Mike HeapMike is an experienced Product Manager who focuses on all the “non-development” areas of My AskAI, from finance and customer success to product design, copywriting, testing and more.
Table of Contents
- Why hiring agents to match ticket volume stops working
- The inverted support team
- Role 1. The AI agent, your new tier-1 front line
- Role 2. The senior human tier, fewer people doing harder work
- Role 3. The AI owner, the genuinely new role
- The capability that feeds it all, data and automation
- What an AI-first team looks like in real rollouts
- How to start building your AI-first team this week
- How do I map out my own AI-first team?
- When an AI-first team is the wrong fit
- The takeaway
- FAQs
Created time
Jun 4, 2026 02:53 PM
Title length (<60)
Author
Ecomm?
Tags
Image
Publish date
Jun 8, 2026
Slug
building-an-ai-first-support-team
Featured
Type
Article
Ready to Publish
Ready to Publish
Going AI-first rebuilds your support team around the AI: it takes the front line, your humans get more senior, and one new role owns it.
An AI-first support team is a different org chart, built around the AI from the start. The tier-1 headcount you'd have hired as you grew is the headcount you no longer need to add.
Nearly every guide on building a support team gives you the same answer, and I think it quietly misleads you. Hire agents in proportion to your ticket volume, train them, set up shift cover, then layer in seniors and team leads as you scale. That advice is sound for a team where a human touches every ticket, and it quietly assumes a human touches every ticket.
When an AI agent resolves the repetitive 60 to 80% of your tickets on day one, that assumption breaks, and the make-up of the team changes with it (this is the bit the guides all miss). The pyramid flips. You stop hiring tier-1 to match volume, and you start building around a front line that flexes with demand and gets better every week.
I'm Mike, co-founder of My AskAI, and I've spent the best part of three years on calls with support leaders watching exactly this play out. We run AI customer support for 200+ ecommerce and SaaS businesses, our agents have resolved over 1,000,000 tickets, and we sit at a 72%+ resolution rate on a rolling 30-day basis across that base. So this post is the org pattern we keep seeing: what the team looks like, which roles fade, which roles get more senior, and the one role nobody had three years ago.
Why hiring agents to match ticket volume stops working
TL;DR: The standard playbook staffs humans in proportion to ticket volume. Once an AI agent clears most of that volume on day one, the math over-hires for work that no longer reaches a person.
The standard playbook for building a support team is genuinely good, and worth stating fairly before I take it apart. Help Scout's guide gives a seven-step version: define what great service means, pick your channels, hire for empathy, measure the right things, choose your tools, build a knowledge base, and wire support into the rest of the company. Zendesk's org-structure guide lists eight ways to arrange the people, from tiered to functional to geographic, with junior agents on the front line and complex issues escalating up.
They all share one blind spot, in our view: the team is entirely human. AI shows up, if at all, as a product in the nav bar, never as part of the org chart itself. The model underneath all of them is the same, more tickets means more agents, so you hire ahead of growth.
That model breaks the moment an AI agent takes the front line. If the AI handles the repetitive tail on day one, the tier-1 agents you hired for that tail are now sitting on work that never reaches them. You end up paying twice: once for the AI, and once for the humans it just made redundant at that tier.
The teams we work with hit this fast. Customer.io put it plainly:
"The faster we grew, the more we spent. We realized that it was no longer possible to solve problems by throwing more people at it. We needed to explore ways to achieve more operational efficiency while still maintaining a high standard of service."
Freecash reached the same point before they came to us:
"We knew we couldn't keep up with the rising demand by just adding more team members."
So the useful question isn't "how many agents per thousand tickets." It's a structural one: what does a team look like when the AI is the front line rather than a tool bolted onto the side of it?
There's a darker version of this question in the headlines. In September 2025, CNBC reported that Salesforce had cut its support headcount from around 9,000 to about 5,000, with Marc Benioff saying he needed "less heads" now that AI handles roughly half of all support interactions.
That's the layoff story, and at a company that size it's real. It's also not the pattern we see at most of the teams we work with.
The inverted support team
TL;DR: An AI-first team flips the support pyramid. The AI becomes the tier-1 front line, the human team gets smaller and more senior, and one new role owns the AI's improvement.
The framework below is the one thing worth taking away from this post. In a traditional support org the headcount pyramid is wide at the bottom: a big base of tier-1 agents handling volume, narrowing to a few seniors and specialists at the top.
An AI-first team turns it upside down, and in our experience that's the real unlock. The AI becomes the wide base, the human team becomes the narrow, senior, high-judgment top, and one new role appears on the side to keep the whole thing improving.
That gives you three roles plus one capability: the AI agent as your front line, a smaller and more senior human tier above it, an owner for the AI, and a data-and-automation capability feeding all of it. (We'll walk through each.)
Layer | Traditional support team | AI-first support team |
Front line (tier-1) | Agents hired in proportion to ticket volume | The AI agent, flexing with demand and seasonal peaks |
Above the front line | A few seniors and specialists | A smaller, more senior human team on complex and emotional work |
Owns the system | Team leads and managers | An AI owner running the weekly improvement loop |
Builds capability | Rarely a support-team function | A slice of developer time for data connections and automations |
Role 1. The AI agent, your new tier-1 front line
The AI handles the repetitive questions that used to fill the tier-1 queue: where's my order, how do I reset my password, what plan am I on, how do I get a refund. It works nights and weekends, answers in 95 languages, and never sits in a queue. The thing that makes this a structural change rather than a productivity tweak is that the AI flexes with demand.
This is the part that replaces peak-season hiring. As we've seen it, companies don't tend to get rid of staff, they just don't need to hire as many as they scale or for peak periods, because the agent flexes with scale and demand. An ecommerce team that used to staff up for Black Friday can let the AI soak up the spike instead.
The mental model that works best here is onboarding. You're bringing on a junior team member who starts rough and gets better every week, and the first few weeks are where the resolution rate climbs fastest. Swytch, an ecommerce customer, described it in exactly those terms, they "treated the AI like a new hire, continuously training and refining." Day one is the worst it'll ever be.
There's a real limit here, which the counter-arguments later return to. The AI is only as good as the knowledge and data you give it (point it at thin or stale docs and you'll get thin or stale answers).
Role 2. The senior human tier, fewer people doing harder work
Once the repetitive tail is gone, what's left for humans is the work humans are actually good at: complex, multi-system problems, emotional or high-stakes conversations, relationships with key accounts, and the genuine edge cases the AI routes up (the stuff we'd never want to automate away anyway). Swytch found their "agents can now focus on what they do best, building relationships and solving challenging issues, without being bogged down by repetitive tasks." Zeffy's team ended up "spending more time on complex technical support cases."
In the teams we work with, you hire for seniority and judgment over raw throughput. The tier-1 pool you'd have grown from five people to twelve stays at five, and those five get more senior over time. Because so much volume is resolved before it reaches them, each remaining agent can also own more of the tickets that do.
This is also where offshore staffing shifts. Tier-1 is often where outsourced teams sit, doing the very macro-driven work the AI now handles, so that layer becomes less necessary. As tickets escalate to tier two or tier three, they come back inside the company to the people who can actually resolve them, and customers tend to get faster answers overall.
I want to be straight about the cost of this, because no other guide is. Customers have told us directly that while it's great the AI handles the monotonous, repetitive work, it has "actually put more intensity into their work." The agents who stay now face only the hard problems, all day. AI-first support is better for the customer and the budget, and at the same time it makes the human half of the job tougher.
Role 3. The AI owner, the genuinely new role
Every AI-first team needs someone who owns the AI agent. This is the role that doesn't exist on any of the traditional org charts, because their teams have no AI to own.
The way we describe it: each company needs someone to manage the AI agent with a little understanding of how it works. They don't need to be super technical, but they do need to be focused on the areas where it isn't performing as well, or where there's room to improve. The job is to review the questions the AI couldn't answer, close the gaps in your documentation, tune the guidance that controls tone and escalation, decide which workflows are worth automating, and run a regular quality check.
It's usually not a full-time hire, at least at first. Time needs to be carved out for it as a percentage of someone's role, and it's a different skill set to traditional support, more time spent understanding what the AI gets wrong and improving the underlying knowledge. In practice it runs to about 30 minutes a week, up to an hour if the owner is hands-on (RecruitCRM, for example, ran a disciplined weekly review to keep their resolution rate climbing).
The toolkit for this role is where a good AI agent really stands out. Ours gives the owner an Insights view that scores every conversation for CSAT rather than the usual small sample, an Inspect view where the team can open any conversation and ask the AI why it gave an answer and where it got the information, and Self-Learning that drafts new knowledge articles by comparing the AI's replies to the human agent's. The title is starting to settle in the market too, somewhere around "AI support lead" or "AI operations."
The capability that feeds it all, data and automation
The fourth piece is less a permanent seat on the org chart and more a capability the AI owner pulls in when they need it. The biggest single lift in resolution rate almost always comes from connecting live customer data, so the AI can answer "where's my order" with the real order rather than a generic policy, and from building automated actions for things like refunds and address changes.
We think of the dev side as the second new role to plan for: getting time and prioritisation from the development team to build the APIs and connections that push data into the agent. Sometimes that's carved out of the existing engineering team, and sometimes it's a developer who sits inside the support team and builds these things on an ongoing basis.
The good news (we tell teams this constantly) is it's rarely as big a job as they fear. Connecting a read-only data lookup is often one to three hours of work, done once and useful forever, and you can now point an AI coding tool at your own codebase to scaffold most of it. Edel Optics added a customer-data connection themselves and watched their resolution rate jump from around 25% to 79% almost overnight.
What an AI-first team looks like in real rollouts
TL;DR: Across real rollouts, teams hold headcount flat, move people onto complex work, and put one person in charge of the AI. Zinc held service steady without hiring; Zeffy's seven-person team stayed seven.
The pattern above isn't a thought experiment. Across our rollouts the same picture keeps showing up: teams hold headcount flat or shrink the tier-1 plan, move humans onto harder work, and put one person in charge of the AI. The numbers here all come from our published customer case studies.
Zinc, a background-checking platform, started from exactly the right question. The challenge their leadership set the support team was, in their own words, "how can we provide the same or better service, without hiring anyone?" The AI now resolves at least 68% of queries, response times dropped under 60 seconds (a goal they'd chased for six months), and CSAT held at 97%. The clearest org signal of all: they didn't need to backfill a team member who left, despite the company still growing.
Zeffy, a free fundraising platform for non-profits, shows the redeployment side. With 84% of tickets deflected, their seven-person CX team has stayed seven people, and they've kept splitting their time 50/50 between support and strategic projects even as the company grew quickly. They describe "receiving a lot less FAQ questions and spending more time on complex technical support cases." The AI soaked up the growth so the humans could keep doing the high-value half of their job.
Swytch, the e-bike conversion company, is the cleanest example of people moving up the value chain. The AI deflects 81% of their volume, more than 4,050 tickets a month handled end to end, which freed their agents up for relationships and the hard problems (and yes, they're the team that onboarded the AI "like a new hire").
YouGarden, an online garden center, puts a number on the headcount you avoid. Their AI resolves 66% of tickets, peaking around 82%, and saves roughly 965 hours a month. That's about six full-time agents' worth of capacity they scaled into without hiring six people, while holding a 78% CSAT score across nearly 12,000 tickets.
Customer.io tells the same story from the SaaS side: 68% deflection, 55 hours of human time saved in the first week alone, and bandwidth "reinvested into other strategic areas."
How to start building your AI-first team this week
TL;DR: You don't reorganise on day one. Audit your ticket mix, appoint an AI owner, run the AI in notes-mode behind your team, and let the structure change as resolution climbs.
You don't reorganise the team on day one, and we wouldn't push you to. The make-up changes gradually as the AI's resolution rate climbs, so the work this week is about setting that in motion, and the restructuring takes care of itself over the following months.
Start by auditing your ticket mix. Pull last month's tickets and tag them into two buckets: the repetitive tier-1 work, and the things that genuinely need a human. The repetitive percentage is the share the AI front line will take, and it's the headcount you can stop scaling (budget about two hours for this).
Appoint your AI owner before you switch anything on. Name the person who'll run the weekly improvement loop, and protect about 30 minutes a week of their time for it. The single most common mistake we see is treating AI support as set-and-forget, and the antidote is simply having an owner from the start.
Onboard the AI in internal-notes mode, behind your existing team. In this mode the AI drafts a reply as an internal note on every ticket without sending anything to the customer, so your agents can check the quality at zero risk and compare it against whatever you run today.
Flip to direct replies on the ticket types it's clearly nailing. On knowledge alone you can be live within minutes to hours, and most teams get to live-and-direct within about a month (this is the bit I'd never rush).
Redraw the escalation path rather than the headcount plan. Decide what routes straight to a human: low-confidence answers, signs of frustration, and any sensitive topics you name. Making escalation easy is also what keeps your resolution number honest, because anything the AI can't handle goes to a person, so the rate reflects real resolutions rather than customers who gave up.
Connect one data source or build one automated action. Pick your highest-volume unresolved topic and close it (we've watched a single data connection do more for a resolution rate than weeks of fiddling), usually one to three hours of API work for the lift it delivers. Then put a 30-minute AI review on your weekly support stand-up, where the owner reports what the AI missed and what they fixed, and watch the resolution rate move over four weeks.
How do I map out my own AI-first team?
Want a head start? Paste your numbers into the prompt below and an AI model will draft the first version of your inverted team. I'd still check the output against your own read of the team, but it gets you off a blank page (and it'll flag anything it can't infer so you're not taking guesses as fact).
You are helping me design an AI-first customer support team using the "inverted support team" model: the AI handles tier-1, a smaller and more senior human team handles escalations and complex work, one person owns the AI, and a slice of developer time connects data and builds automations.
Here is my situation:
- Monthly ticket volume: [e.g. 4,000]
- Current support headcount and roles: [e.g. 6 tier-1 agents, 1 team lead]
- Helpdesk: [e.g. Zendesk]
- Roughly what % of tickets are repetitive — order status, password resets, FAQs? [your estimate, or "unsure"]
- Do you have written documentation / a help centre, and how complete is it? [yes/no + detail]
- Can you connect customer data via an API, e.g. orders or accounts? [yes / no / not sure]
Please give me:
1. My likely repetitive % and which tier-1 work the AI should take first.
2. Which roles I should stop scaling, and which people to move onto more complex work.
3. Who should own the AI, and roughly how much of their week it needs.
4. The first data source or automation worth building, and why.
5. Anything you can't infer from what I gave you — write "ask your team" instead of guessing.When an AI-first team is the wrong fit
TL;DR: If your volume is tiny, your knowledge is thin, or your support is genuinely bespoke, there's no repetitive tail to invert. Build the human team first.
In our experience this model only changes how a team is built when that team has a repetitive tail. If you don't have that tail, there's nothing to invert, and a few situations make the whole approach a poor fit.
Very low volume is the first. Under a few hundred tickets a month, your team doesn't have a tier-1 base big enough to invert, so a small human team is simpler and the AI is a nice addition rather than the front line.
Thin documentation is the second. If your answers aren't written down anywhere, the AI has nothing to learn from, and you get the familiar problem of poor inputs producing poor outputs. But if this is an issue you can always use our historic ticket training feature to bootstrap your docs.
Zinc spent twelve months documenting their processes before they went live, which is a big part of why their rollout landed overnight. Build the knowledge first, and the team follows.
Genuinely bespoke, high-touch, or heavily regulated support is the third. If most of your tickets are already complex, relationship-led, or compliance-bound, your "tier-1" is already senior work (we see this most with high-touch B2B teams), so there was never a wide base to hand the AI. A team already resolving most tickets well has less room to gain, because there's a limit to how far any resolution rate can climb.
And the last one, worth repeating: the work that's left for your humans gets harder. If your team isn't ready to spend their days on only the difficult problems, that intensity is a real cost you should plan for upfront.
The takeaway
TL;DR: Building an AI-first support team means inverting the pyramid: the AI on the front line, senior humans on judgment, and one owner on the improvement loop.
Building an AI-first support team has very little to do with hiring fewer people. The real move is turning the pyramid upside down: the AI on the front line flexing with demand, a smaller and more senior human team on judgment and relationships, and one owner keeping the whole thing improving.
The framework I'd want you to remember is the inverted support team, and the single most important move is to appoint an AI owner and audit your ticket mix before you change anyone's job. The team you build is built around an asset that gets better every week, which is exactly why it isn't a set-and-forget project.
If you want to see this with the numbers attached, the customer rollouts above are the proof, and our guide to AI-first customer support walks through the rollout itself.
FAQs
How do you structure a customer service team?
The traditional answer is a tiered human structure: junior agents on the front line, escalating to seniors and specialists, with several valid arrangements of that (Zendesk alone lists eight). On an AI-first team it inverts: the AI agent becomes the front line, a smaller and more senior human tier handles escalations and complex work, and one person owns the AI's ongoing improvement. The old "5 C's" and "7 C's" of service still matter for the humans, they just apply to a smaller, more senior group.
How many support agents do you need per number of tickets?
There's no clean industry ratio for this, and on an AI-first team the question changes entirely. Your repetitive percentage decides it, the share of tickets the AI can resolve on its own, because that's the volume you no longer staff humans against. In our rollouts that share is commonly 60 to 80%, and because the AI flexes with demand you also stop hiring for seasonal peaks.
Do you still need human agents on an AI-first support team?
Yes, and arguably better ones (we've yet to see an AI-first team that didn't still need its humans). The repetitive work goes to the AI, which leaves your people on the complex, emotional, and relationship-heavy tickets that need human judgment. Teams like Swytch and Zeffy kept their agents and moved them up to harder, higher-value work instead of cutting the team.
What roles does an AI-first customer support team need?
Three, plus a capability: the AI agent as the tier-1 front line, a senior human tier for escalations and complex work, and an AI owner who runs the weekly improvement loop (the new role we keep coming back to). Feeding all of it is a data-and-automation capability, usually a slice of developer time to connect live customer data and build automated actions.
Who owns the AI, and what is an AI support lead?
The AI owner is the person responsible for keeping the agent improving: reviewing what it couldn't answer, closing the gaps in your documentation, tuning its guidance, and running quality checks. We've found this doesn't need to be a full-time or deeply technical role, just carved-out time as a percentage of someone's week and a willingness to learn a different skill set. In the market this role is starting to be called an AI support lead or AI operations specialist.
How do you build a customer service department from scratch with AI?
Start with your knowledge, because the AI can only answer from what's written down, so documenting your common questions and policies comes first. From there, we usually connect the AI in internal-notes mode behind a small human team, appoint someone to own it, then add live customer data and automated actions as you grow. You end up scaling capacity through the AI rather than through headcount.
Does AI customer service mean fewer support jobs?
At the largest incumbents it has meant cuts (Salesforce reduced its support team from around 9,000 to 5,000 in 2025). For most of the teams we work with the pattern is different: they hold their headcount, stop hiring as fast, avoid backfilling, and move people onto harder work. Zinc, for instance, simply didn't need to replace a team member who left.
How long until an AI-first team is up and running?
If you start from existing knowledge like help centers and websites, you can be live within minutes to hours. Most teams reach the point where the AI replies directly to customers within about a month, with the bulk of the setup effort in that first month. After that it settles into roughly 30 minutes a week of the AI owner's time, up to an hour if they're hands-on.
