Why almost nobody gets ROI from ChatGPT (despite paying $20/month)
73% of US companies say they use generative AI in some form. Only 12% can show measurable savings. The difference isn't the tool — they all pay the same $20/month for ChatGPT Plus or $25 for Team. The difference is three things: how it gets rolled out, how the team is trained, and how it's measured. What follows is what the 12% actually do — and what the other 61% skip.
The three conditions for ChatGPT to work in a company
- Clear, role-specific use cases. Not "use ChatGPT for whatever you want" — but "for this specific task, this is the exact flow".
- A written, communicated usage policy. What data stays out, what decisions stay human, what happens if someone breaks the policy.
- Real impact measurement. Hours saved per month per process — not "% of the team with an active account".
Use cases by department
Sales
Pre-meeting research (LinkedIn + prospect site + brief), first drafts of proposals, summaries of long calls, generating discovery questions. Typical ROI: 8-12 hours/week per senior SDR.
Customer support
Copilot for the human agent: response drafts, summaries of the customer's history, tone suggestions for the situation. Typical ROI: 50-70% less drafting time.
Operations
Document processing, report generation, internal translations, meeting transcription with action-item extraction. Typical ROI: 20-40 hours/month per person in ops.
Finance
Budget variance analysis, management-report commentary, first pass at reconciling discrepancies. Important caveat: don't put sensitive financial data into ChatGPT Free/Plus.
HR
Job description drafts, first-pass resume screening (carefully — bias risk and emerging AI regulation), internal comms, employee FAQs. Heavily regulated territory — human oversight is non-negotiable.
ChatGPT vs. Claude vs. Gemini vs. Copilot — which one and when
| Model | Strength | When to pick it |
|---|---|---|
| ChatGPT (GPT-4) | Versatility, most mature ecosystem | Reasonable default for most teams |
| Claude (Anthropic) | Reasoning, long-form writing, code | Technical work, substantive writing |
| Gemini (Google) | Workspace integration, multimodal | Google-first companies |
| Copilot (Microsoft) | Native M365 integration | Microsoft-first companies without custom builds |
Governance: usage policy, sensitive data, audit trail
A usage policy is mandatory from day one — not optional. Minimum: one clear page with what's allowed, what isn't, and what happens if it's broken. Essential components:
- What data CANNOT go into external chats (closed list).
- Which plan the company pays for and how to get an account.
- What decisions CANNOT be delegated to AI (sensitive HR calls, binding legal/financial decisions).
- How to report misuse or an incident.
- Consequences of misuse.
- Who owns the policy inside the company.
How to train the team (without turning it into "look at the keys")
General training — "this is ChatGPT, this is a prompt" — lasts two hours and gets forgotten. The training that actually works is role-specific, with concrete cases and tested prompts:
- Identify 8-12 real use cases for the role with the process owner.
- Design tested prompts for each case — not generic ones.
- Hands-on workshop of 3-4 hours (no more) per team.
- A 1-2 page quick reference doc with the prompts and when to use them.
- Two-week check-in: review adoption and adjust the cases.
Metrics that tell you if it's actually working
| Level | Metric | Healthy read |
|---|---|---|
| 1 · Adoption | % of team with weekly active use | >60% by day 90 |
| 2 · Frequency | Sessions/person/week | >5 |
| 3 · Impact | Hours saved/month per process owner | >10 h/person involved |
Typical mistakes (and what they actually cost)
- Free plan for company use. Your data can train the model. Real cost: one reputational or data-leak incident.
- No written policy. You can't apply fair consequences when an incident hits.
- Generic training. Wasted investment — the team remembers nothing in two weeks.
- Not measuring impact. Impossible to justify continued investment — the project dies at budget review.
- Department-by-department rollout with no coordination. Duplicated effort, inconsistent practice, regulatory exposure.
Free material · PDF
ChatGPT use-policy template (ready for your company, GDPR-ready)
The document your DPO will ask for before you deploy ChatGPT for real. Editable, covered by GDPR and AI Act, with examples in English.
What you get
- Editable use-policy template (1 page + 5 annexes)
- List of forbidden data by sensitivity
- Incident protocol with deadlines