AI ROI (return on investment) is the net value an AI system creates, in savings or extra revenue, divided by everything it costs to build and run, over a stated period. For an AI agent in customer service, the value usually comes from contacts resolved without a person; the costs include the build, model usage, maintenance, evaluation, and change management.
The formula is ordinary. The difficulty is in the inputs. AI projects tend to overstate the benefit, because a demo resolves more than live traffic does, and to understate the cost, because the work that keeps a model reliable after launch is easy to leave out of a spreadsheet. Machine learning systems are known to carry large ongoing maintenance costs that early wins hide.4
Finance teams have reason to be skeptical. Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, with escalating costs and unclear business value among the causes.1 A credible ROI case is the main defense against both.
The formula, with a worked example
For a customer service agent, start from the contacts it resolves on its own. Each one saves the cost of a human-handled contact, minus what the AI costs to handle it.
- Contacts per year
- 600,000
- Cost per human-handled contact
- €4.00
- Automation rate (resolved without a person)
- 50%
- Value: 600,000 × 50% × €4.00
- €1,200,000
- Model and hosting: 300,000 × €0.40
- €120,000
- Maintenance, evals, monitoring (year)
- €180,000
- Change management and training (one-off)
- €60,000
- Build (one-off)
- €300,000
- Total cost, year 1
- €660,000
- Net value, year 1
- €540,000
- ROI, year 1: €540,000 / €660,000
- 82%
Illustrative example. Every input is invented to show the arithmetic, not a benchmark. Replace each one with your own figures from the helpdesk, finance, and the vendor quote.
Two other views help. Payback period: here the running margin is about €75,000 a month (€1,200,000 minus €300,000 of yearly run costs, divided by 12), so the €360,000 of one-off costs pays back in under five months. Year 2 ROI is higher, because the one-off costs drop out. Present all three, and say which assumptions they depend on.
What to count, and what people forget
Most weak ROI cases are missing a line, not wrong on arithmetic. The right-hand column is where business cases usually come up short.
| Category | Usually counted | Often forgotten |
|---|---|---|
| Value | Contacts resolved without a person, times the cost per contact | Faster resolution for customers, after-hours coverage, and revenue from shopping or lead agents |
| Build | Vendor or team cost to design and integrate the agent | Security review, legal and DPO time, data preparation, and integration changes in other systems |
| Model usage | Price per conversation at launch volume | Longer conversations, retries, tool calls, and growth in volume. Providers price per million input and output tokens.6 |
| Maintenance | Bug fixes | Updating prompts and knowledge when policies change, model version upgrades, and integration breakage |
| Evaluation | Testing before launch | Ongoing scoring of live conversations and a regression test set for every change5 |
| Change management | Agent training on the new tool | Redesigned workflows, new roles for reviewing the AI, and time for the team to trust it |
| Risk | Not usually priced | Wrong actions, refunds issued in error, complaints, and compliance work |
Where the value comes from
Cost per contact is the easiest value to measure, so it dominates most business cases. It is not the only one.
Contacts resolved
The direct saving: contacts that no longer need a person. This is the core of the formula above.
Faster human work
AI that assists people also counts. In a study of 5,179 customer support agents, access to a generative AI assistant raised issues resolved per hour by 14% on average, and by 34% for novice and low-skilled workers.3
Revenue
Shopping and lead agents add value through conversion, not cost: in a Tidio Agents case study, ADT saw +17% conversion. Count revenue as margin, not gross sales.
Capacity
Handling peak seasons without temporary hiring, or covering nights and weekends. Value it at the cost of the alternative you would otherwise pay for.
How the ROI case differs by agent type
The structure of the formula stays the same. The value line and the hardest cost change with the agent.
Retail and marketplaces
Value is contacts resolved on order status, returns, and delivery problems. The hardest cost is keeping policies and integrations current.
volume × automation × cost per contactOnline retail
Value is extra margin from conversion and basket size. Measure it with a controlled test, not before-and-after totals that include seasonality.
sessions × conversion lift × marginTelecoms
Value is call minutes removed from the contact center. Count telephony and speech costs on top of model costs.
calls × automation × cost per minutePayments and claims
Value is resolved contacts plus faster cycle times. Compliance review and audit logging are larger cost lines than in retail.
contacts + cycle time − compliance costInvoices and procurement
Value is hours of manual entry and matching saved, and fewer payment errors. Count the review queue that stays with people.
documents × touchless rate × minutes savedInbound sales
Value is qualified meetings booked outside office hours and faster response to inbound leads, priced at pipeline conversion.
leads × qualification × win rateCommon ROI mistakes, and how to avoid them
Gartner reported that generative AI deployment approaches it studied came with costs ranging from $5 million to $20 million, and that costs are less predictable than for other technologies.1 Agentic AI projects face the same scrutiny: Gartner predicted over 40% would be canceled by the end of 2027, with escalating costs and unclear business value among the reasons.2
Automation rate from a demo
ControlUse the rate measured on real traffic in a proof of concept, and model a lower case alongside it.
Deflection counted as resolution
ControlA customer who gives up or calls instead still costs money. Count repeat contacts within a few days against the automation rate.
Year 1 only
ControlOne-off costs make year 1 look worse and later years better. Show at least three years, with model price and volume assumptions stated.
Missing run costs
ControlAdd maintenance, evaluation, monitoring, and model upgrades as yearly lines, not a contingency percentage.
Savings that never materialize
ControlSaved agent hours are only a saving if staffing, overtime, or outsourcing spend actually changes. Agree with operations how capacity will be used.
Common questions
How do you calculate the ROI of an AI agent?
Multiply contact volume by the automation rate and the cost per human-handled contact to get the value. Subtract build, model, maintenance, evaluation, and change management costs. Divide the result by the total cost for the same period.
What is a good ROI for AI in customer service?
There is no reliable industry benchmark, because cost per contact and automation rates vary widely by company and request type. Build the case from your own volumes and costs, and test the automation rate on your own data.
Which AI costs are most often missed?
Ongoing maintenance when policies and systems change, evaluation of live conversations, model version upgrades, integration work in other systems, and the time it takes teams to change how they work.
How long does it take for an AI agent to pay back?
It depends on volume, cost per contact, and how much of the traffic the agent resolves. High-volume support operations with repetitive requests pay back fastest. Calculate it with the formula above rather than relying on a general figure.
References
- Gartner. Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025. Press release, 29 July 2024.
- Gartner. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. Press release, 25 June 2025.
- Brynjolfsson, E., Li, D., and Raymond, L. Generative AI at Work. NBER Working Paper 31161; Quarterly Journal of Economics 140(2), 2025.
- Sculley, D. et al. Hidden Technical Debt in Machine Learning Systems. NeurIPS 2015.
- Anthropic. Define success criteria and build evaluations. Claude documentation.
- Anthropic. Pricing. Claude documentation.