TL;DR & Quick Summary
Decide on AI support automation with your own cost per ticket, not a vendor's percentage. The method:
- Cost per ticket today = loaded hourly agent cost ÷ tickets resolved per hour, per ticket type if possible
- Cost of an AI-resolved ticket = model usage per conversation + a share of running costs
- Automate first the tickets that are high-volume, low-risk and answerable from your systems
- Payback = build cost ÷ monthly net saving
In the worked example below, a team handling 5,000 tickets a month at $5 each, with AI resolving 30%, pays back a $40,000–$60,000 build in about 7–11 months.
- Key Takeaway: The business case is made ticket type by ticket type. An average containment rate across all tickets hides whether automation is working where it matters.
- Get Started: Want your own payback worked out? Schedule a strategy call with Cogniq AI or see our AI customer support services.
Step 1: Work Out Your Cost per Ticket
Cost per ticket is the fully loaded cost of resolving one support request. Two ways to calculate it:
- From totals: monthly support cost (salaries, benefits, software, management, outsourcing) ÷ tickets resolved that month
- From time: loaded hourly cost of an agent ÷ tickets an agent resolves per hour
For illustration: an agent costing $30 an hour fully loaded, resolving one ticket every 10 minutes (6 an hour), costs $5 per ticket.
Do it by ticket type if your help desk tags tickets. A password reset might take 3 minutes and a billing dispute 25. The average hides exactly the difference automation exploits.
Step 2: Work Out What an AI-Resolved Ticket Costs
Two parts:
- Model usage per conversation. Language models charge per token. A short support conversation typically costs a few cents to a few tens of cents, depending on length, how much documentation is sent with each turn, and the model. Our chatbot development cost guide shows how to calculate it. The worked example there comes to about 8 cents for a six-turn conversation at illustrative prices.
- A share of running costs: hosting, search infrastructure, monitoring, keeping help content current, and reviewing conversations the AI handled badly.
The build cost is separate. It's an investment recovered from the monthly savings.
Step 3: Decide What to Automate First
Score each ticket type on three things:
| Factor | Automate early | Keep human-led |
|---|---|---|
| Volume | Many tickets a month | Rare |
| Definition | Clear question, clear answer | Ambiguous, investigative |
| Risk | Low cost if wrong, easy to correct | Money, legal, safety or reputation at stake |
Typical good first candidates:
- Order, delivery and booking status: answered from your systems once the customer is identified
- Account access and password help: follows a fixed process
- Policy and how-to questions covered by help articles
- Simple changes within rules: rescheduling, address updates, cancellations inside a window
Typical ones to keep with people, at least at first:
- Complaints and escalations
- Refunds and credits above a threshold, or sent for human approval
- Technical faults that need investigation
- Anything legally sensitive
Status and account questions need the AI connected to your order, booking or account systems. That integration is what separates an AI that answers from one that resolves.
A Worked Payback Example
For illustration, a support team with these numbers:
- 5,000 tickets a month at $5 per ticket: $25,000 a month in handling cost
- 40% of tickets fall into automatable categories (status, account, policy)
- The AI fully resolves three-quarters of those, so 30% of all tickets: 1,500 a month
- AI running costs: about $400 a month in model usage at 8 cents per conversation, plus about $1,500 a month for hosting, monitoring, content upkeep and review
| Item | Monthly |
|---|---|
| Handling cost avoided (1,500 × $5) | $7,500 |
| AI running costs | −$1,900 |
| Net monthly saving | $5,600 |
With a build cost of $40,000–$60,000, typical for an account-aware support assistant, payback is about 7–11 months.
Halve the volume and payback roughly doubles. Raise it, or include faster responses, longer support hours or retained customers, and payback shortens. Test the case with your own numbers before trusting anyone else's.
How to Find Your Automatable Share
The most important number in the business case is the share of tickets that AI can genuinely resolve. Don't guess it. Measure it with a ticket audit:
- Export three months of tickets with their tags, channel, first message and resolution notes
- Sample 200–300 tickets across all categories. Reading them takes a day and is the most valuable day of the project.
- Mark each one: could it be resolved with information from our help content or systems, without judgement or exceptions?
- Note what each automatable ticket needs: just help content, or a lookup in the order, booking or account system
- Calculate the share by category, weighted by volume
The audit almost always reveals two things. Some high-volume categories are more automatable than expected, usually status and account questions. And some tags are so broad that they mix easy and hard requests, which is worth fixing in the help desk regardless of AI.
A 90-Day Rollout
- Days 1–30: ticket audit, scope the first three categories, connect read-only to the systems they need, build a test set from real tickets
- Days 31–60: build and test; launch to a small share of customers on one channel, with easy escalation to a person
- Days 61–90: expand to all customers on that channel; review failed conversations weekly; add the next category once resolution is steady
Measure cost per resolved ticket from the first live day. By day 90 you'll have real numbers to replace the assumptions in the payback calculation.
Make the Savings Real
Freed agent time only becomes money if it's used. Decide in advance which of these you'll do:
- Absorb growth without hiring
- Reduce overtime or outsourced overflow
- Extend support hours, covering nights and weekends with AI and escalating to people in the morning
- Move agents to higher-value work: complex cases, retention, upselling
If none of these happens, the time disappears into slack and the business case never shows up in the accounts.
Measuring After Launch
- Resolution rate by ticket type, not just overall
- Escalation quality: do handoffs arrive with full context, or do agents start from scratch?
- Customer satisfaction on AI-handled versus human-handled tickets
- Repeat contacts: the same customer asking again means the first answer failed
- Cost per resolved ticket, AI versus human, monthly
Review failed and escalated conversations weekly at first. Most improvements come from fixing help content and adding missing integrations, not from changing the model.
Beyond Cost: Value the Payback Leaves Out
The payback calculation only counts handling cost. Three other effects are often worth as much, and are worth estimating separately so they don't inflate the core case:
- Faster answers. Customers asking about an order at 11pm get an answer at 11pm, not the next morning. For businesses where slow replies lose sales, such as bookings, quotes and pre-purchase questions, that has a direct revenue value.
- Longer coverage without night shifts. AI can handle routine requests around the clock and leave a clean handoff for the morning team.
- Better agent focus. With routine tickets handled, agents spend more time on complex cases, retention conversations and the customers most at risk of leaving.
Keep these as upside rather than part of the payback figure. If the case works on handling cost alone, these make it stronger. If it only works with them included, test them in the pilot before relying on them.
Build or Buy?
Support platforms increasingly include AI features, and for standard FAQ deflection they may be enough. A custom build earns its cost when the AI must act in your own systems, follow your specific rules, run across channels, or operate at a volume where per-resolution platform fees add up. The three-year comparison in custom AI agent vs SaaS applies directly. For channel-specific setups, see WhatsApp AI customer support.
Common Mistakes
- Using an average cost per ticket, which overstates the saving on easy tickets and understates it on hard ones
- Counting deflection as resolution. A customer who gives up isn't a resolved ticket.
- Automating complaints first because they're painful, rather than starting with high-volume, low-risk tickets
- No connection to your systems, so the AI can only repeat help articles
- No plan for freed time, so the saving never materialises
Conclusion
AI support automation pays off when it resolves high-volume, low-risk tickets that are answerable from your systems, at a cost per ticket well below your agents'. Work out your cost per ticket by type, estimate AI cost per conversation, start with the right ticket types, and calculate payback with your own volume. In our example, 5,000 tickets a month paid back the build in under a year.
Schedule a strategy call with Cogniq AI to work out your cost per ticket and payback, or explore our AI customer support services.
