How Much of Our Customer Support Can AI Actually Handle?

How much of our customer support can AI actually handle?

AI can handle between 40% and 70% of routine, well-scoped customer support contacts, measured as genuine resolution rather than as deflection. That range comes from the benchmark surveys set out further down, and where you land inside it depends on three things about your queue rather than on the software.

The first is how repetitive your contacts are. A business where most inbound messages are five questions with the same five answers is a different proposition from one where every ticket is a different account in a different state of exception.

The second is whether the answer exists in a system a machine can read. "Where is my order" is answerable if the system holds a tracking number. "Can you make an exception on this" is not answerable by anything, because the answer is a judgment somebody has to own.

The third is how much authority you are willing to give it. A system that can read is useful. A system that can act, issue the refund, change the appointment, reschedule the delivery, resolves things. Most disappointing deployments are read-only systems being measured against act-and-resolve expectations.

The published figures bear that out if you read them carefully. The two largest independent samples sit at the bottom of the range: Comm100's 2026 AI Live Chat Benchmark Report, drawn from more than 220 million live chat interactions, found AI chatbots resolving 44.8% of them, and Crisp's 2026 state-of-AI-support survey put it at about 40% of conversations closed without a human for teams running an AI agent in production. Vendor-published figures sit much higher: a dataset of 195 deployments across 38 vendors, compiled by My AskAI, gives a median of 70% with half the field between 56% and 80%, and individual vendors publish their own averages above that. Both sets describe real systems. They are not measuring the same thing, which is the subject of the next section.

What is the difference between deflection and resolution, and why does it matter?

Deflection counts contacts that did not reach a person. Resolution counts problems that were solved. The gap between those two numbers is the part of the project that failed, and it is usually invisible.

The mechanism is ordinary. A customer asks something the system cannot do. The system answers with the closest thing in its knowledge base. The customer gives up and closes the window. No agent was involved, so the ticket counts as deflected. On a deflection dashboard that is a win. In the business it is a customer who did not get helped and did not tell you.

There is now a number on how much the label changes the answer. Lorikeet, analysing published deployment figures, found deployments whose figure was labelled "resolution" had a median of 72.5% across 108 cases, against 61% across 52 cases for figures labelled "automation". Same category, same year, eleven and a half points of difference produced by the choice of word. Several vendors say this openly, which is to their credit, and it is still the most common way these projects report success they did not earn.

Metric	What it counts	What it hides
Deflection or containment rate	Contacts that did not reach a human	Customers who gave up
Resolution rate	Contacts where the issue was actually closed out	Nothing, if measured on the customer's confirmation rather than the system's
Escalation rate	Contacts handed to a person	Nothing, and a healthy number here is good news
Abandonment rate	Conversations the customer left mid-flow	Usually not measured at all, which is why it belongs on the list
Repeat contact rate within 7 days	The customer coming back about the same thing	A resolution that was not one

The two to insist on are resolution confirmed by the customer and repeat contact rate within a week. Together they are difficult to fake, and a vendor who will report both is a vendor whose numbers you can use.

The customer's own account is the harshest version of this. Ada's research with NewtonX, published on 24 March 2026 from fieldwork between December 2025 and February 2026, surveyed 2,000 consumers who had used AI-driven support in the previous year plus 500 enterprise decision-makers, and found 24% saying their most recent AI support interaction was fully resolved by AI alone. The other 76% escalated to a human, got partial resolution, or gave up, and NewtonX's own write-up puts 11% as abandoning entirely without complaining or following up. That 11% is the deflection gap with a number on it. Two caveats you should carry: Ada sells AI customer service software and commissioned the research, and the consumer sample was screened to people who had already used AI support, so it does not describe consumers in general.

What does the independent evidence actually say about AI in support?

The independent evidence on AI in support says three things worth more than the statistics pages, and only one of them is about resolution rates.

On the agent side the evidence is strong and positive. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the staggered rollout of a generative AI assistant to 5,172 customer support agents, published as Generative AI at Work in the Quarterly Journal of Economics in May 2025. Access to the assistant raised issues resolved per hour by 15% on average. The gain was concentrated: less experienced and lower-skilled agents improved both speed and quality, while the most experienced saw small speed gains and small declines in quality. This is the best-evidenced finding in the whole category, and it is about helping agents rather than replacing them.

On the customer side, the Ada and NewtonX study above found what people actually object to, and it is not the software. Consumers who had a bad AI interaction described the system failing to understand them, hitting the edge of what it could do, and repeating the same unhelpful answer. Meanwhile 59% said they would prefer instant round-the-clock AI support over waiting for a human, as long as it can resolve the issue. Worth noting that Ada's own research landing page shows 47% for that same preference where its press release shows 59%, which is a useful reminder to read the methodology on any figure you plan to repeat, this page included. People are not against AI support. They are against AI support that cannot finish the job.

On self-service generally, the long view is the most useful thing on this page. Gartner reported in a press release dated 25 September 2019, from studies of more than 8,000 customer journeys, that about 70% of customers used self-service somewhere in their journey while only 9% resolved their issue completely that way. A later Gartner survey of more than 5,700 customers put full self-service resolution at 14%. Ada's 2026 figure of 24% fully resolved by AI alone is the third point on that line. The number is moving, and it is moving more slowly than the marketing implies.

Forecasts are a different genre and should be read as one. Gartner predicted in a press release dated 5 March 2025 that agentic AI would autonomously resolve 80% of common customer service issues without human intervention by 2029, with a 30% reduction in operational costs. That is a forecast about a category four years out, not a benchmark for your deployment this year.

What does it cost us to answer a ticket today?

What it costs you to answer a ticket today is almost certainly more than you have calculated, and the calculation is the part worth doing before you buy anything.

The durable published figure is old and primary. Gartner's 2019 customer service and support leader poll put the average cost of a live contact, meaning phone, live chat or email handled by a person, at $8.01, against roughly ten cents for a self-service contact, as reported by DestinationCRM at the time. Newer medians circulate widely, $13.50 for an assisted contact against $1.84 for self-service, and we could not trace either to a Gartner publication, so they are not on this page.

Your own number is better and takes ten minutes. Count contacts per month. Divide total support payroll, loaded, by contacts handled. For the loaded part, the US Bureau of Labor Statistics puts private industry total compensation at $46.89 per hour worked as of June 2026, in a release dated 9 September 2026, with wages at $32.82 and benefits at $14.07, so benefits run about 30% on top of pay. Apply that ratio to what you actually pay, not to the national average.

Then split the result. If four hundred of a thousand monthly contacts are five repeated questions, those four hundred are the automatable population, and the value of automating them is four hundred times your cost per contact, minus what the system costs to run, minus the contacts it fails and hands back. That last term is the one vendors leave out of the business case, and it is the difference between a projection and a forecast.

One warning on the arithmetic. A support cost per contact that includes your best agents' time on hard problems will overstate the saving available on easy ones, because easy contacts are cheap contacts. Price the automation against the cost of the tickets it will actually take. The general method, including how to avoid counting time that moved rather than disappeared, is in how to measure whether AI automation is saving money.

What happened when a company went all in on AI support?

The best-documented case of a company going all in on AI support is Klarna, which went most of the way and then bought some humans back. Both the company's version and its chief executive's version are worth reading.

Klarna announced in late February 2024 that its AI assistant had handled about 2.3 million conversations in its first month, roughly two thirds of its customer service chats, equivalent to the work of about 700 full-time agents. Those were company figures, not externally audited, and they made Klarna the reference case for the whole category for a year.

In May 2025 the chief executive, Sebastian Siemiatkowski, told Bloomberg that cost had become too dominant a factor in how the support operation was designed and that the result was lower quality. He said customers should always be told a human is available, and the company began recruiting customer service staff again on a flexible remote model.

Klarna disputes the word reversal and its position is on the record. It told Fortune it remained "very much still AI-first", that it was keeping its policy of not replacing departing employees apart from freelance agents in its outsourcing division, and the company has noted that its first-line support before the assistant was largely handled by outsourcing partners whose staff were reassigned rather than by Klarna employees who lost jobs.

Both accounts point at the same lesson, which is why the dispute matters less than it looks. The failure was not the technology. It was optimising a support operation for cost per contact and discovering that quality was the thing being traded away, with no measurement in place that would have caught it earlier. Design the measurement so a quality regression shows up before a customer tells you about it on social media. What that looks like in practice, including sampling and rollback, is in what happens when AI gets it wrong.

Which support tickets should AI never answer?

Five categories of support contact should never be answered by AI, and they are the same five in almost every business.

Anything where the answer is a judgment about an exception. Goodwill, a refund outside policy, a deadline waived: the system can gather the facts and present them, and a person decides. A machine that learns to say yes to exceptions has quietly rewritten your policy.

Anything with money moving that the customer did not initiate. Reading a balance is fine. Taking an action against it without an explicit instruction is not.

Anything that is actually a complaint. A customer who is angry needs acknowledgement from somebody with authority, and the fastest way to make a bad situation permanent is to answer it correctly and impersonally. Route these to a person immediately and measure how fast that happens.

Anything regulated. Medical advice, legal advice, insurance coverage statements, anything a regulator would read as a representation from your business. The adjacent cases are on the dental and medical practices board, where the line between scheduling and advice decides what can run.

Anything safety related. A gas smell, a flood, a locked-out tenant, a patient in pain. These are the highest-value contacts your business receives and they belong with a person in seconds. On the trades side, the capabilities and where they stop are on the home and trade services board.

How do we put AI into support without making customers angry?

Put AI into support in four steps without making customers angry, and the first two involve no customer-facing automation at all.

Step one, agent assist. The system drafts replies, summarises the account and surfaces the relevant policy, and a person sends everything. Nobody outside the company notices anything except faster answers. This is the step with the strongest evidence behind it and the lowest risk, and plenty of businesses should stop here for a quarter.

Step two, triage and routing. Classify, prioritise, attach account context, flag urgency. Still no customer-facing text. Measure how often the classification is right before trusting it.

Step three, resolution on a narrow set. Pick the three or four highest-volume questions where the answer lives in a system and the action is reversible, and let it handle those end to end. Narrow and complete beats broad and partial, every time.

Step four, widen slowly, with two rules that do not bend. A human is always reachable, visibly, in one step, which is the point Klarna's chief executive made after the fact. And the system says what it is, because being helped by software annoys far fewer people than being misled about it.

Two conditions sit underneath all four steps. Your knowledge base has to be true, because a system working from an out-of-date policy will apply it confidently a thousand times before anybody notices, and is our data good enough for AI covers how to tell. And somebody has to own the quality number after launch. Systems BLACKSIG builds run on our infrastructure and we operate them from there, so the monitoring and the drift checks sit with us, and the policy calls stay with your people.

How do we measure whether it is working?

Measure whether it is working on five numbers, reported weekly, and resolution rate is only one of them.

Resolution rate, confirmed by the customer rather than inferred by the system. Escalation rate, which should be healthy rather than low. Repeat contact rate within seven days, which catches resolutions that were not. Abandonment rate inside automated conversations, which almost nobody measures and which is where the deflection gap hides. And customer satisfaction split by handling path, so an AI-handled score and a human-handled score are visible separately instead of blended into one number that hides both.

Add one qualitative check that costs an hour a week: read twenty AI-handled conversations. Not the flagged ones, a random twenty. Teams that do this find problems weeks before the dashboard does, because a dashboard cannot see a technically correct answer that reads as dismissive.

Set the baseline before anything launches. Current cost per contact, current resolution time, current satisfaction, current repeat contact rate. A project without a before is a project whose results are an argument rather than a measurement, and that argument is why most support automation cannot prove its own value. How to build that baseline properly is in how to measure whether AI automation is saving money, and the difference between the kind of system that can resolve a contact and the kind that can only answer one is in AI agents versus chatbots.

Frequently asked questions

How much does it cost to automate customer support?

What moves the number is how many contact channels you run, whether the answers live in systems a machine can read and write, how many of your contacts are genuinely repetitive, how much authority the system is given to act, and how much review stays in the path. A business with a thousand monthly emails and five recurring questions is a different job from one with phone, chat, email and a marketplace inbox across three brands. BLACKSIG does not publish a rate card; scope is set on the call after the strategy work has said which parts are worth doing. The drivers are in how much AI automation costs.

How does AI customer support actually work?

A contact arrives, the system classifies it, pulls the relevant account context, and either answers it, acts on it or hands it to a person, with a confidence score on each decision. Anything above your threshold is handled automatically. Anything below it goes to a queue with the reason attached. A human stays reachable in one step throughout, and every automated resolution is logged so the quality can be sampled afterwards.

Is an AI support agent better than outsourcing to a support provider?

Different trades, and plenty of businesses run both. An outsourced team gives you judgment, languages and coverage, and it needs managing, training and quality control. A system gives you instant answers on the repeatable share and does nothing outside its scope. The honest comparison is per contact type rather than wholesale: automate the questions that repeat, outsource or staff the ones that need a person, and refuse any proposal that treats all contacts as one population.

Can our support be automated if our answers are in people's heads and not a knowledge base?

Not yet, and that is the real finding rather than a blocker. A system can only answer from what exists in a form it can read, so the first phase of most of these projects is turning what the team knows into something written down and current. That work has value on its own: it also makes onboarding faster and reduces how often the same question reaches your most senior person. Is our data good enough for AI covers how to assess it.

What resolution rate should we expect in the first month?

Lower than the number in the proposal, and that is normal rather than a failure. The figures from production deployments cluster around 40% of conversations closed without a human, per Crisp's 2026 survey of teams running an AI agent in production, and they improve over months with active tuning, which means the first month measures your scoping rather than the technology. Judge the trend over a quarter, on resolution confirmed by customers and repeat contact rate, and hold the vendor to reporting both.

How Much of Our Customer Support Can AI Actually Handle?

The short version