When to Hire a Support Agent vs AI: What Your Tickets Say

Created time

Sep 29, 2026 09:12 AM

Title length (<60)

Author

Mike Heap

Last optimised

Ecomm?

Why do ticket ratios and salary comparisons both point to the wrong hire?

⚡

TL;DR: A tickets-per-agent trigger and a salary-versus-software sum are the two usual answers. Both skip the reason a ticket needed a person, and that reason decides whether a hire fixes it.

The ratio answer has a long history in support. A support metrics study drawn from one helpdesk vendor's own customers found an average of 21 tickets per agent per day (in an older dataset that assumed mostly manual work), and its author suggests thinking about a new hire at around 30. The number gives a quick read on how busy your agents are.

The salary answer is newer. US government labor figures put the median customer service rep's pay at $21.53 an hour, and next to that an AI agent's cost per ticket looks tiny.

To see where the ratio goes wrong, picture a team at 35 tickets per agent per day that starts hiring while half of what reaches its agents is either a question the help center doesn't answer yet or an order-status check the AI can't see.

A two-column graphic titled What the ratio and salary numbers miss. On the numbers alone: a team at 35 tickets per agent a day, past the rough 30 where one study suggests thinking about a hire, and $21.53 an hour median pay for a support rep against an AI cost per ticket that looks tiny by comparison. Sorted by reason: half are a question the help center does not answer yet or an order status the AI cannot see, fixed by new articles and one data connection, and a complaint about money still needs a person with the authority to make an exception.

The new hire spends their day on work that a few new articles and one data connection would have removed. We saw how big that effect can be at Edel Optics, where resolution went from 25% to 79% once the AI could see live customer data.

The salary comparison assumes any ticket can go to the AI if the price is right, so it never flags the tickets where a person is the right answer. A complaint about money is the obvious example (the customer needs someone with the authority to make an exception). On the repeat work, the salary comparison holds up. One practitioner on Reddit found most of their support cost there:

"The expensive stuff was never the hard questions. It was the same easy one arriving forty times a month."

Start with the tickets your AI already hands to your team. We ask each of those tickets one question (why did this need a person?), because each answer points to a different fix.

Our Insights dashboard groups every conversation by topic and shows the handover rate for each, so you can read in a few minutes which topics your AI hands over most.

What decides whether the next hire is a person or more AI?

⚡

TL;DR: A ticket reaches a person because the AI doesn't have the answer, because it isn't able to take the action, or because you've decided a person should handle it. The third is the hiring signal, once that group is growing.

Contact centers have long reviewed transferred calls to find out why each transfer was necessary, and trade guidance on transfer scorecards is to make sure transfers happen for the right reasons. We apply the same habit to the tickets an AI hands over, to find out whether the fix is a person or more AI coverage. One practitioner on Reddit put the test in one line:

"I draw the line at where the answer lives, not how hard the question sounds. A doc or a record the agent can read, I automate. Money or someone's day, I hand it to a person."

Here's how each of the three reasons shows up in your tickets, and what fixes it.

Reason a ticket reaches a person	What it looks like in your tickets	What fixes it	Is it a hire?
The AI doesn't have the answer yet	A policy only one agent knows, a missing help article, an order status the AI can't see	Write the answer down, or connect the customer data	No
The AI isn't able to take the action yet	Refunds, cancellations, address changes, payment checks	Some developer time, done once and then used for every customer	No
You've decided a person should handle it	Exceptions to policy, disputes about money, high-value accounts, regulated advice	A person with the authority to decide	Yes, when this group keeps growing and the team is stretched on it alone

A fourth check applies to all three: whether anyone is improving the AI at all. If nobody is, the hire you need may be someone to own the AI.

The AI doesn't have the answer yet

This is usually the biggest group, and a ticket ratio counts every one of these tickets as a reason to hire (it can't tell a missing article from a busy agent). The customer asks something reasonable, the answer exists in someone's head or a system the AI can't reach, and the ticket goes to a person.

It comes in two forms:

Knowledge - a returns rule that's only in one agent's memory, a shipping exception nobody wrote up, a product question the help center skips.

Customer data - an order's status, a subscription's renewal date, the plan an account is on, none of which the AI can answer unless it can look them up.

We usually see knowledge alone get teams to somewhere between 40% and 70% resolution. Connecting customer data then resolves roughly another 15 to 50 tickets in every 100, with ecommerce at the high end because so many tickets are about orders.

Your current team can handle both steps. Writing the answer down should take a few hours if AI drafts it from your agents' past replies, and connecting order or account data is a one-off job for whoever looks after your store or product. For teams with no help center or written docs, we train your AI on your historic tickets and draft starter articles from the last 5,000 of them by default for your team to review.

The AI isn't able to take the action yet

Here the customer needs something done (a refund, a cancellation, a new delivery address) and the AI hasn't been given access to do it, so the customer waits for an agent to click the same button they've clicked all week. That means a slower reply to a simple request, and hours of your agents' time on steps a system could do.

The fix is some developer time so the AI can take the action, either by itself or by proposing it for an agent to approve (you choose per action). Actions often resolve another 5 to 20 tickets in every 100, and once built, an action keeps working when people leave the team. Take it to your engineering lead as a question of priorities.

You've decided a person should handle it

This is where hires go. It covers the tickets your business has chosen to keep with people today: exceptions outside policy, disputes and complaints where money or trust is at stake, high-value accounts you can't afford to lose, and advice you're regulated on.

I treat it as today's choice, and I expect it to shrink over time. As the AI gets more access and you see how it handles edge cases, some of these tickets move over, the same way order lookups did.

The hiring signal is this group growing month on month while your team is already stretched on these tickets alone. A new agent hired while most handed-over tickets are missing answers or missing actions spends most of their week on work the AI should be doing. If most are tickets you've chosen to keep with people and the count keeps rising, you've found the hire. Our AI vs human support guide goes into which ticket types teams keep with people today.

When is the right hire the person who owns the AI?

The fourth check is whether missing answers and missing actions are shrinking at all. If resolution hasn't moved for weeks and nobody reviews the tickets the AI couldn't answer, the role you're missing is someone who owns the AI. We ask every team we work with to name that person.

Anyone on the support team can do it with a rough idea of how the AI works and some protected time (usually half an hour a week, up to an hour if they're very involved) to review what came through and fix the gaps. That time usually comes out of an existing role, and in a larger team it can be the next hire.

A yellow card headed "Want to make these AI replies even better?" with four linked actions: inspect this conversation to see what knowledge was used, add guidance, create custom answers for questions the AI couldn't answer, and connect your internal systems for live customer data.

Our Self-Learning feature drafts new help articles by comparing the AI's reply with what your agent sent on each handed-over ticket, so missing answers get written up without anyone starting from a blank page.

What does this look like in real rollouts?

⚡

TL;DR: Edel Optics fixed a missing-data problem by connecting live customer data, an iGaming operator let the AI handle payment checks, and a live-music events company keeps most tickets with people by choice. At YouGarden, AI covers the work of about six full-time agents.

We've seen four customers make different staffing calls, each driven by one of the three reasons.

Rollout	Which reason	What they did	Result
Edel Optics	The AI didn't have the answer	Connected live customer data	Resolution from 25% to 79%
An iGaming operator	The AI couldn't take the action	Let the AI run a deposit task (covering appeals and receipt checks) and a withdrawal task	Those two tasks handle 55% of incoming tickets
A live-music events company	They decided a person should handle it	Set the AI to hand most tickets over with context	About 555 of 750 monthly tickets go to a person by design
YouGarden	They decided a person should handle follow-ups	AI answers the first message, people take follow-ups	965 hours saved a month

Edel Optics: resolution from 25% to 79% after connecting customer data

Edel Optics handles about 4,000 tickets a month, and their jump from 25% to 79% AI resolution came from one change, plugging in live customer data. Before that, questions about orders, deliveries and returns reached a person because the AI had nothing to look up.

Hiring would only have added another person to look those orders up by hand. Their AI now saves the team 150 hours a month at 92% CSAT. They still keep some tickets with people on purpose (anything to do with faulty items may need a person).

An iGaming operator: two payments tasks handle 55% of tickets

An iGaming operator had a large share of tickets about deposits and withdrawals, and answering each one meant working in the payments system. Our agent now handles them itself with two payments tasks, one for deposits that also deals with a player's appeal or receipt check, and one that tells a player where their withdrawal stands. Each task asks the player for what it needs, looks the transaction up and confirms the result back to them. Those two tasks now handle 55% of incoming tickets.

The AI now resolves 44% of conversations at 72% AI CSAT and saves the team about 170 hours a month. About a third of conversations still reach a person. I'd check that split before opening any new role, because it shows what a new hire would spend their week on.

A live-music events company: most tickets go to a person by design

A live-music events company handles about 750 tickets a month, and the AI resolves about 195 directly. The other 555 or so go to a person with context the AI has already gathered, because the business chose that split.

The AI's replies score 85% AI CSAT, and the people who pick up the rest start with the customer's details already in front of them. If the tickets they keep with people grew faster than the team, that would be the right place to hire.

YouGarden: 965 hours a month, about six full-time agents

When people ask me about headcount, I point them to YouGarden. Their AI resolves 66% of tickets and saves 965 hours a month, which is the work of about six full-time agents.

Video preview — I Let AI Agents Resolve 10,000 Support Tickets, Here's How Much It Cost

They chose to let the AI reply only to the first message a customer sends. Their people handle every follow-up.

Each of these rollouts has a full write-up in our case studies, with the setup and numbers in more detail.

What should you do this week?

⚡

TL;DR: Take last month's tickets that reached a person, sort them by the three reasons, and hire for the part your team has chosen to keep and can't keep up with.

You can start with what's already in your helpdesk. Work through these five steps in order, since the first gives you the numbers for the rest.

Sort last month's handed-over tickets by reason - export them from your helpdesk and have AI group them into the three reasons, which takes about an hour. You finish with three numbers, one per reason.

Close the gaps where the AI didn't have the answer - draft the missing articles with AI from your agents' past replies, and list any customer data the AI couldn't see. Expect the articles to take a few hours, and those tickets should stop reaching people once the answers are live.

List the actions the AI couldn't take - pick the two or three actions behind most of those tickets, then book a half-hour with your engineering lead this week and leave with a date for the work.

Check whether the tickets you keep with people are growing - if that count is up on last month and the team is stretched on those tickets alone, write the job ad around judgment work. Our guides to building an AI-first support team and to support staffing models cover what kind of person to hire and how many.

Name who owns the AI - block half an hour a week for them to review what the AI couldn't answer. Watch whether the first two numbers fall month on month.

How do I get AI to sort my handed-over tickets by reason?

Here's a prompt that does step 1 for you. Paste it into ChatGPT or Claude along with last month's handed-over tickets (an export with each customer's messages and your team's reply is enough). We've kept it to the three reasons above, in the same words, so its numbers drop straight into steps 2 to 4. It can only judge what's written in each ticket, so check how it sorted a dozen of them before you trust the totals.

You are helping me decide whether my support team needs another hire or more AI coverage.

Below is an export of last month's support tickets that our AI agent handed over to a person. For each ticket, read the customer's messages and our team's reply, then put the ticket into exactly one of these three groups:

1. The AI didn't have the answer: the answer was missing from our help center, lived only in one person's head, or needed customer data the AI couldn't see (an order status, a renewal date, the plan an account is on).
2. The AI couldn't take the action: the customer needed something done, such as a refund, a cancellation, an address change or a payment check.
3. We decided a person should handle it: an exception to policy, a dispute or complaint about money, a high-value account, or advice we're regulated on.

If a ticket doesn't clearly fit one group, mark it "unclear" and say why. Don't guess.

Then give me:
- A table with each group, its ticket count and its share of all the tickets.
- For group 1, the missing answers that came up most often, and a separate list of the customer data the AI would have needed.
- For group 2, the two or three actions behind most of those tickets, most common first.
- For group 3, the most common ticket types.

[Optional: I've also pasted the previous month's export. Sort that the same way and tell me whether group 3 went up or down.]

Tickets:
[paste your export of last month's handed-over tickets]

Our AI Tagging labels each ticket's contact reason as it arrives in Zendesk, Intercom, Freshdesk, Freshchat or Gorgias, so you can start step 1 by filtering your helpdesk on those tags.

The My AskAI AI agent setup Tagging page, showing three conversation properties from Freshchat (Reason Code, Sentiment and Type) and a switch to pause AI tagging.

When does sorting tickets by reason not settle it?

⚡

TL;DR: Hire first when there's no ticket history and nothing written down, when the work is mostly phone calls or licensed advice, or when a team of two can't cover a holiday. Sorting by reason assumes there's something for the AI to work from.

The three reasons work when the AI has something to learn from and the work suits an AI agent. A few situations change the answer:

Nothing is written down and there's no ticket history - a brand-new business has nothing for the AI to learn from, so the first hire answers tickets and writes the answers down as they go. If you have past tickets, the AI can start from those even with no help center.

The work is mostly voice, licensed advice or being on site - high call volumes, regulated advice and in-person work are still jobs for people, and they belong in your hiring plan.

Your team is very small - the human work that's left after AI is harder and more concentrated, so a team of two may hire for cover and resilience even when the numbers say the AI could take more. I think that's a sensible call when one person's holiday leaves nobody on the hard tickets.

Your volume spikes for a few weeks a year - the AI can absorb a seasonal peak, and peak planning needs a separate playbook.

Outsourcing is a third option (it suits cover for hours or languages you can't staff). For a small business making its very first hire, general small-business guidance on first hires covers the basics that apply to any role.

Whoever you hire will spend more of their time on the tickets that need judgment. AI help for the agent pays off most on those tickets, as one Reddit commenter put it:

"Then the human can spend less time gathering information and more time making the right call."

Our AI Copilot Chrome Extension drafts replies for your agents inside your helpdesk, with no seat charges. It works from the same knowledge your AI agent uses, so a new hire starts from what the AI knows on day one.

The takeaway

⚡

TL;DR: Before you hire, find out why tickets reach a person. Fix the AI when it lacks an answer or permission, and hire when the work is the kind you've chosen to keep with people and there's more of it each month.

There are three reasons a ticket still reaches your team: the AI doesn't have the answer, it isn't able to take the action, or you've decided a person should handle it. In our rollouts, missing answers and missing actions get fixed by writing things down, connecting data and some developer time. The tickets you've chosen to keep with people are where hires go, once there are more of them each month and your team is stretched on them alone.

This week, sort last month's handed-over tickets by those three reasons. If most turn out to be missing answers or actions, our 30-day free trial lets you test the fix on your own tickets, with every feature unlocked and no card needed.

FAQs

Will AI replace customer service?

AI already takes most of the repeat work in teams that run it well, and the share keeps rising as the AI gets more knowledge, data and actions. People handle the work the business chooses to keep with them, which today means exceptions, disputes and your most valuable accounts. That group gets smaller over time as teams hand more to the AI.

Will AI replace customer service jobs?

US government labor projections expect customer service rep employment to shrink by 5% over the next decade. In our customers' rollouts, teams keep their people and stop hiring for the repeat work as they grow, and the human job gets more senior. One analyst forecast expects half of the companies that cut support staff because of AI to rehire people for similar work under different job titles.

Should I hire another support agent or add AI?

Sort the tickets that reached a person last month by why they needed one. If most were missing answers, missing data or actions the AI couldn't take, add AI coverage first, because a hire would spend their week on that work. If most were tickets you've chosen to keep with people and that count is growing, hire.

When does it make sense to hire a support agent instead of using AI?

Hire when the tickets your business keeps with people (exceptions, disputes, high-value accounts, regulated advice) are growing month on month and your team is stretched on them alone. Also hire first when there's nothing written down and no ticket history, when most of the work is phone calls or on site, or when a very small team needs cover for holidays and sick days.

We're growing fast and can't hire support agents fast enough. What should I use?

An AI agent trained on your help center can go live within hours, and almost every team we work with is answering customers directly within a month. As volume grows, the AI answers the extra repeat questions, so you only need more people when the tickets that need a person grow too. Our guide to scaling customer support without hiring goes through the full playbook.

We're spending too much on customer support agents. How can AI help?

Start by finding which handed-over tickets are missing answers, data or actions, since those are the ones AI can take off your team's plate. On My AskAI, a typical helpdesk chat ticket works out at about $0.12 on the Pro plan and about $0.10 on the Scale plan, with add-ons for extras such as actions and tagging. Before you pay anything, our 30-day free trial lets you prove it on your own tickets, with every feature unlocked, unlimited tickets and no card needed. Our guide to calculating AI customer service ROI shows how to put the saving into numbers for your finance team.

When to Hire a Support Agent vs AI (and the Signals That Say Which)

Why do ticket ratios and salary comparisons both point to the wrong hire?