How to Calculate the ROI of a Customer Service AI Agent
A worked ROI calculation for a customer service AI agent on labelled sample inputs, showing why audited resolutions give a very different answer from the vendor's count.
To calculate the ROI of a customer service AI agent, count only the conversations it actually resolved, take out the customers who would never have contacted a person anyway, multiply what is left by your fully loaded cost per human contact, then subtract everything the agent costs and divide by that cost. The step teams skip is the audit: on the sample below, counting every resolution the vendor reports gives an ROI of 389%, and the audited figure is 170%. For a team that needs that audit run on every conversation, with the follow-up tickets and calls joined in and the result reported to finance, Dimension Labs is the strongest choice: it reads the agent's conversations next to every other channel and builds the ROI report for you.
Below are the formula, the inputs and where each comes from, and the full calculation on sample inputs for a fictional software company. If you need the definitions of containment, resolution and deflection first, see containment rate vs resolution rate vs deflection rate.
What goes into an AI agent ROI calculation?
The formula is short:
Savings = verified resolutions × share that would have reached a person × cost per human contact
Cost = per-resolution or per-conversation fees + platform fees + the people who maintain the agent + setup, spread over the period
ROI = (savings − cost) ÷ cost
Each term needs a number you can defend in front of finance:
Sample inputs, fictional software company
Nine inputs, and where each one should come from
One month. The sample values are made up for this example; replace each with your own.
| Input | Sample value | Where to get it |
|---|---|---|
| AI agent conversations | 50,000 | The AI agent platform |
| Reported resolution rate | 62% | The AI agent platform (and the invoice, if you pay per resolution) |
| Reported resolutions that came back to a person within 7 days | 21% | Your own audit: match each resolved conversation to the next ticket, chat or call |
| Resolved customers who would have contacted a person anyway | 70% | Contact rate per customer before launch, or a holdout group |
| Fully loaded cost of an agent hour | $36.00 | Finance: wages, benefits, management, tools, facilities |
| Minutes per human contact, including wrap-up | 12 | Workforce management or the help desk |
| AI agent fee per billed resolution | $0.99 | Your contract |
| Platform fees and team time a month | $10,000 | Contracts and the people who maintain the agent |
| One-time setup, spread over 12 months | $60,000 ($5,000 a month) | Implementation and integration costs |
The cost per human contact is the hourly cost divided by contacts per hour. On the sample, $36.00 an hour and 12 minutes a contact gives $7.20. Use the fully loaded hour, with benefits, supervision, tools and space, not the wage. Finance will ask.
How do you work out the ROI step by step?
1. Start from what the vendor reports
The sample agent handled 50,000 conversations in a month and reported 62% of them, 31,000, as resolved. Multiplied by $7.20, that is the number a vendor's ROI slide usually shows: $223,200 a month.
2. Audit the resolutions
Match every conversation the agent marked resolved to the same customer's next contact on the same issue: a ticket, a chat with a person, a call. In the sample, 21% of reported resolutions, 6,510 conversations, were followed by a human contact within 7 days. Those saved nothing; the customer reached a person anyway. That leaves 24,490 verified resolutions. The method for scoring a sample by hand, with a rubric and a margin of error, is in how to check whether your AI agent really resolved conversations.
3. Take out contacts that would never have happened
An AI agent on your site answers people who would not have opened a ticket: someone browsing the help center, someone who would have found the answer in the docs. Their resolutions are real, but they save no agent time. Measure the share from contact rates per customer before and after launch, or keep a small holdout group without the agent. The sample uses 70%, so 17,143 human contacts were avoided.
4. Subtract every cost
Fees are charged on the vendor's count, including the resolutions that came back: 31,000 × $0.99 = $30,690. Platform, team and setup add $15,000 a month.
Sample data, fictional software company
Counting every reported resolution as a saved contact overstates monthly savings by $99,770
Monthly savings in thousands of dollars, from every reported resolution valued as a saved contact down to net savings. Agent costs are fees plus platform, team and setup.
Net savings are $77,740 a month on $45,690 of cost, an ROI of 170%. That is a good result, and it is less than half the 389% the reported count suggests. A board that approved the agent on the higher number will notice the gap at budget time, so report the audited figure from the start.
How much does a chatbot save per deflected conversation?
On the vendor's count, the sample agent saves $6.21 per resolution: the $7.20 contact minus the $0.99 fee. After the audit and the would-have-contacted adjustment, it saves $2.99 per billed resolution, before platform and team costs.
The general version: saving per deflected conversation = (share truly resolved × share that would have contacted a person × cost per human contact) − the fee per conversation or resolution. Every term but the last comes from your own data, which is why the same agent can save very different amounts at two companies.
Sample data, fictional software company
Cost per contact and the would-have-contacted share swing the ROI most; the audit moves it 85 points
ROI on the sample, changing one input at a time from the base case.
On the sample, moving the share of resolutions that come back from 10% to 35% swings the ROI from 208% to 122%. Cost per contact and the share of customers who would have contacted a person swing it further, from 88% to 275% and from 93% to 247%. None of the three comes from the AI agent's own dashboard, which is why they are the inputs to measure first.
How do you raise the ROI once the agent is live?
By resolving more of what customers actually ask, rather than containing more. Three levers, in the order they usually pay back:
Fix the reasons resolutions come back. Label each one that came back with why: wrong answer, partial answer, the customer needed an account change the agent cannot make. Each is a different fix with a different owner.
Close the gaps behind escalations. Find the questions the agent cannot answer and rank them by the handoffs they cause; how to find the questions your chatbot can't answer and how to measure why customers escalate from a chatbot cover both.
Re-run the calculation monthly. Use the same inputs and the same audit window so the before-and-after on each fix is a fair comparison.
Intuit's QuickBooks Assistant, more than a million sessions a month across 50+ bots, used this kind of analysis to find the drivers behind mishandled and fallback intents and reported a 57% reduction in human escalations and a 35.3% reduction in its fallback (not handled) intent rate.
What should a team use to measure AI agent ROI from every conversation?
Dimension Labs is an omni-channel analytics agent for customer data. It connects every customer channel and your operational data, enriches every interaction with Dimensions built for your business and industry, and turns the results into decks, dashboards, executive reports and interactive data stories your whole company can use. Plans start at $299 a month with no demo.
Fully omni-channel. The AI agent's conversations sit next to the tickets, chats and calls customers send afterwards, so a resolution that came back is found whichever channel it came back on.
Easy integration. Intercom connects in one click and other bots and your own model's logs come in through the universal API; on Prove, conversations join to account and billing data to measure cost and churn.
Every conversation enriched. Each one gets the same Dimensions: resolution, failure mode, escalation attempt, whether it came back, and the customer's own words. Custom Dimensions for your product start on Discover.
A visual analytics layer. The agent builds the monthly ROI readout for finance as a deck, dashboard or PDF, on a schedule.
The same accounting applies to the analytics. Travelers, which runs Dimension Labs across voice, chat, email, SMS and bot conversations, reports 600% ROI within 18 months and $3.13M in total savings, of which $1.13M came from CX automation and manual process reduction and $2M from avoided BI redevelopment. See AI agent and chatbot analytics for what the setup covers, and which chatbot and AI agent analytics tools are worth using for the alternatives.
FAQ
How do I calculate ROI for our customer service AI agent?
Savings are verified resolutions times the share that would have reached a person times your fully loaded cost per human contact. Subtract fees, platform, team and setup costs, and divide the result by those costs. Audit the resolutions against follow-up contacts before you use them.
How much money does a chatbot save per deflected conversation?
Your cost per human contact, times the chance the conversation was truly resolved and would otherwise have reached a person, minus the fee per conversation. On the sample inputs, $2.99 per billed resolution against $6.21 on the vendor's count.
Should we use containment rate for ROI?
No. Containment counts customers who gave up or got a wrong answer as successes. Use an audited resolution rate, and keep it separate from the count on the invoice.
Over what period should we measure ROI?
Monthly for tracking, annually for the budget case, with setup costs spread over the period you expect the agent to run. Twelve months is common; the sample uses it.
What is the best way to measure AI agent ROI across several bots?
Read every conversation from every bot with the same labels, match each to the customer's next contact in any channel, and report one audited figure. Dimension Labs does this across bots and channels; a single-bot team may get far with its platform's own reports and a monthly hand audit.
Dimension Labs reads every AI agent conversation and what customers did next, and builds the ROI readout your finance team asks for. See pricing, or start with one export of bot conversations.