do you measure AI adoption?
- Count weekly use in the workflow the business case depends on, at month six, not logins. Name one owner per use case, track use from day one and read it every month, not once at renewal.
ATTACHMENTS
CALCULATOR#0006 Adoption-first ROI modelOPENHow I get paid: I'm CRO at an AWS partner. We sell cloud and AI delivery. Nothing here is sponsored, and I flag it when a post touches what we sell.
- Enterprise surveys reported by Fortune, 29 May 2026, via Digital Trends and Windows Latest. The coverage doesn't name the survey or its sample.