AI automation ROI calculator — methodology and benchmarks
TL;DR Honest ROI on AI automation comes from a simple formula: (annual labour saved + annual revenue gained − annual run cost) ÷ project setup cost. Most well-scoped SMB projects return 200–500% in year one. The trick is using realistic inputs, not vendor-supplied benchmarks. This article gives the formula, the inputs you need, and validated benchmarks from real deployments.
Key takeaways
- Annual ROI formula: (Labour saved + Revenue gained − Run cost) ÷ Project setup cost
- Realistic year-one ROI for well-scoped SMB projects: 200–500%
- Use real labour rates including overhead (1.4× salary), not headline salary
- Include all hidden run costs (model, telephony, vector DB, monitoring) — typically 8–15% of project cost annually
- Discount vendor benchmarks by 30–50% for your own modelling
The formula
Annual ROI % = ((annual labour saved + annual net new revenue − annual run cost) ÷ project setup cost) × 100
Each input deserves a realistic estimate, not a wishful one. The discipline is to put numbers on paper, then halve the upside and double the downside, and check whether the project still pencils out. If yes, do it. If no, reduce scope until it does.
Calculating annual labour saved
- Hours saved per week × 52 × loaded hourly rate
- Loaded hourly rate = (annual salary × 1.4 overhead multiplier) ÷ 1,800 working hours
- Example: support agent at £30,000/yr salary → £23/hr loaded. 15 hours/week saved → £18,000/yr labour saved.
- Discount the saved hours by 30% to account for the time still spent on AI-assisted exception handling
Calculating annual revenue gained
- Voice agent: extra calls answered × answer rate × meeting conversion × deal close rate × average deal value
- AI SDR: extra meetings booked × pipeline conversion × average deal value
- Be conservative — use your existing team's conversion rates, not the AI agent's vendor-promised lift
- If you can't tie a dollar number to revenue, set this to zero — labour savings alone often justify the project
Calculating annual run cost
| Cost item | Typical SMB range |
|---|---|
| Maintenance retainer | £4,800–£21,600/yr |
| LLM token usage | £360–£6,000/yr |
| Voice telephony (if applicable) | £600–£10,000/yr |
| Vector database hosting | £360–£2,400/yr |
| Monitoring (Langfuse, Sentry) | £0–£1,200/yr |
| Specialist data (Apollo, Clay for AI SDR) | £1,200–£6,000/yr |
Validated benchmarks from real 2026 deployments
| Project type | Typical year-1 ROI | Payback window |
|---|---|---|
| Voice agent (service business) | 300–600% | 30–60 days |
| AI SDR (B2B services) | 200–400% | 60–120 days |
| Tier-1 support agent (e-commerce) | 150–300% | 90–180 days |
| Invoice processing automation | 250–500% | 60–120 days |
| Custom AI-powered CRM | 100–250% | 180–365 days |
How to sanity-check a vendor's ROI claims
When a consultancy or vendor presents an ROI projection, run it through three checks. One: are they using your actual labour cost or a generic UK average? Two: does the revenue lift assume the same conversion rates your team currently achieves, or a hypothetical lift? Three: are all run costs included, or only the consultancy retainer?
If any of those three is generous, halve the projected ROI and re-decide. A project that still pencils out at half the projected return is a confident yes. A project that only works at the optimistic projection is a maybe.
The three numbers people get wrong
- Time saved per task — measured optimistically, from the best case rather than the average including exceptions. Time the task properly for a week before using the number.
- Fully loaded hourly cost — salary alone understates it substantially once employer NI, pension, software, management overhead and non-productive hours are included.
- Adoption rate — the assumption that everyone uses it from day one. Real adoption curves are gradual and some people never switch. Model 60-70% for the first quarter, not 100%.
Costs to include on the other side
An ROI model that counts only the build price against the salary saved will always look excellent and will usually be wrong.
- Build cost, and integration work specifically, which is often 30-50% of it
- Monthly token spend at realistic production volume, not pilot volume
- Internal time on setup, data cleanup and oversight — the largest hidden line
- Ongoing maintenance and periodic re-testing as models change
- The cost of errors during the ramp period
A sanity check before you commit
If the model only works at 100% adoption, best-case time savings and zero maintenance, it does not work. Re-run it at 60% adoption, average-case savings and realistic running costs. If it still pays back inside a year, the project is probably sound.
Also check what happens if the agent handles only the straightforward two-thirds of cases and humans keep the rest. That is the likely real outcome, and a business case that survives it is one you can defend afterwards.
Frequently asked
What's a good ROI to commit to?
Conservatively projected 150% year-one ROI is the floor most SMBs should require. Below that, the project doesn't justify the management time and integration risk. Above 300% conservatively projected is a strong case to greenlight without much further analysis.
How long should I project for?
Year one is the right time horizon for the green-light decision. After year one, the run cost continues but the setup cost is recovered, so cumulative ROI compounds. Most projects we ship are still operational and net-positive in year three with minor model and prompt updates.
What if my labour costs are lower than UK averages?
Use your actual loaded rate, not a benchmark. If you're paying £18/hr instead of £25/hr, the labour-saved component shrinks and revenue-gained becomes more important to the case. AI SDR and voice agent projects (revenue-led) hold up better than pure cost-saving projects in low-cost-labour environments.
What payback period is realistic?
Run the model at 60-70% adoption, average-case rather than best-case time savings, and realistic running costs including internal oversight. If it still pays back inside a year on those assumptions, it is probably sound. Models that only work at 100% adoption and best-case savings do not survive contact with reality.
What is the most commonly missed cost?
Internal time — setup, data cleanup and oversight during the ramp. It rarely appears in a business case and is frequently the largest line after the build itself.