"We Replaced 40% of Tier-1 Tickets": A Support Director on What Agents Actually Handle

When Maya Chen, Director of Customer Support at a mid-market B2B SaaS company, tells people her team "automated away 40% of Tier-1 support tickets," she's learned to brace for two reactions: skepticism from peers who've heard the claim before, and enthusiasm from executives who want the same number on next quarter's dashboard.

Neither reaction is quite right. Chen's company—a B2B platform with roughly 180,000 monthly customer contacts across email, chat, and an in-app help widget—didn't eliminate 40% of its support work. It automated a precisely scoped slice of repetitive, well-defined requests, and in doing so, changed what the remaining work looks like for the humans who stayed.

That distinction is the whole story. In this interview, Chen walks through the baseline, the task-by-task breakdown, the rollout missteps, and the metrics that actually told the truth about whether the AI agent deployment worked. For support operations leaders planning their own rollouts, her core message is blunt: the headline percentage is the least interesting number in the room.

The Setup: What "40% of Tier-1 Support Tickets" Actually Meant

The Baseline Before AI Automation

Before the first AI agent went live in early 2025, Chen's org handled about 180,000 monthly contacts with 62 frontline agents across three channels: email (55%), live chat (30%), and in-app widget (15%). Average handle time ran 11 minutes for email and 7 minutes for chat. First-contact resolution hovered at 68%, and CSAT sat at 4.3 out of 5.

Those numbers matter because they define the denominator. "When I say 40%, I mean 40% of Tier-1 resolved-without-human-touch volume, measured against our own historical Tier-1 baseline," Chen says. "Not 40% of all contacts. Not 40% of headcount. Anyone who quotes a percentage without telling you the denominator is selling something."

The baseline also included a cost figure the finance team cared about more than any service metric: a fully loaded cost per Tier-1 resolution of roughly $6.40, driven mostly by labor, with a median queue wait of just under nine minutes on email and under two minutes on chat. Those two numbers—cost and wait—became the real scoreboard, because they're what the business could feel. (See related: how to calculate cost per ticket in your support operation.)

One caveat worth stating up front: a lower cost per resolution only becomes savings when volume drops or agents are redeployed to higher-value work. Chen's team reports both gross AI cost per resolution (inference, licensing, and an amortized share of build) and net savings after redeployment, precisely so finance doesn't double-count.

Defining Tier-1 in This Organization

Chen's team classified Tier-1 as requests solvable with documented knowledge and no account-level judgment: password resets and MFA lockouts, order and subscription status, billing FAQ questions, how-to and configuration guidance, and basic integration troubleshooting with known error codes.

That definition is deliberately narrow. "We excluded anything requiring a refund decision, a policy exception, or a conversation about a customer's frustration," she says. "Those are human categories by design, not by accident." Because the org published its Tier-1 taxonomy internally, the 40% figure is reproducible—an engineer can pull the query and verify it. That auditability, Chen argues, is what separates a real support metric from a slide deck.

The taxonomy is a living document, versioned in the same repository as the help center and reviewed quarterly by a standing group of one support lead, one engineer, and one finance analyst. Every contact type carries an owner, a definition, and an example ticket. "If two people can't classify the same ticket the same way," Chen says, "the category isn't ready to automate." (Related: building a support ticket taxonomy that scales.)

Reconciling the numbers. Tier-1 accounts for roughly 70% of total contact volume in this org. Of that, about 40% now resolves without human touch—which works out to 28% of all contacts. The remaining 30% of total volume was never Tier-1 to begin with. If the arithmetic doesn't close in your own operation, the headline number is hiding something.

What AI Agents Actually Handle Now: A Task-by-Task Breakdown

Before the categories, a definitional note that shapes everything below. Chen's team tracks three distinct outcomes:

Containment — the AI resolves the request and the customer does not reopen or re-contact within 7 days.
Deflection — the contact never reaches a human, whether or not it stayed resolved.
Assisted resolution — the AI drafts or gathers context; a human reviews and sends.

The headline 40% is containment, not deflection. "Deflection numbers look great and lie quietly," Chen says. "A ticket the bot closes and the customer reopens two hours later was never resolved." Her team measures recidivism explicitly: any contact that returns within the 7-day window is subtracted from containment and logged as a failure for weekly review.

Fully Automated: The High-Volume, Low-Ambiguity Work

Roughly 28% of total contact volume now resolves end-to-end without a human. The categories are unglamorous, and that's the point. Containment figures below are measured over a trailing 90-day window and are this org's own numbers—not industry benchmarks.

Password resets and MFA lockouts — 94% containment, with automatic escalation if the account shows recent security flags. The agent authenticates against the identity provider through a read-only scope, triggers the reset, and never sees a credential. Because these flows are a favorite target for account takeover, every reset requires a step-up verification challenge, and all agent actions are written to an immutable audit log.
Subscription and order status — 91% containment; agents read from the billing system via a scoped, read-only API and never write. The most common sub-intent, "why did my invoice change," resolves by diffing the last two billing periods and explaining the delta.
Billing FAQ — invoice explanation, plan comparisons, and tax documentation requests — 82% containment. Plan comparisons are generated from the live pricing table rather than a static article, which eliminated a recurring source of wrong answers.
Known-error troubleshooting — the agent matches the error code to a runbook and walks the customer through documented fixes. Coverage is limited to error codes with a documented resolution path; anything unmapped escalates with the code attached.

Escalation triggers are explicit: any request touching refunds, credits, contract terms, or security incidents routes to a human immediately. "The agent's job is to know what it doesn't know," Chen says. "We tuned for precision over coverage. A contained ticket that should have escalated costs more trust than ten escalations that should have contained."

That tuning had a concrete cost. Chen estimates the team deliberately left 8–10 points of potential containment on the table by keeping thresholds conservative in the first two quarters. "We could have shown a bigger number in month three," she says. "We'd have spent the next year explaining the CSAT dip." (Related: setting escalation thresholds for AI support agents.)

Assisted Resolution: The Middle Band Most Rollouts Forget

The most underreported part of the deployment isn't the automation—it's the ~12% of total volume where the AI drafts and a human sends. Here the agent assembles context (account history, prior tickets, relevant knowledge articles, suggested resolution) and a human reviews, edits, and delivers. Median handle time on these contacts fell from about 9 minutes to under 4, because the agent does the retrieval and the human does the judgment.

This band matters for two reasons. First, it's where trust is built: agents see the AI's suggestions succeed before they're asked to trust it end-to-end. Second, it's where the org catches its own errors—assisted drafts are reviewed and tagged, and recurring corrections feed directly back into the knowledge base and the escalation thresholds. "The assisted tier is our quality control," Chen says. "If we'd gone straight to full automation, we'd have shipped our mistakes to customers instead of to reviewers."

What Still Requires a Human, By Design

The remaining volume—non-Tier-1 work plus escalations—stays human. Refunds and credits, contract and policy exceptions, security incidents, frustrated customers, and anything an agent flags as ambiguous. Chen is emphatic that this isn't a failure of the AI: "We didn't shrink the human job. We changed it. Agents spend less time on resets and more time on the conversations that actually need a person, which is also the work that retains them."

The Rollout: What Went Wrong First

The clean numbers above are the second attempt. The first launch, in Q1 2025, was scoped too broadly and shipped without the assisted tier. Two failures stand out:

Over-broad intent matching. The agent attempted to handle "billing questions" as a single category, including ones that bordered on refunds. It produced confident, wrong answers on proration edge cases. The fix was the taxonomy discipline described above: every automatable intent got an owner, a definition, an example, and a hard escalation boundary.
No recidivism tracking. Early containment looked strong because the agent closed tickets. Reopen rate wasn't measured until a customer-success leader noticed repeat contacts spiking. Adding the 7-day reopen window cut the reported containment by roughly six points overnight—and made the number honest.

The lesson Chen draws is unglamorous: the agent is only as good as the boundary around it. Broad scoping produces impressive demo numbers and expensive cleanup.

The Metrics That Actually Told the Truth

Chen's team tracks five numbers, in priority order:

Containment net of 7-day reopen — the headline, defined precisely.
Escalation-appropriateness rate — the share of escalations a reviewer agrees should have escalated. This is the precision metric behind "precision over coverage"; the team targets ≥95% and treats misses as P1 quality bugs.
CSAT on AI-touched vs. human-only contacts — tracked separately, never blended. A blended CSAT hides a CSAT gap.
Cost per resolution, gross and net — with inference, licensing, and amortized build broken out.
Agent time reallocated — hours moved from Tier-1 to escalated and non-Tier-1 work, which is where the real retention and quality gains show up.

"Executives want one number," Chen says. "Give them one number, but make it the honest one, and keep the other four underneath it. The 40% is real. It's just the least interesting thing about what changed."