Here are the nine that survived, with their dates on them. Agentic AI to resolve 80% of common customer service issues by 2029, Gartner, March 2025. Half of the companies that blamed AI for headcount cuts to be rehiring by 2027, Gartner again, February 2026, off 321 leaders polled the previous October. A 14% lift in issues resolved per hour, from the only large measurement of live support agents anybody has published. 64% of customers would rather you did not use AI on them at all, off 5,728 people. 2,814,000 Americans doing the job in 2024 and 341,700 openings a year regardless. CA$650.88, which is what one wrong chatbot answer cost in a tribunal. The rest are below with the same treatment.
What you will not find here is forty numbers. There are pages on this search carrying forty and better, and if you go and pull the thread on any given one of them you tend to arrive at another listicle, then a third, then a vendor’s landing page from 2023 with no sample size on it and no month. The number survives all that travelling, funnily enough, and none of the rest of it does. The year it was about, the firm that made it and the count of people it came off do not, on account of nobody quoting a statistic ever having room for the boring half of it.
How a number loses its year between the study and your slide
Say you go looking for one figure to put on a board deck. Not a research project, ten minutes on a Tuesday. The first page gives you a percentage in a big font with no date beside it, the second gives you the same percentage attached to a different year, and the third has it in a sentence that starts “studies show”, which is the phrase that means the writer never opened the study either.
None of those three went and did anything wrong exactly. Each one copied accurately, mind you. What drops off is a qualifier at a time, and the qualifiers were the part that made the number mean anything. The word forecast goes first. A prediction about 2029 turns up on the next page with the 2029 mislaid, then the poll of a couple hundred managers stops mentioning the couple hundred managers, and somewhere along the line a survey that a software company ran on its own prospects quietly stops being a survey that a software company ran on its own prospects. By the fifth telling you have a bare claim about what AI does to a support queue this morning, and the thing it actually came from was one analyst’s view of the back end of this decade.
That is the whole of the mechanism, and it is why this page is short on numbers and long on provenance, on account of provenance being the part that decides whether a figure is any use to you. Every figure below carries who collected it, when, off how many people, and the thing it does not say.
That 80% figure, with the year left on it
The most-quoted number in this category came out of a Gartner press release dated 5 March 2025, and it says that agentic AI will autonomously resolve 80% of common customer service issues without human intervention by 2029, with a 30% reduction in operational costs going along with it. The analyst named on it is Daniel O’Sullivan, and the release is public, so none of this is hard to check.
Read the qualifiers your own self, because there are four of them doing heavy lifting and the citations drop the lot. It is 2029, mind you, not now. It is common issues as well, and that is a category nobody has gone and defined for your desk. It is a prediction rather than a count of anything that has happened. And the 30% is on operational cost, which is not the same animal as your bill.
Fair is fair, it is a serious forecast from a serious firm and it may well come in. Just do not let it onto a slide about this year’s budget, on account of it not being about this year at all.
The same firm, pointing the other way
Here is the part that gets left out of every roundup we read while writing this one. Gartner also put out a release on 3 February 2026 predicting that by 2027, half of the companies that attributed headcount reduction to AI will go and rehire staff for similar functions, under different job titles. That came off a poll of 321 customer service and support leaders run in October 2025, with Kathy Ross and Emily Potosky named on it.
The same release carries a duller figure that we would put above the forecast if anybody asked us, truth be told, which is that only 20% of customer service leaders had actually reduced agent staffing because of AI. Most of the cutting that got blamed on AI was not really AI at all, it was the ordinary economy going through a bad stretch and picking up a more modern explanation somewhere on the way.
And there was an earlier one in the same direction, 10 June 2025, predicting that by 2027 half of organisations that expected to significantly reduce their customer service workforce would abandon the plan, off 163 leaders polled that March, with 95% of them saying they intended to keep human agents in the loop. So one firm has published both the 80%-by-2029 line and the you-will-be-hiring-them-back line inside eleven months. Anybody quoting one of those at you and not the other is not lying, they are picking, and picking is the ordinary condition of this entire subject.
The one number that came off a measurement
Nearly everything above is somebody’s forecast or somebody’s poll of managers. There is one large piece of work that instead went and measured what happened to real support agents with a real assistant, and it is worth more than the rest of this page put together.
Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the staggered rollout of a generative AI assistant across a customer support operation. The working paper, NBER 31161, issued April 2023 and revised that November, reports 5,179 agents and a 14% average increase in issues resolved per hour. The published version in the Quarterly Journal of Economics, volume 140 issue 2, 2025, reports 5,172 agents and 15%. Same study, two numbers.
Both are printed here on purpose. The sample moved by seven people and the headline moved by a point between the working paper and the journal, which is entirely normal in academic work and is also precisely why you see 14% on some pages and 15% on others, with neither one telling you there is another version. That is the honest texture of a real number, and it is the sort of thing that gets sanded off in a roundup.
The distribution underneath the average is the useful bit of it anyway, and it is worth going and sitting with for a minute. Novice and low-skilled agents improved by 34%. The most experienced and highest-skilled saw very little, and in the published analysis a small decline in quality. The mechanism the authors suggest is that the assistant spreads the habits of the best people on the floor to the newest ones, which means the thing is functioning as training that happens to arrive inside the workflow. Customer sentiment improved. Retention went up as well, which is the finding that gets quoted least of the lot and probably matters most to whoever does your hiring. During software outages, when no suggestions came through at all, the agents who had used it most held onto part of their gains, which reads like people having genuinely learned something rather than merely leaning on a crutch.
Notice what the study did not test. One company, one assistant, one stretch of time, and the models in the field now are not the models that were in the field then. It is not a deflection rate, it is not a resolution rate, and it says nothing whatever about replacing anybody. It says your least experienced people get much better and your best people mostly do not. Which, if you have ever run a desk, sounds about right.
What the people on the other end said about it
Gartner ran a customer survey of 5,728 customers, published on 9 July 2024. The fieldwork was December 2023. 64% said they would prefer that companies did not use AI in their customer service. 53% said they would consider switching to a competitor if they found out a company was going to use AI for service. The concern that came up first was not accuracy or privacy, it was the difficulty of reaching a human being. The same release put 60% of customer service leaders under pressure to adopt AI, which is the sentence that makes the other two land properly.
That survey is getting on for three years old at time of writing and attitudes move, so treat it as the last good measurement of a mood rather than today’s mood. We went and looked for a newer one of the same quality and did not find it.
There is a newer read, and we may as well tell you who paid for it. The Zendesk CX Trends 2026 report was fielded in June 2025 across 22 countries, with more than 11,000 respondents, 6,182 of them consumers and 5,115 on the business side, and it reports that 83% of consumers think experiences should be better than they are today. That is a large, properly documented survey, and it was published by a company that sells AI customer service software. Both halves of that sentence are true at once, so use the number and keep the interest in mind while you do.
Employment, which is a decline and not a collapse
The US Bureau of Labor Statistics had 2,814,000 people employed as customer service representatives in 2024, with employment projected to decline 5% through 2034 and median pay of $42,830 a year, or $20.59 an hour, as of May 2024. The page was last revised on 28 August 2025. Both figures sit on the one page.
Now the number that never makes it into anybody’s roundup, on account of it being no use to either side of the argument. About 341,700 openings for the role are projected each year, on average, across that same decade. Turnover and retirement create far more vacancies than the automation removes, mind you, and they will go on doing it all the same. A 5% decline over ten years on a base near three million is a real contraction and it is also nothing whatsoever like the disappearance the headlines have been selling since 2023, and the two facts sit together perfectly comfortably once you stop needing the subject to be dramatic.
What a wrong answer cost, the once it was properly counted
Most of what gets written about AI accuracy is either a vendor’s benchmark or a screenshot of a bot behaving badly. There is one figure with a tribunal behind it. It is a small figure.
In Moffatt v. Air Canada, decided by the British Columbia Civil Resolution Tribunal on 14 February 2024, a customer was awarded CA$650.88 in damages for negligent misrepresentation after the airline’s chatbot gave him wrong information about bereavement fares, coming to roughly CA$812 with interest and fees. The airline argued, and this is the part worth the whole of the case, that the chatbot was a separate legal entity responsible for its own statements. The tribunal did not accept it.
The money is small enough to be a rounding error, no two ways about it. The principle underneath it is not small in the least, on account of it meaning that whatever the thing says on your behalf gets treated as a thing you said. Whoever is signing off on an AI customer support agent should read that one before they read another benchmark.
Two figures we went and left out
There is a line doing the rounds that only 14% of customer service issues get fully resolved in self-service today, and it is attributed to Gartner nearly everywhere it appears. We could not trace it to a dated release from here. It may well be theirs and it may well be sound, but the rule over here is that we print no benchmark we cannot define, on account of an untraceable number being indistinguishable from an invented one, and so it is not up there with the nine.
A second figure went the same way, that being the per-resolution cost comparison you see everywhere, some cents for an AI answer against several dollars for a human one. Every version of it we could find traced back to another content page. What you can verify your own self is the list price. At least one major platform publishes 99 cents per resolution on its public pricing page, and a per-resolution model is a straightforward thing to model against your own volume. Take your monthly conversation count, decide honestly what share of it is the simple repetitive kind, and multiply. That arithmetic is yours and it beats anybody’s benchmark, including ours.
Where we sit in all this, said plainly
Over here at Maxdesk we make a help desk sat on one shared address, and the ticketing system puts states and owners onto the mail. The free plan runs at $0 with our ads on it, Pro costs $20 in the month for the workspace entire, however many people are sat at it, and Elite at $99 brings the AI layer, which is two agents rather than one. Triage does the sorting, so it sets a priority, files the category and picks whoever carries the ticket. The resolver drafts a reply and sends only where its confidence is high, and everywhere else that draft sits still for a person to read. There is an allowance of 5,000 AI responses inside Elite with paid packs beyond it, and mail is the only door we watch.
The honest bit, which cuts against us
We are a vendor publishing statistics about our own category, and you should read this page in that light the same way we asked you to read the vendor survey above.
There is no resolution rate of our own on this page either. We have not run a study of the sort Brynjolfsson and his colleagues ran, we do not have five thousand agents to run it on, and any accuracy percentage we printed about our own resolver would be a marketing number with our thumb on it. A vendor scoring its self is not much of a measurement. So there is not one here, and where you find one on somebody else’s page, put the same three questions to it that this page has been putting to everybody all the way down, which are who counted the thing, when they counted it, and off how many people they counted it.
The other thing we will concede is that the AI comes only at $99, bundled into Elite, and the responses are metered while the rest of our bill is flat. If your interest in this subject is purely the arithmetic of automating a queue, that metered line is the one to model, not the seat count.
Four numbers of your own, this month
Before you quote anybody’s percentage, go and get four of your own, and none of it needs a tool or a project.
Take last month’s mail and count the total. Then count how many of those were the same handful of questions, the ones a knowledge base page would have answered, and that share is the only honest ceiling on what automation can do for you, whoever sells it to you. Then count how many needed a person to make a judgement call. Then count how many came back a second time after you had answered them, because a repeat contact is the failure that all the deflection arithmetic in this industry quietly forgets.
Four counts, an afternoon with a spreadsheet, and no vendor involved in any of it. You will have collected the lot your own self, which is more than can be said for most of what is written on this subject. Whatever you learn from your own four beats every figure above, on account of the numbers on this page being about the industry and those ones being about you. And if the second and fourth counts come back large together, that is the shape that says your problem is a written-down-answers problem rather than an AI problem, and the fix is customer self service and cleaner helpdesk automation before anybody’s agent. Track it with the same customer satisfaction metrics you already run, mind you, and if csat meaning is still fuzzy on your team, settle that first, on account of a satisfaction score nobody agrees on being a poor referee for an automation argument. The export off your email management software is all any of it takes.