Updated
How do you reduce support ticket volume with AI?
Put the answers to your repeat questions in sources an AI agent can read, let it resolve those intents end to end, and hand everything else to a person with context. Then review weekly: every gap the agent reports is a source you write once instead of a reply you type forever.
This playbook does that in three parts, sized for a team of one to twenty. Launch sets the goal and the baseline. Evaluate tests the agent on your real tickets before a customer sees it. Scale turns the hours it gives back into work your team never had time for.
Part 1
Launch: build the case before the bot.
Is an AI agent worth it for a small team?
It is worth it when a large share of your tickets are questions you have answered before. Pull the last 90 days of tickets and tag each one with an intent. If a handful of intents (password resets, order status, plan questions, how-tos) make up a big slice of the queue, an AI agent can close them. If most tickets need judgment, start with drafts, not resolutions.
Set one goal, not five.
Pick the goal you will judge the agent on: fewer tickets reaching a person, faster first replies, or coverage after hours. Write it as a number you will check at 30, 60, and 90 days. Everything else is a side effect.
Record your baseline metrics.
You cannot prove a reduction without a before. Capture these first:
| Metric | Why it matters |
|---|---|
| Tickets per month | Volume you are trying to reduce |
| Top 20 intents and their share | What the agent should own first |
| First response time | How long customers wait today |
| Time to resolution | How long a ticket stays open |
| Minutes per ticket | The hours an AI resolution gives back |
| Reopen rate | Catches answers that looked resolved but were not |
Run your numbers.
Plug in your volume and the share you expect the agent to resolve. Minutes per ticket is your own average handle time. The estimate is yours, not a promise from us.
Under the hood
How does an AI agent process a support ticket?
Here is the path a ticket takes through Kal, the AI agent in SupportWire. Most good agents follow the same five steps. Ask any vendor to show you each one.
Intake
A message lands from live chat, email, or a form. It joins one shared inbox with the customer, their plan, and their past threads attached.
Intent
Kal reads what the customer actually wants, plus sentiment and urgency. A refund request, a bug, and an angry churn risk take different paths.
Answer from the Knowledge Store, with citations
Kal answers only from the sources in your Knowledge Store (help articles, PDFs, Notion, Drive, past replies) and cites them, so you can check every claim.
Handover to a human, with context
When the rules say a person should take it, Handover passes the thread with a written summary on top: what was asked, what Kal tried, which sources it read.
Learn from the correction
When a teammate edits or overrides a reply, Kal holds the correction. Missing answers become Knowledge Store gaps you fill once instead of answering forever.
The Knowledge Store is a repository of input sources Kal reads from: URLs, PDFs, screenshots, Notion, Drive, Confluence, and exports from tools like Zendesk or Help Scout. It is not a separate knowledge base you have to write. If you are switching tools, Wire Through moves you mid-conversation so customers keep their thread.
Part 2
Evaluate: test on your tickets, not the demo.
What should you look for in a customer support AI agent?
Score every agent you try on the same eight criteria.
| Criterion | How to check | What good looks like |
|---|---|---|
| Answer accuracy | Run your test set. Count correct, partly correct, wrong. | Wrong answers, not unanswered ones, are the failure that costs you. |
| Citations | Every answer links the source it used. | You can audit a reply in seconds instead of guessing. |
| Knows when to stop | Ask questions with no answer in your sources. | It should say so and hand over, not invent. |
| Handover quality | Read what the human receives. | A summary and the sources tried, so nobody re-asks the customer. |
| Setup time | Time from signup to first real answer. | A small team cannot spend a quarter on onboarding. |
| Pricing model | Seat price plus what you pay per AI outcome. | Pay on resolution, not on every conversation the bot touches. |
| Switching cost | Can you move without dropping live threads? | Wire Through keeps conversations intact mid-switch. |
| Data handling | Is your data used to train public models? | Ask for it in writing. Kal never trains public models on your data. |
AI chatbot or AI agent?
Most "AI chatbot for customer service" searches end up here. The difference shows on the first question the script did not plan for.
| Trait | Rule-based chatbot | AI agent |
|---|---|---|
| How it answers | Scripted flows and keyword matches | Reads intent and writes an answer from your sources |
| Unknown question | Falls back to a menu or "I did not understand" | Says it does not know and hands over with context |
| Upkeep | Every new flow is built by hand | Add a source to the Knowledge Store |
| What it can close | Simple FAQs inside its script | Repeat questions end to end, plus drafts for the rest |
Build a test set from real tickets.
Copy 30 to 50 real customer messages into a sheet. Cover your top intents, a few billing and angry messages, and at least five questions your sources cannot answer. Write the expected outcome next to each: resolve, ask a clarifying question, or hand over. Our test set template has the columns ready.
How to run a trial.
- Connect your help content to the Knowledge Store. Leave out anything outdated.
- Run the whole test set and score each answer pass, partial, or fail.
- Fix the source behind every fail, then run the set again.
- Write your handover rules before go-live, not after the first bad reply.
- Go live on two or three intents first. Add more as each one holds up.
SupportWire has a 14-day trial with no card, long enough to run this whole loop on real tickets.
Pricing
What do 1,000 AI resolutions cost?
On SupportWire, 1,000 AI resolutions a month cost the same at any team size. Only the seats change. You pay when Kal closes a ticket, never for a handover.
| Team | Seats | AI resolutions | Total per month | Per resolution, all in |
|---|---|---|---|---|
| 1 seat | $19 | $465.50 | $484.50 | $0.48 |
| 3 seats | $57 | $465.50 | $522.50 | $0.52 |
| 5 seats | $95 | $465.50 | $560.50 | $0.56 |
| 10 seats | $190 | $465.50 | $655.50 | $0.66 |
$19 per seat per month, first 50 resolutions included, then $0.49 each. Lite seats are free. Full details on pricing. Comparing against Intercom Fin? Use the Intercom pricing calculator.
Part 3
Scale: roles, milestones, and the hours you get back.
Who runs an AI agent on a small team?
Enterprise playbooks hire new roles for this. A small team needs three hats, not three hires.
- Owner (about 2 hours a week)
- One person owns the agent: reads the weekly review, decides what Kal handles next, and signs off on handover rules. Usually the founder or support lead.
- Knowledge keeper (about 1 hour a week)
- Fills the gaps Kal reports. Every unanswered question becomes a source added to the Knowledge Store, written once.
- Everyone on the inbox
- Takes handovers, corrects replies in place, and flags anything that felt off. Corrections are the training loop, not a side job.
Hold one 30-minute review a week: read every wrong answer, every handover Kal should have resolved, and the list of gaps. Fix sources, adjust rules, move on.
What should you measure at 30, 60, and 90 days?
| Metric | Day 30 | Day 60 | Day 90 |
|---|---|---|---|
| AI resolution rate | Measured on 2 to 3 intents | Rising as sources fill gaps | Covers your top 20 intents |
| Handover rate | High on purpose | Falling on owned intents | Stable, by rule not by accident |
| Wrong answers | Reviewed one by one | Rare, each one fixed at the source | Near zero on owned intents |
| First response time | Baseline vs now | Seconds on AI-owned intents | Seconds for every first reply |
| Reopen rate | Baseline vs now | No higher than baseline | No higher than baseline |
| Hours given back | Counted, not spent yet | Assigned to one project | Part of the weekly plan |
Targets are directional. Your baseline sets the bar, not an industry average.
Where do the freed hours go?
If you do not assign the hours, the queue absorbs them. Name where they go before day 60. Good homes on a small team:
- Writing the sources Kal reports as missing.
- Proactive outreach to customers who are stuck in onboarding.
- Turning repeat bug reports into fixed bugs with engineering.
- Faster, deeper replies on the conversations that need a person.
Templates
Two templates. No email required.
- Handover rules template (CSV)
- Ten common triggers, from refunds to churn signals, with who takes over and the context they need. Edit the routes to match your team.
- AI agent test set template (CSV)
- Twelve sample tickets across intents, with the expected outcome and a column to score each run. Replace them with your own.
Method
How we wrote this, and why you can trust it.
We build SupportWire and Kal, so we have a stake. Here is what we did to keep this useful anyway:
- We read the public launch and scale guides other vendors publish, then rewrote the steps for a team of one to twenty instead of a support org with an AI operations department.
- There are no survey statistics on this page. We did not find numbers we could verify at the source, so we left them out. Your baseline is the only benchmark that counts.
- Every price comes from the same constants that power our pricing page, so this page cannot drift from what you are billed.
- The evaluation criteria apply to any vendor. Run them on us and on anyone else.
FAQ
AI support agents, answered.
Updated November 2026