How AI Agents Evolved for Customer Support in 2026

A year ago, an AI Agent in customer support meant a chatbot that answered a question from a knowledge base, and handed off anything it did not recognize.
That description no longer covers what shipped in 2026.

Across the Customer Experience products this year, Agents started fixing their own mistakes, closing sales inside a support conversation, working through business logic that used to require a form, coordinating with sales and success on the same account, investigating a ticket before a Support Specialist opens it, dialing outside companies and sitting through their hold music, and even patching a software bug before an engineer sees the report.

From answering to self-improving

For most of the AI Agent’s short history, improving it meant a person reading through failed conversations, spotting a pattern, and rewriting an instruction by hand. Decagon’s Duet Autopilot, introduced in June, automated that loop. Autopilot analyzes production conversations, traces a failure back to its root cause in the Agent’s operating procedures, proposes a fix, and tests that fix against a curated set of past conversations before a human ever sees it. If the fix does not hold up, Autopilot keeps iterating until it does. A reviewer’s job shifts from writing the update to approving a diff with test results already attached.

Decagon has kept building on that foundation. The governance update shipped in September added Vault, a place for a team to store the brand voice, policy, and off-limits guidance that shapes what Autopilot is allowed to propose, along with a way to specify which outcome, deflection, CSAT, or a custom metric, should win when Autopilot finds more than one valid fix. The company also split Autopilot’s work into two tracks: one that reacts to errors the moment an alert fires, and a separate audit process aimed specifically at the subtler deflection and CSAT gains that error-fixing tends to crowd out.

Adam Liu at GlossGenius has described the value in practical terms: Duet and Autopilot catch opportunities to improve Agent logic that would be nearly impossible to find by hand. That is the shift in one line. The Agent used to be the thing a team maintained. Now the Agent maintains itself, with a person reviewing the result.

From resolving to selling

A second shift moved in the opposite direction: instead of narrowing what the Agent does, some vendors widened it into revenue.

Maven AGI’s Maven Commerce, announced in September, gives an existing support Agent the ability to sell a product, a subscription, or a service inside the same conversation that started as a support issue. A customer calling about a broken charger can leave the call having also bought a spare, because the Agent that verified the warranty is the same one that noticed a pattern worth mentioning. Maven scopes this with a Charter that defines exactly what the Agent can offer, at what price, and to whom, so the boundary between a helpful upsell and an invented discount is enforced rather than left to the model’s judgment.

Gorgias took a related bet from the other direction: rather than teaching a support Agent to sell, it rebuilt Chat itself into a shopping surface. The redesign, detailed in a product update, replaced external product links with shoppable cards, clickable replies, and instant cart and order access inside the conversation. Gorgias reports a 36 percent increase in engagement with recommended products and more than double the add-to-cart rate following the change, attributing the lift to keeping the entire shopping journey inside one conversation instead of routing shoppers out to a separate page.

Both moves point at the same underlying assumption: a customer does not experience support and sales as separate systems, and an Agent that already holds the account context is better positioned to notice a sales moment than a separate tool bolted on afterward.

From scripts to procedures

Complex, multi-step requests, the kind that touch billing systems, verification logic, and more than one internal tool, used to sit outside what a support Agent could safely automate. Fin’s Procedures are built specifically for that gap. A team describes a process in natural language, the way they would train a new hire, and Fin combines that outline with existing content and past conversations to draft a Procedure the team can refine. Inside a Procedure, teams can add branching logic, use code to enforce rules like eligibility checks or date calculations, and connect Fin to systems like Stripe or Shopify through data connectors or MCP.

The distinguishing idea is deterministic control layered on top of agentic reasoning. MONY Group’s AI and Solutions Manager described moving cashback claim investigations from a static web form to a Procedure that checks account information through live API calls and guides the customer through the entire claims process dynamically, adjusting when the conversation does not follow a straight line. Fin validates every Procedure against simulated conversations before it ships, and reruns that test library whenever a Procedure changes, functioning as regression testing for conversational logic.

From a single silo to a revenue-wide system

Support has traditionally been the last stop in the customer relationship, disconnected from what sales and success already know about an account. Attio’s Workflows, launched in June, is aimed at that gap from the CRM side: an engine for orchestrating Agents across prospecting, triage, briefing, follow-up, and stalled-deal recovery, all grounded in what Attio calls Universal Context, a shared picture of every account across conversations, signals, and connected tools including Linear and Fin.

The premise is that an Agent’s output is only as good as what it can see. A churn-prevention Agent that only reads sales call history misses the customer who stopped engaging after a bad support experience. Attio’s context layer is built to surface that kind of cross-functional pattern automatically, rather than depending on someone in sales remembering to check the support queue. For CX leaders, the relevant detail is less the CRM feature set and more the direction it signals: support data is increasingly treated as an input to revenue Agents outside the support org entirely, whether or not the support team was consulted on that integration.

From deflection to full investigation

Most AI Agent products optimize for one number: how much volume the Agent resolves without a person. Pylon built its Agentic Customer Support platform, launched in open beta in July, around a different premise entirely: deflection Agents solve a small slice of the actual work. The majority of support labor is not answering the easy questions an Agent already handles. It is gathering context, investigating a hard issue across half a dozen systems, coordinating a handoff, and following through, and that work has stayed manual even at companies with a mature deflection Agent in place.

Pylon’s Background Agents start investigating the moment an issue arrives, checking connected systems, searching related issues, and pulling account and product context before a human opens the ticket. Support teams write their own investigative processes as reusable Skills in natural language, so the strongest technical support engineer’s mental checklist becomes something any Agent, or any teammate, can run.
Early-access customers including Cognition, Hex, and Nominal reported one team cutting L1-to-L2 escalations by 70 percent, another reducing full-day investigations to one or two hours, and a third handling growing volume without the additional headcount it had already budgeted for. Pylon says, “the goal is not a few more deflected tickets, it is changing what a support team spends its day doing”.

From assisting the human to coaching the Agent

The AI copilot, an assistant that drafts a reply for a person to review and send, has become close to standard across the category. Gorgias used its comparison of AI copilots to make a structural argument about the category’s limit: a copilot makes the human faster, but every ticket it touches still reaches a person and still consumes handle time. When an automation rate plateaus, a copilot does nothing to lift it, because it was never pointed at the automation in the first place.

Gorgias built Gaia to sit one level up from that pattern. Instead of drafting a reply for a Support Specialist, Gaia reads a team’s ticket data to find where the AI Agent itself is failing, recurring requests it does not cover, unnecessary escalations, knowledge gaps, and proposes or generates the Skills and Actions to close those gaps, with every change requiring explicit approval before it goes live. Nicole Parker at Bare Necessities described Gaia auditing an existing knowledge base, flagging duplication and missing information, and producing a ready-to-review draft within minutes. The distinction Gorgias draws is a useful one for evaluating any AI investment in this category: a tool that makes a person faster is a linear gain, while a tool that improves the Agent doing the resolving compounds every cycle.

Into the codebase

Support and engineering have always shared a border, usually crossed by a person copying a bug report into an issue tracker and writing a summary of what a customer experienced. Linear’s engineering team has spent months narrowing where its own Agent can safely cross that border unsupervised. In a first-person account from the team, engineer Igor Sechyn described the two categories the Agent handles reliably today: cleaning up dead feature flags once a flag has evaluated the same way long enough to confirm the gated code is unused, and patching failing background tasks, where the Agent gets the fix right on the first attempt roughly a third of the time and still saves the assigned engineer a chunk of investigation time even when it does not.

The team added confidence gates that make the Agent demonstrate it understands a problem before it attempts a fix, both to cut down on off-target pull requests and to avoid burning tokens on a low-probability attempt. Sechyn’s broader point applies well beyond Linear’s own codebase: an Agent’s capability is not a fixed property of the technology, it is a function of how much context and how many guardrails a team has built around it, discovered through direct experimentation rather than assumed from the start. For CX leaders whose escalations already flow into Linear, that lower-level engineering work is part of what determines how fast a bug reported by a customer actually gets fixed.

Onto the phone network

Some of the hardest support work has never happened inside a helpdesk at all. It happens when a company has to call another company: a healthcare provider confirming a prior authorization with an insurer, a pharmacy transferring a prescription, a claims team checking status with a payer.
Sierra’s IVR navigation work, published in August, tackles a problem that sounds mechanical and turns out to be genuinely hard: getting a voice Agent through a phone tree built for a person.

The obstacles are specific. An IVR often opens with a disclosure that requires silence, not a response, and answering too early derails the call. Some prompts expect a spoken keyword, others accept only touch-tone input, and confusing the two fails immediately. Detecting a live person at the end of hold music is unreliable, since recorded greetings during a hold can sound identical to an actual associate answering the phone. Sierra built navigation, memory of which paths work on a given line, and retry logic directly into its Agents, and measured the effect in a controlled benchmark: adding IVR navigation lifted the pass rate on a standardized set of phone trees from 57 percent to 85 percent. A revenue cycle management company using the capability reports a 70 percent overall success rate reaching payers by phone, rising to 94 percent for specific payers.

The work is a reminder that a meaningful share of support automation still depends on infrastructure built decades before the Agent doing the calling existed.

The execution gap underneath all of it

Every capability above assumes the Agent’s job ends when it resolves the conversation. Brainfish’s Ballet, launched publicly in August by Brainfish CEO Daniel Kimber and Ajain Vivek, is built on the observation that resolving a ticket and finishing the work behind it are two different things. A refund marked resolved still needs processing. An account update still needs to land correctly across a billing tool, a CRM, and whatever else needs to agree on what happened. Kimber’s research into this gap surfaced ops teams describing themselves as human middleware, manually moving information between systems that do not talk to each other, a description that will be familiar to anyone running a support queue.

Ballet is positioned as an orchestration layer built for ops teams to describe and adjust directly, rather than a tool that requires engineering support to maintain. One mid-market B2B SaaS company is auto-resolving 40 percent of its support volume through a Ballet workflow, not deflecting it away from a human but actually completing the follow-through work, with roughly $120,000 in annual savings and two to three FTEs of freed capacity, according to Brainfish. Whatever a support Agent’s answer accuracy looks like, if nobody, human or Agent, does the record-keeping and system updates behind that answer, the ticket is not actually finished, and 2026 was the year that gap started getting its own dedicated product category.

Where this leaves support leaders

None of this changes the constraint we wrote about back in August: data quality is still the rate limiter on every one of these capabilities.

• A self-improving Agent still needs well-written operating procedures to improve against
• A Procedure still needs accurate live system data to reason over
• An orchestration layer still needs a support team willing to write down its own process before an Agent can run it

The products got substantially more capable in 2026. Whether that capability turns into results still depends on whether the operational groundwork underneath it gets built, and that part still has not shipped in a box.