Most chatbot ROI claims are marketing arithmetic: multiply deflected tickets by a made-up cost per ticket, announce a huge number. If you're going to spend real money on a support chatbot, you deserve a calculation you can defend to your CFO.
Here's the framework we use to scope projects honestly — including the costs vendors don't mention and the metrics that actually predict whether the bot is working.
Start with your real cost per ticket
Cost per ticket = fully loaded support cost (salaries, tools, management overhead) divided by tickets resolved per period. For most SMB teams this lands between $3 and $15 per ticket; for technical products it can be far higher.
Compute your own number. It's the foundation of the whole calculation, and using an industry average instead of your actual figure is how ROI theatre starts.
Deflection: the number everything hinges on
Deflection rate = conversations fully resolved by the bot with no human touch, as a fraction of conversations that would otherwise have become tickets. Honest deflection excludes abandoned chats and "resolved" conversations where the user gave up.
Realistic ranges: 40–60% for a well-built RAG bot over good documentation; 60–80% when the question mix is heavily repetitive. Any vendor promising 90% on day one is quoting the dishonest version of the metric.
The quality check: CSAT on bot-resolved conversations should be within a few points of human-resolved CSAT. Deflection that tanks satisfaction isn't savings — it's churn risk wearing a cost-savings costume.
The costs on the other side of the ledger
Build cost (one-time), model/API usage (scales with volume, usually modest), hosting, and — the one everyone forgets — knowledge maintenance. Someone has to keep the docs the bot answers from current. Budget a few hours a month, or watch answer quality decay.
Include an evaluation pass after launch: measuring answer accuracy on a real question sample. It's how you catch drift before customers do.
Second-order returns people forget
24/7 coverage without night shifts. Faster first-response for the tickets humans do handle, because queues are shorter. Lower agent churn when the job stops being repetitive. Every bot conversation is also structured data about what confuses your customers — free product research.
These are harder to put numbers on, but in mature deployments they often matter as much as the deflection line.
The metrics to track weekly
Five numbers: honest deflection rate, CSAT on bot conversations, escalation quality (did the human get context, or did the customer repeat themselves?), answer accuracy on a sampled set, and cost per bot-resolved conversation.
If you only track one: deflection × CSAT. Either number alone is gameable; together they tell the truth.
A worked example
A team handling 2,000 tickets/month at $6 per ticket spends $12,000/month on resolution. A bot honestly deflecting 50% saves ~$6,000/month before its own costs — call it ~$5,000/month net after API, hosting, and maintenance.
Against a mid-range build cost, that's payback inside 3–6 months, and everything after is margin. Run the same math with your numbers — it either clears the bar convincingly or it doesn't, and both answers are useful.
Frequently asked questions
Conclusion
Support chatbot ROI is real, but it's earned with honest metrics — deflection paired with CSAT, costs counted fully, maintenance budgeted.
If you want the worked example run on your actual ticket volume and costs before committing to anything, that's exactly what our discovery call is for.