Why do most AI agent ROI claims fail?
Because nobody wrote down the starting point. Without a baseline, any improvement is a guess, and any vendor average becomes your forecast by default.
What should each agent type be measured on?
One outcome per agent, chosen before it goes live, plus a guardrail that shows whether quality held.
| Agent type | Outcome metric | Guardrail |
|---|---|---|
| Customer service | Conversations resolved without a person, time to resolution | Customer satisfaction, reopen rate |
| Prospecting | Meetings booked, pipeline created | Reply quality, unsubscribe rate |
| Data and research | Research time per account, records enriched | Accuracy on a checked sample |
| Marketing content | Pipeline from campaigns, time to publish | Edits per draft, brand review fails |
| Revenue and billing | Days to collect, overdue invoices | Customer complaints |
| Custom agents | The one task it owns, in hours or volume | Error rate, escalations |
How do you set a baseline?
- Measure before launch. Four to eight weeks of the current numbers, by team, market and channel.
- Keep a comparison group. Where you can, run the agent in one market or team and not another.
- Count the full cost. Licences, HubSpot Credits, setup, and the people time spent reviewing agent output.
- Agree the formula. Finance signs off how time saved or revenue added turns into money, before results arrive.
What does the ROI calculation look like?
Value is the change in the outcome metric against baseline, converted to money with the formula finance agreed. Cost is everything in step three. Report it monthly.
Where does governance come in?
An agent with no owner cannot be measured. Before any agent goes live, set:
- An owner. One named person accountable for the agent's number.
- Limits. What it may do alone, what needs review, and a cap on credit spend. HubSpot lets admins cap credits per feature and account-wide (HubSpot).
- Review. A sample of outputs checked every week against the guardrail.
- A stop rule. The result that ends the pilot, agreed in advance.
HubSpot's Agent Hub shows live status and outcomes for each agent (HubSpot). Use it alongside your own revenue reports, not instead of them.
What about published averages?
HubSpot publishes averages for its agents, such as tickets closed and leads created (HubSpot). This is the approach in from AI pilots to P&L.
