Show your sources
Why these defaults
The preset describes a written support queue handled by an agent you build yourself. Each number is sourced, calculated or marked as an assumption.
- Monthly ticket volume8,000 ticketsAssumption
- Chat and email tickets after removing spam, duplicates and auto replies.
- Human cost per ticket$6Assumption
- Anchored to a Gartner poll that put live channels such as phone, live chat and email at $8.01 per contact. That average includes phone, so the preset trims it for written channels. Replace it with your own payroll and tooling cost divided by tickets handled.
- Handle time10 minutesAssumption
- Reading, replying and tagging a written ticket. Shown for context only.
- Automation (resolution rate)45%Assumption
- Tickets closed without a human reply. Set below most vendor headlines on purpose.
- Human review15%Assumption
- Resolved tickets a person still spot checks or reopens.
- Retry rate10%Assumption
- Extra model calls from failed tool calls, validation and clarifying turns.
- Tokens per ticket3,500 in, 400 outAssumption
- One ticket thread plus the help articles retrieved to answer it.
- ModelSonnet 5Sourced
- A mid tier model for policy heavy answers. Rates come from the site model table, which cites the Anthropic pricing page.
- Other cost per ticket$0.05Assumption
- Retrieval, embeddings, help desk API calls and logging for an agent you run yourself. To model a product priced per outcome, enter the fee times your expected resolution rate instead.
- Build and maintenance$60,000 build, $5,000 per monthAssumption
- Help desk integration, evals on past tickets, guardrails and monitoring.
- Automation decay5% per yearAssumption
- Help articles go stale and resolution drifts down unless someone owns the knowledge base.
- Discount and ramp12% per year, 3 monthsAssumption
- Kept equal to the general calculator so results stay comparable.
About this tool
What this customer support AI agent ROI calculator answers
This customer support AI agent ROI calculator estimates payback, monthly net savings and 36 month NPV for an agent that resolves chat and email tickets before they reach a person. Use it to compare building your own agent with buying one priced per outcome, on the same ticket volume and the same deflection assumptions.
Support is where most teams try agents first, and where vendor ROI claims are loudest. The preset models a written queue of 8,000 tickets a month with a resolution rate below most published headlines. Deflection only counts as savings when a ticket closes without anyone touching it, so the review share trims the headline rate before any money is counted.
How to use it
Pull the inputs from a help desk export before trusting any output.
- Enter ticket volume. Enter monthly chat and email tickets from your help desk, excluding spam, duplicates and auto replies.
- Set the cost per ticket. Divide fully loaded support payroll and tooling by tickets handled to get a real cost per ticket.
- Set resolution and review. Set automation to tickets closed without a human reply, and review to the share a person still checks.
- Choose the pricing model. For a product priced per outcome, enter the fee times your expected resolution rate as other cost per task.
- Size tokens and model. Size tokens from one ticket thread plus the help articles retrieved for it, then pick the model you would run.
- Share and challenge. Share the scenario with your support lead and ask which ticket categories the resolution rate really covers.
How deflection turns into savings
Labour saved equals ticket volume times resolution rate, reduced by the share of resolved tickets a person still reviews, times cost per ticket. Monthly LLM cost is tokens per ticket priced from the model table, multiplied by volume and the retry rate. The other cost field covers everything else charged per ticket, which is where a vendor's outcome fee belongs if you buy instead of build. Net savings subtract both lines and maintenance, and payback divides build cost by what remains.
The preset also sets automation decay to 5% a year. Products change, prices move and old answers become wrong. If someone owns the knowledge base, set decay to zero and add their time to maintenance instead.
Where support agent ROI usually breaks down
Resolution rate is the number that breaks. Some dashboards count a conversation as resolved when the customer stops replying, which includes people who gave up and opened a new ticket or picked up the phone. Count a ticket as resolved only when it stays closed for a week with no repeat contact from the same customer.
The second break is channel shift: a deflected chat that turns into a phone call is a more expensive ticket, not a saving. The third is the cost of wrong answers. A confident reply about refunds or warranty terms that contradicts policy creates escalations, credits and sometimes legal exposure that the labour line never shows.
When ticket deflection ROI is real
It is real for repetitive, documented questions: order status, password resets, plan and billing explanations, how to steps that already exist in the help center, and returns that follow a clear policy. Teams with clean help articles and tagged ticket history see the strongest results.
It is weak when most tickets are bugs that need engineering, account investigations, or enterprise customers who expect a named person. There the agent earns more by drafting replies and summarising threads, which saves minutes per ticket rather than whole tickets. Model that by lowering cost per ticket instead of raising automation.
How this differs from vendor ROI calculators
Help desk vendors publish calculators that assume their own resolution rates and often leave out their own fees. This page asks for both. Outcome pricing is common: Intercom lists Fin from $0.99 per outcome, and Salesforce lists Agentforce at $2 per conversation or $0.10 per action through Flex Credits. These are public list prices, not recommendations, and either can go into the other cost field. The cost per ticket default leans on a Gartner poll published in 2019 that put live channels at $8.01 per contact, so treat it as a dated anchor.
Test the build or buy decision with an architect
Bring your ticket mix and the scenario link, and get a second opinion on resolution rate, guardrails and whether a vendor or your own agent fits the queue.
Book a strategy callFrequently asked questions
- How do I calculate ROI for a customer support AI agent?
- Multiply monthly tickets by the share the agent resolves without a human, reduce that by the share a person still reviews, and multiply by your cost per ticket. Subtract model cost, any per ticket platform fee and monthly maintenance to get net savings. Divide the build or onboarding cost by net savings for payback. The calculator repeats this monthly and discounts three years into NPV.
- What resolution rate is realistic for chat and email tickets?
- It depends on how much of your volume is repetitive and documented. The preset assumes 45%, an assumption below many vendor headlines. Pull three months of tickets, tag the categories an agent could close from existing help articles, and use that share as a ceiling. Then measure the real rate in a pilot, counting only tickets that stay closed.
- How do I model per resolution pricing in this calculator?
- Enter the vendor fee multiplied by your expected resolution rate in the other cost per task field, because the calculator applies that field to every ticket. For example, a $0.99 fee at a 45% resolution rate averages about $0.45 per ticket. Then lower build cost to your onboarding and integration effort, since you are not building the agent yourself.
- Should I build or buy a customer support AI agent?
- Buying usually wins at lower volume, because outcome fees scale down and there is little to build. Building tends to win at high volume or when the agent must act inside internal systems a vendor cannot reach. Run the calculator twice with the same ticket assumptions, once with a vendor fee and low build cost and once with your own build, then compare 36 month NPV.
- Why is automation decay set above zero for support?
- Support knowledge goes stale. New features, pricing changes and policy updates make old help articles wrong, and the agent's resolution rate falls unless someone keeps the content current. The preset assumes a 5% decline per year as a placeholder. If you assign a knowledge owner, set decay to zero and add that person's time to the monthly maintenance figure instead.
- Is my ticket data sent anywhere?
- No. The calculator runs in your browser and only asks for aggregate numbers such as monthly volume and cost per ticket, never ticket content. The share button stores those numbers in the link after the hash sign, which browsers do not send to the server. There is no login, no form and no tracking of the values you enter.
Related services and reading
From a scenario to a scoped build.