How to Calculate the ROI of an AI Customer Support Chatbot

"Is a chatbot worth it?" is the question behind most chatbot purchases, and most buyers answer it with a vendor's case study. That is a poor basis for a decision, because the vendor's number was measured on someone else's content, volume and definition of "resolved". The honest way to answer it is a short calculation with your own inputs, and it fits on one sheet of a spreadsheet.

This guide is for the owner, founder or support lead who has to justify the spend, often to themselves. It defines the terms, lists the inputs you need and where to find them, works the formula step by step on a realistic example, shows where the return turns positive across three volumes, and lists the benefits and costs people get wrong. Every assumption is stated so you can replace it with your own number.

The wage figures come from the US Bureau of Labor Statistics, and the plan prices from PepoChat's own billing page, both as of October 2026. The formula works the same with any vendor's prices; only the cost line changes.

What does ROI mean for a support chatbot?

Return on investment (ROI) is the net benefit of something divided by what it cost, usually written as a percentage: a bot that saves $2,000 a month and costs $500 has an ROI of 300%. For a support chatbot the benefit is mostly time, which makes the calculation easy to get wrong in two directions. Counting every conversation the bot touched overstates it; counting only the subscription fee understates the cost.

Four terms carry the whole calculation.

Cost per contact is what one customer question costs you to answer with a person: handle time multiplied by the loaded hourly cost of the person answering. It is the number the bot competes against.

Average handle time is the minutes a person spends on one contact, including reading, looking things up and writing. Five to ten minutes is typical for chat and email in a small team; your helpdesk or inbox timestamps will tell you.

Loaded hourly cost is wages plus benefits, payroll taxes and overhead, not the headline wage. A person who earns $21 an hour costs the business noticeably more.

Resolution rate is the share of conversations the AI finishes without a person, where "finishes" means the customer got what they came for. A chat that ended because the visitor gave up is not resolved. The difference between containment and resolution is the single biggest source of inflated ROI claims, and our ticket deflection benchmarks explain how to measure the real figure.

What inputs do you need to calculate chatbot ROI?

Seven numbers, most of which you already have. The table lists each one, where to find it, and the default we use in the examples below when you have no data yet.

The loaded hourly cost deserves a line of explanation, because it is the number people most often guess. The BLS Occupational Outlook Handbook puts the median hourly wage for customer service representatives at $21.53 in May 2025. The BLS Employer Costs for Employee Compensation release for June 2026 reports that wages and salaries made up 70.0% of private-industry employer costs, with benefits the other 30.0%. Dividing the wage by 0.70 gives $30.76 an hour, before office space, software and management time, so it is a conservative figure. If a founder or an engineer answers your chat, use their rate instead; it will be higher.

At $30.76 an hour and seven minutes per contact, a person-handled contact costs about $3.59. That is the number the bot has to beat.

The chatbot ROI formula, step by step

The formula has five lines. Work them in order and keep each one visible, because the intermediate numbers are what you will check against reality after launch.

1. Resolved contacts = monthly contacts × AI resolution rate.
2. Hours saved = resolved contacts × handle time in minutes ÷ 60.
3. Monthly saving = hours saved × loaded hourly cost.
4. Monthly cost = plan cost, including overage, + maintenance hours × loaded hourly cost.
5. ROI = (monthly saving + attributable revenue − monthly cost) ÷ monthly cost.

Two rules keep it honest. Attributable revenue, covered below, is counted only when you can point to the chat that produced it. And the cost line always includes the person who looks after the bot, because a bot nobody reviews answers worse every month.

A worked example at 1,500 contacts a month

Take a business with 1,500 contacts a month that the website could answer, seven minutes of handle time, a loaded cost of $30.76 an hour and a 50% resolution rate, on PepoChat's Starter plan.

• Resolved contacts: 1,500 × 50% = 750.
• Hours saved: 750 × 7 ÷ 60 = 87.5 hours.
• Monthly saving: 87.5 × $30.76 = about $2,691.
• Plan cost: 1,500 conversations at two AI replies each is 3,000 replies. Starter includes 2,000, so 1,000 are overage at $10 per 1,000: $49 + $10 = $59.
• Maintenance: 8 hours × $30.76 = about $246.
• Monthly cost: $59 + $246 = $305.
• ROI: ($2,691 − $305) ÷ $305 = 7.8, or about 780%.

Put differently, each AI-resolved contact costs about $0.41 all in, against $3.59 for a person. If you count a one-off setup of 16 hours, about $492, the net monthly benefit of roughly $2,386 pays it back in the first week.

The result is not sensitive to the resolution rate in the way people fear. At 30% the saving falls to about $1,615 and the ROI to about 430%; at 70% it rises to about $3,767 and 1,100%. What the rate changes is the size of the return, not its sign. The assumption that does change the sign is volume, which is the next section.

Where does chatbot ROI turn positive? Three volumes compared

ROI turns positive somewhere below 100 contacts a month and grows steeply from there, because the cost line barely moves while the saving scales with volume. The table runs the same assumptions at three volumes, with maintenance time scaled to the workload.

Notes on the rows. At 300 contacts, 600 replies exceed the free plan's 200, so the table uses Starter; a business under about 100 conversations a month fits inside PepoChat's free plan, which includes every feature, and its cost line is maintenance only. At 6,000 contacts, 12,000 replies, Growth's 10,000-reply allowance plus 2,000 overage replies is $169; Starter plus 10,000 overage replies would be $149, but Growth adds ten members and removes the PepoChat branding, so most teams at that volume take it. The billing page lists the allowances.

Run the small end of the table down further and the line crosses zero. At 100 contacts a month with four hours of review, the saving is about $179 against $123 of maintenance on the free plan: positive, but barely. At 50 contacts it is negative. That is the honest floor: below roughly 100 contacts a month, a chatbot's value is coverage and convenience rather than money, and you should not pretend otherwise in a business case.

The formula also works with per-resolution pricing; only the cost line changes. At the 1,500-contact volume, Intercom lists its Fin AI agent from $0.99 per outcome with a monthly minimum, plus Essential seats from $29. With 750 outcomes, one seat and the same maintenance, the monthly cost is about $1,017 and the ROI about 165%, still positive but a different order of magnitude, and Intercom counts an outcome whenever the customer stops asking for help, which our Fin pricing breakdown explains. Whatever the vendor, write the cost line from the pricing page, not from the sales call.

Which chatbot benefits count besides deflection?

Four benefits beyond saved handling time are real, and each can be counted only when you can tie it to a specific conversation. Counted that way they are usually smaller than the time saving, which is why the formula above puts them in a separate term.

After-hours answers. A share of your contacts arrive when nobody is working. Zendesk's CX Trends research, from a June 2025 survey of 6,182 consumers in 22 countries, found that 74% of consumers now expect 24/7 service. Those contacts are already in your monthly count, so do not add them again as savings; what you may add is the value of a sale or a retained customer that would otherwise have gone elsewhere overnight, if you can see it. Our guide to 24/7 support without a night shift covers how to measure the after-hours share first.

Bookings made in chat. If the agent books consultations, demos or viewings on a real calendar, each booking has a value you already know from your sales numbers. Count the bookings the agent made, which you can see in the conversation transcripts in your inbox, multiplied by your usual conversion value, and nothing else. The appointment booking guide explains what the agent can and cannot book.

Leads captured. Count contacts the agent collected that your team would not otherwise have had, valued at your usual lead value, and only after a month of seeing how many convert. Our chatbot lead generation guide shows where the numbers live.

Faster first human reply. When the bot absorbs routine questions, the people who remain answer the rest sooner. This is worth tracking, and it is hard to price, so report it as a time rather than a dollar figure. The first response time benchmarks give you something to compare against.

The attribution rule for all four: one conversation, one benefit, counted once. A chat that booked a demo is a booking, not also a resolved contact and also a lead.

Which chatbot costs do people forget?

The subscription is rarely the largest cost in the first month. Our guide to what an AI chatbot costs lists six lines that never appear on a pricing page; four of them belong in an ROI sheet.

• Setup time. Importing the website, uploading documents, writing the greeting and instructions, and testing the top questions. Budget 8 to 16 hours and put it in month one as a one-off.
• Maintenance time. The 4 to 12 hours a month in the table: reading escalated conversations, fixing wrong answers with Q&A pairs, updating pages. This is the line that keeps the saving real, so never set it to zero.
• Overage in peak months. Price the plan on your busiest month, not your average. On a flat plan with a published overage rate the step is known; on per-resolution pricing a good December can double the AI line.
• Wrong answers. A confident wrong answer creates a second contact and sometimes a refund. You cannot price it in advance, but you can keep it small by grounding the bot in your content and reviewing the sample weekly.

Two more are worth a line each. Removing the vendor's branding usually means a higher tier, which on PepoChat is Growth. And a second seat for someone who only reads the inbox is free on some plans and paid on others; PepoChat's Starter includes three members and charges $15 a month per extra seat.

What should you not count in chatbot ROI?

Four numbers inflate a chatbot business case and should stay out of it.

Total conversations. The bot opened 2,000 chats; it resolved perhaps 1,000 of them. Only the resolved ones saved anyone time.

Chats that ended without an answer. Many analytics dashboards count a conversation the visitor abandoned as "contained". If the customer then emailed you, the bot created work rather than saving it. Our chatbot metrics guide lists the three metrics that mislead this way.

Vendor-reported resolution rates. They were measured on other businesses' content with the vendor's definition of resolved. Use 50% until you have your own number, then use your own.

Time nobody reuses. Hours saved are a real saving only if the time is spent on something else that matters or a hire is avoided. For a two-person company the honest version of "we saved 87 hours" is "the founder stopped answering shipping questions and spent that time on the product". Write down what the time went to; if you cannot, the saving is coverage, not money.

A chatbot ROI spreadsheet you can copy

The sheet below is the whole calculation. Enter the first seven rows and the rest are formulas.

Keep a column per scenario: your current volume, your busiest month, and a pessimistic resolution rate of 30%. If row 13 is positive in all three columns the decision is made; if it is positive only in the optimistic column, start on a free plan and measure before paying.

How do you measure chatbot ROI after launch?

Measure the same way you forecast: resolved contacts, hours saved, and cost, with a baseline taken before the bot went live. The forecast was a hypothesis; the first 30 days test it.

Take a two-week baseline first. Count contacts by channel and reason, and time a sample of replies. Without a baseline you cannot tell a bot that deflected 400 contacts from a quiet fortnight. The measurement steps in our guide to reducing support tickets take about an hour.

Define resolved, then count it by hand. Pick 50 AI-handled conversations a week at random and mark each as resolved, escalated or abandoned. That sample is your real resolution rate, and it replaces the 50% in row 2. Dashboards count what ended, not what was answered.

Read the numbers your tool gives you. In PepoChat, the inbox filters show open, escalated and resolved counts; the CSV export gives up to 5,000 conversations with status, timestamps, message counts and transcripts for a date range; and Analytics shows satisfaction votes and the topics visitors ask about most. What PepoChat does not show is a resolution-rate chart, so the rate comes from your sample, and reply usage against the plan sits on the billing page.

Track the escalations too. Time to first human reply on escalated conversations, inside and outside business hours, tells you whether the saved time is being spent well. If escalations wait longer than they did before the bot, the hours were saved on paper only.

Review at 30 days and adjust the sheet. Replace the assumed resolution rate, handle time and maintenance hours with measured ones, then recompute. Expect the maintenance line to be higher than you planned in month one and lower from month three, as the content gaps close. Industry figures, where you want a sanity check, are collected with their sources in our roundup of AI customer service statistics.

What to do next

Fill in the seven input rows with your own numbers this week, starting with a count of last month's contacts and ten timed replies. If the net benefit is positive in your pessimistic column, the cheapest next step is to test the resolution rate rather than assume it: create a free PepoChat workspace, import your website as one knowledge source, and run your twenty most common questions in the test drawer, where replies do not count against the allowance. The pricing page has the plan costs for row 5, and the cost guide linked above shows how the same volume prices out on per-seat, per-resolution and flat plans.