Conversational AI for support that resolves, then hands off cleanly.

The economics explain why this matters so much. An AI resolution costs around $0.62 against roughly $7.40 for a human-handled ticket, and a blended model of AI plus fast human escalation cuts cost per resolution by some 71% at a satisfaction cost of just five hundredths of a point. The prize is real, which is exactly why the temptation to claim it dishonestly, by counting contained contacts as solved, is so strong, and why the line between deflection and resolution is the one that decides whether the saving is genuine or borrowed from next month.

Containment is the most misleading number of all: the customer did not reach a human, so the system logs a success, whether or not the problem was solved. The data is blunt about the gap, with AI containing 45% or more of contacts but only around 14% genuinely resolved end to end without a person. A bot optimised for containment becomes a barrier customers fight to get past, which is why the only honest design has no dead end, only resolution or a clean escalation. Gartner expects agentic AI to resolve 80% of common service issues by 2029; reaching that is about real resolution rather than a higher containment score.

The old chatbot followed a fixed decision tree: it matched a keyword, offered a menu, and dead-ended when the customer said anything it had not been scripted for. That is the experience that taught a generation of customers to type agent immediately, and it deflected by exhausting people rather than helping them. A modern support agent is a different thing: it reasons about the request, retrieves the relevant policy, and takes the action, a refund, an account update, an order change, by connecting to your help desk and order systems rather than reading out a script.

That shift, from matching keywords to resolving requests, is what moved automated support from a cost-cutting annoyance to something that genuinely resolves a large share of contacts. The agentic systems now reach autonomous resolution on the majority of routine requests precisely because they can do the task rather than describe it. We build the resolving kind, connected to the systems where the work really happens, because a support agent that can answer but not act is a more articulate version of the chatbot customers already learned to skip past.

We measure the things that predict whether customers are helped, not the ones that flatter a report. Resolution rate, verified by what happened after the conversation rather than by the bot marking its own homework. Re-contact rate within a couple of days, which catches forced closures. Knowledge coverage, the share of questions your content can genuinely answer, because an agent cannot resolve what your knowledge base never documented. And the CSAT gap between AI and human handling, measured separately rather than blended into one figure that hides which is which.

Measuring AI satisfaction separately matters because automation handles the easy contacts and humans take the hard ones, so a blended score tells you nothing about either. Tracked apart, the data often shows AI satisfaction matching or beating humans on the contacts it should handle, and falling short on the ones it should not, which is exactly the signal you need to set the escalation line. We build the measurement in from the start, with audit-ready logging of resolutions and handoffs, because a support system you cannot honestly measure is one you cannot improve, and one whose vendor reports only the flattering number is one to distrust.

The satisfaction gap between AI and human support has effectively closed for the right contacts, and remains real for the wrong ones. On structured, routine requests, AI now scores around 4.1 out of 5 against a human 4.3, a gap that narrows to five hundredths of a point under good hybrid escalation, and on the most structured intents like password resets and order status it matches or beats human handling outright. The reason is simple: those problems get solved instantly, without a queue or an availability window.

The honesty is in naming where it falls down. On emotionally charged contacts, a complaint, a cancellation, a customer who is already angry, autonomous AI scores far lower, around 3.3, because what those customers need is acknowledgement and judgement rather than a correct answer. We tune the system to recognise sentiment and route those contacts to a person quickly, so AI handles what it is good at and humans handle what they are better at. Forcing an agent to handle an emotional complaint to protect a containment number is how a support operation damages the relationships that matter most, and we design against it deliberately.

The savings are large and well documented. An AI-handled resolution runs around $0.62 against roughly $7.40 for a human one, first response times have dropped from hours to minutes, and total resolution time has compressed by something like 87% in AI-native operations. A blended model, AI handling the routine and humans the rest, cuts cost per resolution by around 71% at a negligible satisfaction cost, which is the configuration that truly works rather than the all-AI fantasy that damages CSAT.

The reason the case holds up only with resolution, not deflection, is that a deflected-but-unsolved contact comes back, and a returning customer costs more than one helped the first time. Savings built on containment are borrowed against future re-contacts and lost goodwill; savings built on genuine resolution are real and compounding. We model the economics on resolved contacts and the blended human-plus-AI flow, because that is where the documented numbers really come from, and we would rather show you a credible 71% on resolution than an impressive number on a metric that unravels the moment the deflected customers come back.

In support, a confident wrong answer is worse than no answer, because the customer acts on it. An ungrounded model invents answers 15 to 27% of the time, which in a support setting means telling customers things about your policies, your products and their accounts that are simply untrue. Grounding the agent in your own help content, policies and resolved tickets through retrieval brings that rate down to somewhere around 1%, and lets every answer cite the source it came from, so a claim can be checked rather than trusted blindly.

Grounding only works if the knowledge is there and current, which is why knowledge coverage, the share of real questions your content can answer, is one of the metrics we track from the start. Where the knowledge base has a gap, the honest behaviour is to escalate rather than improvise, so the agent knows what it does not know and hands off instead of guessing. We build on the same governed retrieval as our knowledge systems work, kept current as your policies change, because a support agent grounded in last year's refund policy is a liability dressed up as automation.

Almost everyone has tried this and far fewer have shipped it: around 64% of customer-experience teams ran an agentic AI pilot in 2026, but only about 27% had even one channel fully in production. The gap is the same one that defeats AI everywhere, the distance between a demo that impresses and a system that holds up under real, varied, high-stakes customer contact, where the cost of a wrong answer is a damaged relationship rather than a bad test result.

Closing it means building the accuracy and the guardrails in from the start rather than retrofitting them after the first incident with a real customer. The grounding, the escalation, the measurement and the sovereignty are what turn a promising pilot into a channel you can trust with your customers, and they are exactly the parts a quick proof-of-concept skips. We build for production from the first day, because a support agent that works in a demo and embarrasses you in front of a customer is worse than no agent at all, and the 27% who shipped are the ones who treated it as a production system rather than an experiment to bolt onto the queue.

Conversational AI fits a support operation with real volume and a core of routine, repeatable contacts: the password resets, order statuses, refund requests and policy questions that make up most of the queue and need answering fast rather than thoughtfully. A business fielding the same questions thousands of times, or one whose support cannot keep up at acceptable cost, has exactly the profile where automation resolves a large share at a fraction of the cost and frees the human team for the contacts that need them.

Where the contacts are mostly complex, emotional or high-stakes, we will say so. A support operation that is overwhelmingly complaints, sensitive cases or bespoke problems may get little from autonomous resolution and should keep humans in front, perhaps with AI assisting them rather than replacing them. Low volume is the other case where the build does not pay, and we would rather tell you that than sell an agent for a queue that a person could clear by hand. The technology earns its place against high-volume routine work, and where your support is not that, the honest answer is to keep it human or to wait until it is.

A support agent speaks for your brand on every contact, so how it speaks matters as much as what it resolves. The tone, the formality, the warmth, the house style, is something we tune to your brand rather than leave at a generic default, because a customer notices when the support voice is jarringly off from the rest of the company. An agent that resolves correctly but sounds wrong is still a poor representative, and consistency of voice across thousands of conversations is one of the quiet advantages a well-built system has over a room of differently-trained people.

Just as important is what it must never do. Guardrails define the commitments it cannot make, the topics it must not engage, the claims it is forbidden to invent, and the actions that always need a human, so the agent cannot promise a refund outside policy or improvise an answer on a regulated matter. We set those limits to your business and your compliance obligations, because an agent speaking for you needs the same boundaries you would put on a new hire, and rather more, since it speaks to far more customers far faster than any person could.

A support agent should not be frozen at the quality it launched with. Every conversation is data: the contacts it resolved, the ones it escalated, the questions it could not answer, and the human corrections on the cases it got wrong. Fed back, that stream shows where the knowledge base has gaps, which intents are ready to automate next, and where the agent is drifting, so the system improves with use rather than slowly degrading as products and policies change around a fixed configuration.

We build that improvement loop in, because it is the difference between a support agent that earns more of the volume over time and one that plateaus and then quietly slips. The escalations point to the next intents worth automating; the failed answers point to the content worth writing; the corrections refine the handling. This is the same discipline of measurement and iteration we apply across our AI work, applied to support, and it is what turns a launch into a system that is resolving meaningfully more a year later than on its first day, rather than one nobody has touched since the project closed.