AI adoption graph has to go up and right: a corporate satire with real lessons
A viral Reddit post skewers a familiar pattern in enterprise AI: spend big, measure little, present glossy graphs. It follows an executive who rolled out Microsoft Copilot to 4,000 staff at $30 per seat per month, declared “digital transformation”, and then quietly discovered almost nobody used it. The punchline: expand the licences anyway, and make the dashboards look good.
You can read the original here: AI adoption graph has to go up and right. Treat it as satire – but it lands because it mirrors how AI is sometimes bought, sold and reported.
“I told everyone it would 10x productivity.”
“The graph went up and to the right.”
Copilot rollouts, costs and the ROI question
In the post, 4,000 Copilot licences at $30 per month add up to roughly $1.4 million a year. Three months later, 47 people had opened it, 12 used it more than once. That’s not an adoption curve – it’s a warning light.
Microsoft Copilot for Microsoft 365 is a serious investment. If you’re making that call in the UK, the right questions are boring but vital: which roles will actually use it, how will we measure time saved or quality gains, and what will we stop doing as a result? Vanity metrics like “AI enablement” are not outcomes.
For official capabilities and pricing, see Microsoft Copilot for Microsoft 365 (pricing varies by region and plan).
Why this matters for UK organisations
Beyond the banter, there are UK-specific realities:
- Data protection – You’ll need a Data Protection Impact Assessment (DPIA) under UK GDPR, clarity on data handling, and a defensible retention policy.
- Public sector and regulated industries – Procurement, value-for-money, and audit trails matter. If you can’t show utilisation and impact, expect questions.
- Change management – UK workforces are diverse in digital maturity. Role-based onboarding, union/works council engagement, and accessibility are not optional extras.
- Cost control – With per-seat pricing, low adoption is expensive. Track cost per successful use-case, not licence counts.
From vanity metrics to meaningful outcomes
Replace “AI enablement” with metrics that reflect real work. Hallucinations (confident-sounding wrong answers) are a known risk with large language models – measure the cost of corrections, not just usage.
| Metric | What it actually measures | How to collect | Risks/Notes |
|---|---|---|---|
| Weekly active users (WAU) | Real adoption by unique users | Admin/tenant analytics | Low WAU suggests poor fit or onboarding |
| Task time saved | Minutes saved on specific tasks | Time-and-motion sampling before/after | Self-reported time can be inflated |
| Quality outcomes | Readability, accuracy, compliance | Peer review, QA scores | Define “good” in advance |
| Rework due to hallucinations | Time spent correcting AI output | Survey + annotated examples | Tracks the hidden cost of errors |
| Adoption by role | Who actually benefits | Role mapping + usage logs | Use for seat reallocation |
| Cost per successful task | Spend divided by verified wins | Costs + validated outcomes | Great for CFO conversations |
| Support/ticket rate | Friction and failure points | Helpdesk categories | Inform training and guardrails |
A practical adoption playbook for Copilot in the enterprise
- Start with three high-value, repetitive use-cases per function. Examples: summarising long email threads for client services; first-draft bid responses for sales; meeting note summaries for project teams.
- Baseline before rollout. Measure current time/quality so you can compare.
- Run a 6–8 week pilot with power users. Cap seats, instrument everything, and publish transparent findings – including what didn’t work.
- Secure-by-default. Apply least privilege, clear data boundaries, and turn off risky connectors until they’ve passed review.
- Enable in the flow of work. Short role-based training, in-product tips, prompt libraries, and office hours – not a one-off webinar.
- Incentivise outcomes. Recognise teams that retire old processes or hit defined time-saved targets.
- Reallocate or retire licences. If a role doesn’t use it after two months of enablement, pull the seat and try a different team.
Security and compliance – what “enterprise-grade” actually means
The post jokes about “compliance: all of them”. In reality you should be explicit. Typical UK controls include:
- DPIA and a lawful basis for processing under UK GDPR and the Data Protection Act 2018.
- Data residency and transfer assurances that meet your regulatory needs.
- Retention, DLP, and eDiscovery aligned to your information governance.
- Admin controls, audit logs, and access reviews tied to least privilege.
- Clear user guidance on sensitive data, confidential material, and export rules.
See Microsoft’s official guidance: Data, privacy, and security for Copilot for Microsoft 365.
Executive dashboards that don’t lie
If the board wants a graph, give them one – but make it honest:
- Show an adoption funnel: licences – enabled users – weekly active – tasks completed – tasks validated.
- Report cost per validated outcome and trend it down over time.
- Include “sunset decisions”: seats or use-cases you turned off and the money saved.
- Publish a quarterly learning report with examples, prompts, and before/after comparisons.
Tools and quick wins
Don’t over-engineer analytics. Start with what you have: tenant usage reports, a simple data warehouse or spreadsheet, and a standard taxonomy for tasks. If you’re building lightweight dashboards, this primer may help: How to connect ChatGPT and Google Sheets (Custom GPT).
Critically, maintain a living prompt and template library. What works for your organisation’s tone, data, and risk profile is highly context-specific.
Final thought
The Reddit story is funny because it’s plausible. But AI in the enterprise doesn’t have to be theatre. Pick a few jobs worth doing, measure them properly, and be willing to stop what doesn’t work. If your graph goes up and to the right after that, it will mean something.
