How to Calculate Cost per Chatbot Resolution
Cost per resolution is the number that shows whether chatbot automation actually saves money. This how-to defines a real resolution, tallies every cost including failures, and shows how to compare bots to humans honestly.
Table of contents
Plenty of teams can tell you how many conversations their chatbot handled last month. Far fewer can tell you what each resolved conversation actually cost. That gap matters, because cost per resolution is the single number that tells you whether automation is saving money or just shifting it around. It is also the figure that lets you compare a chatbot honestly against the human support it is meant to relieve. This guide walks through how to calculate it without fooling yourself.
Define a resolution before you count one
The most common mistake is measuring cost against the wrong denominator. You want cost per resolved conversation, not cost per conversation started. A resolution is an interaction where the customer's problem was actually solved by the bot — not abandoned, not silently dropped, and not escalated to a human. Decide your definition explicitly: does a conversation count as resolved if the customer stops replying, or only if they confirm the answer helped or a follow-up survey comes back positive? A loose definition flatters your numbers; a strict one tells the truth.
Add up every cost, not just the platform fee
The denominator is resolutions; the numerator is total cost. Most teams remember the platform subscription and forget everything else. A complete picture includes the items below.
| Cost component | What it covers |
|---|---|
| Platform fee | The chatbot vendor's monthly or seat-based subscription |
| Model / usage cost | Token or message charges for the underlying AI |
| Human review | Staff time spent monitoring, correcting, and approving bot answers |
| Failed conversations | The model spend on chats that ended without resolving anything |
| Escalation cost | The human-agent time spent finishing handed-off conversations |
| Maintenance | Time spent updating the knowledge base, prompts, and integrations |
The two most frequently ignored lines are failed conversations and human review. You pay for every chat the bot attempts, including the ones it gets wrong, and someone has to watch the output during the early months. Leave those out and your cost per resolution will look unrealistically good.
Do the calculation
The formula is straightforward once the inputs are honest:
| Step | Example basis |
|---|---|
| 1. Sum all monthly costs | Platform + model + review + maintenance |
| 2. Count resolved conversations | Only those meeting your resolution definition |
| 3. Divide total cost by resolutions | This is your cost per resolution |
Run the same calculation for your human channel — total agent cost divided by tickets resolved — and you finally have a like-for-like comparison. Only then can you say whether the bot is cheaper, and by how much.
Watch for the metrics that lie
A few traps distort the number. Counting escalations as resolutions double-charges: you pay the bot to fail and the human to finish, yet credit the bot with the win. Ignoring partial automation is the opposite error — if the bot collects the customer's details and context before handing off, it still saved the agent time, and that saving belongs in the analysis. And beware of a falling cost per resolution that comes from a looser resolution definition rather than genuine improvement.
Use the number to improve, not just to report
Once you trust the figure, segment it. Cost per resolution will vary widely by topic: simple order-status questions are cheap to automate, while nuanced billing disputes may cost more to attempt than to send straight to a person. That breakdown tells you where automation pays off and where it is burning model spend on conversations it was never going to win. Route accordingly, and your overall cost per resolution falls without any drop in service quality.
Bottom line
Cost per resolution is only useful if both halves of the fraction are honest — a strict definition of "resolved" and a complete tally of every cost the bot incurs, including its failures. Calculate it the same way for humans and bots, track it by topic, and you will know exactly where automation earns its place and where it is quietly losing money.
