Why Most AI ROI Calculations Go Wrong
Most businesses either don't calculate AI ROI at all — they build because AI feels like the right direction — or they calculate it using assumptions that are too optimistic and metrics that are too vague.
"We expect significant efficiency improvements" is not a business case. "We handle 800 support tickets per month at £12 per ticket; an AI agent that deflects 60% saves £5,760 per month against a £9,000 build cost with a 1.6-month payback" is a business case.
This template gives you the framework to calculate ROI before you build and measure it after. Use it to decide whether to invest, to hold your developer accountable to outcomes, and to report impact to whoever in your organisation cares about the numbers.
Part 1: Pre-Build ROI Calculation
Step 1: Define the Workflow
Start with a specific, bounded workflow. "Improve customer experience" is not a workflow. "Handle inbound support ticket volume via email" is.
Write one sentence: The agent will handle [specific interaction type] via [channel] for [user type].
Example: "The agent will handle inbound order status and return request queries via email for retail customers."
Step 2: Measure Current Volume
How many times per month does this interaction happen? Be precise.
| Interaction type | Monthly volume |
|---|---|
| Order status queries | ___ |
| Return requests | ___ |
| Product questions | ___ |
| Total | ___ |
If you don't have exact data, estimate conservatively. Spend an hour in your inbox or support system actually counting.
Step 3: Calculate Current Cost Per Interaction
Time per interaction: How long does it take a human to handle one of these end-to-end? Include reading, research, writing, and any follow-up.
Fully-loaded staff cost: What does one hour of that person's time actually cost? Include salary, employer contributions, benefits, and an overhead allocation — check gov.uk for current employer National Insurance and pension contribution rates if you need exact figures. As a rough rule: annual salary × 1.4 ÷ 1,800 working hours = hourly cost.
Cost per interaction: Time (hours) × hourly cost = £___ per interaction
Monthly cost: Cost per interaction × monthly volume = £___ per month
Annual cost: Monthly cost × 12 = £___ per year
Step 4: Estimate Automation Rate
What percentage of interactions can the agent handle without human involvement? This depends on:
- How well-defined the interactions are (narrow scope = higher automation rate)
- How many edge cases exist — the same edge cases surfaced through thorough pre-launch testing
- How good the agent's knowledge base is
Conservative estimate for a well-scoped first deployment: 55–65%
Realistic target after 90-day calibration: 65–75%
Use 60% as your planning assumption unless you have specific reasons to expect higher or lower. If your gut says 80%, halve it and recheck the case at that lower number. Better to be pleasantly surprised than to discover the project doesn't pay back in month four.
Monthly interactions automated: Total monthly volume × 0.60 = ___ interactions
Step 5: Calculate Monthly Saving
Monthly saving (gross): Interactions automated × cost per interaction = £___ per month
Less: AI agent running costs: Hosting, LLM API (see provider pricing documentation for usage-based cost structures), maintenance retainer = £___/month (typically £300–£1,500 depending on volume and support level)
Monthly saving (net): Gross saving − running costs = £___ per month
Step 6: Calculate Build Cost and Payback
Build cost: Get a specific quote for your scope. Don't use a range for planning purposes — use a realistic estimate for your specific requirements.
Payback period: Build cost ÷ monthly net saving = ___ months
One-year ROI: (Monthly net saving × 12 − build cost) ÷ build cost × 100 = ___% annual ROI
Decision threshold: Payback under 6 months and one-year ROI above 100% is a strong case to proceed. 6–12 months is worth doing, but the case is more sensitive to your assumptions — so revisit them. Above 12 months, rework the scope before committing.
Pre-Build ROI Template (Completed Example)
| Item | Value |
|---|---|
| Interaction type | Order status and return requests |
| Monthly volume | 650 interactions |
| Time per interaction | 8 minutes |
| Fully-loaded hourly cost | £22/hour |
| Cost per interaction | £2.93 |
| Monthly current cost | £1,905 |
| Automation rate (assumed) | 62% |
| Interactions automated/month | 403 |
| Monthly gross saving | £1,180 |
| Monthly running costs | £380 |
| Monthly net saving | £800 |
| Build cost | £8,500 |
| Payback period | 10.6 months |
| One-year net ROI | 13% |
In this example, the payback is longer than ideal. Before proceeding: can scope expand to include more interaction types (which increases volume and saving)? Is the automation rate assumption too conservative? Or is this genuinely a workflow that doesn't pay back in a reasonable window — in which case the honest answer is to not build it yet.
Part 2: Post-Launch Measurement
The Five Metrics That Matter
1. Deflection Rate
Definition: Percentage of in-scope interactions fully resolved by the agent without human involvement.
How to measure: Total interactions handled by agent ÷ total interactions in scope × 100.
What good looks like: 60–75% for a well-scoped agent after 90 days, with a rising trend as the agent is tuned.
What to do if low: Review which interaction types are failing to automate. Knowledge base gap? Scope issue? Classification error? Each has a different fix.
2. Customer Satisfaction (CSAT)
Definition: Customer rating of their interaction with the agent, typically 1–5.
How to measure: Post-conversation survey sent automatically after agent-handled interactions. Aim for 20%+ response rate.
What good looks like: 4.0+ out of 5, or roughly your human-handled support benchmark.
What to do if low: Read the negative-rated conversations. Speed problem? Accuracy problem? Tone problem? Each has a different fix.
3. First Contact Resolution Rate
Definition: Percentage of interactions fully resolved in a single session — no follow-up required.
How to measure: Interactions resolved in one session ÷ total interactions.
What good looks like: 70%+ for in-scope interactions.
What to do if low: Customers aren't getting complete answers. Review the incomplete interactions and find where the answers fall short.
4. Response Time
Definition: Average time between customer contact and agent first response.
How to measure: Logged in your communication platform or monitoring system.
What good looks like: Under 60 seconds for text-based agents. Under 5 seconds is excellent.
What to do if high: Investigate infrastructure latency, LLM API response time, and whether requests are queuing.
5. Human Hours Saved
Definition: Actual reduction in staff hours spent on in-scope interactions.
How to measure: Track time spent on in-scope interactions before and after deployment. A monthly survey of the responsible team is fine for smaller operations; time-tracking tools for larger ones.
What good looks like: Approaching your pre-build automation rate estimate.
What to do if lower than expected: Deflection rate tells you whether the agent is handling interactions. If deflection is high but hours saved are low, the interactions the agent is handling may be shorter than average. Re-examine your cost-per-interaction assumption — you may have averaged across a workload that wasn't actually uniform.
Monthly Reporting Template
Use this for monthly reporting in the first year:
AI Agent Performance — [Month Year]
| Metric | Target | This Month | Last Month | Trend |
|---|---|---|---|---|
| Deflection rate | 65% | ___% | ___% | ↑/↓/→ |
| CSAT score | 4.0+ | ___ | ___ | ↑/↓/→ |
| First contact resolution | 70% | ___% | ___% | ↑/↓/→ |
| Avg response time | <60s | ___s | ___s | ↑/↓/→ |
| Monthly interactions handled | ___ | ___ | ___ | ↑/↓/→ |
| Human hours saved | ___ hrs | ___ hrs | ___ hrs | ↑/↓/→ |
| Running cost | £___ | £___ | £___ | — |
| Net monthly saving | £___ | £___ | £___ | ↑/↓/→ |
Actions this month: [What was changed or tuned based on last month's data]
Actions planned next month: [What will be changed or tuned based on this month's data]
Calculating Cumulative ROI
Track cumulative ROI monthly to see the payback curve:
| Month | Net saving | Cumulative saving | Build cost remaining |
|---|---|---|---|
| 1 | £___ | £___ | £___ |
| 2 | £___ | £___ | £___ |
| 3 | £___ | £___ | £___ |
| ... | |||
| Payback month | — | = Build cost | £0 |
When cumulative saving equals build cost, payback has happened. Every month after is pure return.
Where the ROI Numbers Lie
Two honest caveats. First, the "human hours saved" only converts into real money if those hours actually go somewhere useful — backfilling a hire you didn't make, redirecting people to higher-value work, or genuinely reducing headcount. We've seen agents that deflected 65% of tickets while the support team's hours stayed exactly the same and the company just got slower at responding to the remaining 35%. The agent worked. The P&L didn't move. Decide before you build where the freed time goes.
Second, ROI math is sensitive to assumptions in a way that's easy to underestimate. A 60% automation rate vs 45% changes the payback period dramatically. If the difference between "great investment" and "borderline" is a 15-point swing in deflection, the case isn't as strong as it looks. Run the numbers at both ends of your plausible range and look at the worse one. If it still works, you're in good shape.
Common ROI Mistakes
Too optimistic an automation rate. Assume 60% unless you have specific evidence for more. Optimism early makes the business case look better than reality.
Ignoring running costs. The agent costs money to operate every month. Include those in every calculation.
Measuring deflection without measuring satisfaction. High deflection with low satisfaction means the agent is blocking customers from getting help, not helping them.
Declaring success at launch. ROI accrues over time. Month one is setup and calibration. Evaluate at 3 months, 6 months, and 12 months.
Not tracking the counterfactual. If your support volume grows 30% over the year — including during seasonal peaks — and your human hours stay flat, the agent is delivering significant value, but you won't see it unless you also track what that growing volume would have cost without the agent.
Related guides
- AI agent development cost: pricing tiers
- How much does it cost to build an AI agent?
- The ROI of AI automation: real first-90-day results
- AI agent analytics: the metrics that matter
- AI agent pricing models: fixed project vs retainer
- Our AI agent development services
If you want help building a business case against realistic assumptions — including the version where we tell you not to build — that's what we do.
Talk to us about building your business case — no commitment, just a conversation.
Frequently Asked Questions
What is a realistic ROI timeline for an AI agent?
Most well-scoped AI agents reach payback between 6 and 12 months from launch. Narrowly defined workflows with high interaction volume — such as order status queries or appointment scheduling — can hit payback in 3 to 6 months. Broader or more complex agents typically take longer as they require more calibration before deflection rates stabilise.
What automation rate should I use in my ROI calculation?
Use 60% as your baseline planning assumption unless you have specific data pointing higher. Agents handling tightly bounded, repetitive queries often achieve 65 to 75% after 90 days of tuning. Using an overly optimistic rate — anything above 75% for a first deployment — makes the business case look stronger than it is and increases the risk of a disappointing post-launch result.
How do I calculate the fully-loaded cost of a human handling a support interaction?
Take the annual salary of the person handling the interaction, multiply by 1.4 to account for employer contributions, benefits, and overhead, then divide by 1,800 working hours per year. That gives you a fully-loaded hourly rate. Multiply by the average time per interaction in hours to get cost per interaction. Avoid using just base salary — it understates the true cost by 30 to 40%.
Should I include AI agent running costs in my ROI calculation?
Yes, always. Running costs — LLM API usage, hosting, and any maintenance or support retainer — typically range from £300 to £1,500 per month depending on volume and your support arrangement. Excluding them makes the monthly net saving look higher than it is and shortens the apparent payback period. Every ROI calculation should show gross saving, running costs as a separate line, and net saving.
What metrics should I track after launching an AI agent?
Track five metrics monthly: deflection rate (percentage of interactions fully resolved without human involvement), CSAT score (customer satisfaction rating post-interaction), first contact resolution rate, average response time, and actual human hours saved. Together these tell you whether the agent is handling volume, whether it is doing so satisfactorily, and whether the time savings are materialising on the P&L.
How do I know if my AI agent is actually saving money versus just handling interactions?
Deflection rate measures whether the agent is handling interactions. Hours saved measures whether that handling is translating into reduced staff time. If deflection is high but hours saved are low, the interactions being automated may be shorter than average, or staff are spending the freed time on lower-priority tasks rather than reducing headcount or backfilling a hire. Agree before launch on exactly where the freed time goes — otherwise the agent can work technically while the P&L stays flat.
What if my ROI calculation shows a payback period longer than 12 months?
Before rejecting the project, check three things: whether expanding scope to include additional interaction types increases volume enough to shorten payback, whether your automation rate assumption is too conservative, and whether running costs can be reduced with a different hosting or support arrangement. If the case still doesn't work after those adjustments, the honest answer is to defer until volumes grow or to find a higher-value workflow to automate first.