How much can an AI chatbot actually cut your support costs?

Annual support saving = (monthly tickets × deflection rate × cost per ticket × 12) − annual tool cost.

That is the whole model. Four inputs, one subtraction. The reason ROI discussions get muddy is that vendors tend to supply the flattering terms (a high deflection rate) and leave you to supply the rest.

Work through the terms in the order below. Three of them you can establish today from data you already have. The fourth — deflection — you should refuse to estimate, and measure instead.

Cost per ticket = total monthly support cost ÷ tickets resolved that month — including salary, employer taxes, benefits, tooling and management overhead, not just base pay.

Published industry benchmarks for cost per ticket vary enormously — by channel, region, complexity and how the study defined a "ticket". Borrowing someone else’s average is the fastest way to build an ROI case that collapses under scrutiny. Your own number takes about ten minutes to produce and is defensible.

1. 1

   Add up the true monthly cost of support

   Fully-loaded salaries for everyone answering tickets (base pay plus employer taxes and benefits — typically well above the headline salary), plus helpdesk licences, plus a fair share of the manager’s time.

2. 2

   Count tickets actually resolved in the same month

   Use resolved rather than received, so the two halves of the fraction describe the same work.

3. 3

   Divide

   That is your blended cost per ticket. Do it separately per channel if email and phone differ sharply — phone almost always costs several times what email does.

This is the step almost every vendor ROI page skips, and it is the one that decides whether the whole business case is real. A chatbot grounded in your website content can only answer questions your website content answers. That excludes a large, and often surprising, share of your ticket volume.

The pattern: repetitive, informational, already-public questions are deflectable. Anything requiring account access, an action on your systems, human judgement or de-escalation is not — and should not be. A chatbot that tries to handle a refund request is a liability, not a saving.

Pull your last 200 tickets and tag each one against those two columns. The resulting percentage is your deflectable ceiling — the absolute maximum deflection you could ever achieve. Your real deflection rate will land somewhere below it, because not every visitor with a deflectable question will find or use the chatbot.

Deflection rate is the share of would-be support contacts that the chatbot resolved, and it can only be established from your own post-launch data — never from a vendor benchmark.

The honest way to establish it is a before-and-after comparison on ticket volume, controlled for traffic. Take your monthly tickets per thousand sessions for the quarter before launch, then the same ratio for the quarter after. The drop is your deflection, expressed in the only currency that matters: tickets that stopped arriving.

While you wait for that quarter of data, your chatbot’s own analytics give you a leading indicator. In SiteMind, the analytics dashboard reports the numbers this calculation needs directly:

• Total questions and total answers — the volume actually flowing through the assistant.
• Answer rate — the share of questions it answered rather than declined. A low rate means your content has gaps, not that the tool is broken.
• Unanswered questions, listed individually — the single most valuable list in the product. Every entry is a real customer question your website does not currently answer.
• Top questions — the repetitive queries that dominate your queue, ranked.
• Conversations and unique visitors — the denominator for your per-session ratios.

The unanswered-questions list is where deflection improves over time. Each entry is either a page you should write or an answer you should edit into the knowledge base directly. Do that regularly and answer rate climbs, which pushes real deflection toward your ceiling.

Suppose a business handles 900 support tickets a month, has worked out a fully-loaded cost of $6 per ticket, and tagged 55% of a recent ticket sample as informational and already covered by its website. It launches an assistant and, after a quarter, measures an actual 30% drop in tickets per thousand sessions.

Two things are worth noticing. First, the tool cost is close to a rounding error against the labour cost — which is generally true, and is why the deflection rate, not the subscription price, is what you should scrutinise. Second, the measured 30% sits well below the 55% ceiling, which is realistic; assuming you will hit your ceiling is the second most common way these business cases get inflated.

ROI maths needs a real number for the subscription, and it needs to be the tier your volume genuinely requires — not the cheapest one on the page. SiteMind’s pricing is public and metered on AI responses per month:

Size the plan against expected chat volume, not deflected tickets — every question asked consumes a response, including the ones that would never have become tickets. If your site gets meaningful traffic, assume chat volume comfortably exceeds your current ticket volume.

1. Counting every answered question as a deflected ticket. Most chat questions were never going to be emails. Only a before-and-after ticket comparison tells you the truth.
2. Using base salary instead of fully-loaded cost. This one understates savings, so it is the rare error that works against you — but it still makes the model wrong.
3. Assuming you hit your deflectable ceiling. Real deflection lands below it, because discovery and willingness-to-use are never 100%.
4. Ignoring the maintenance cost. Reviewing unanswered questions and keeping content current is a real, recurring half-hour a week. Small, but not zero.
5. Forgetting that deflected tickets are the cheap ones. The easy, repetitive questions leave the queue first. Your remaining tickets get harder on average, so cost per ticket rises even as volume falls. Model the saving on volume, and expect your average to drift up.

The honest version of this business case takes one quarter to prove and costs very little to test. Establish your cost per ticket and deflectable ceiling this week from data you already have. Deploy against your real content. Then measure tickets per thousand sessions before and after, and let that number — not a vendor’s — decide whether it worked.

SiteMind exists to make the deflectable share genuinely answerable: it reads your existing website, answers only from what it finds there with clickable citations, and declines honestly when your content does not cover something — which is what turns the unanswered-questions list into a reliable content to-do list rather than a pile of confident guesses.

You can start a free 3-day trial with no card required, or watch it answer live on this site before you model anything.