The support floor I worked on went from 34 people to 19 in eight months. The automation everyone had been warned about arrived four months after the second round of cuts, and for its first six weeks it could not reliably tell a customer whether a shipment had left the warehouse.
So here is the claim I want to make and then spend the rest of this piece earning: most of the customer service roles being eliminated under the heading of AI job displacement were already scheduled to go. The model is not the cause. It is the permission slip.
When a company cites AI in a customer service layoff, the technology is usually real but insufficient, and the reduction is usually a cost decision that predates the technology by a fiscal quarter or more. That is the verdict. Everything below is why I think it holds, and what it changes about how you read the next memo.
Nothing in this piece is affiliated with anything. I have no course to sell you and no certification that makes this go away. I worked a queue for fourteen months at a mid-size logistics software company, watched the team get cut twice, then spent eleven months applying to 341 jobs and getting six phone screens. That is the whole of my authority here.
What the memo says and what the org chart says
Layoff memos that invoke AI share a vocabulary, and the vocabulary is worth learning because it tells you which of two things is happening.
When a memo says the company is simplifying operations, that is a description of spans and layers. It means managers with four reports are being consolidated into managers with eleven, and it happens whether or not a single model is in production. When a memo says the company wants to accelerate output, read it as a ratio: the same ticket volume divided by fewer people. That is a headcount statement wearing a capability statement's clothes.
And when a memo announces AI investment and a return-to-office requirement in the same breath, notice what has been bundled. A relocation requirement is a second reduction mechanism, quieter than the first, and it does not appear in the layoff number. People who cannot move resign. Resignations do not come with severance, do not appear in the WARN filing, and do not show up in the percentage the press picks up. Uber's cut to its community operations team came packaged this way, roughly a tenth of the team out and the rest called back in, both framed as part of the same AI-forward strategy.
I am not saying the AI is fake. I am saying the memo is doing two jobs at once, and only one of them is about technology.
The three things a company means when it says AI
When you hear that customer service is being automated, the sentence covers at least three distinct situations that deserve different amounts of your anxiety.
The first is deflection that genuinely works. On my queue, order-status questions were 41% of inbound volume. Where is my shipment, has it left, why does tracking say one thing and the site say another. That volume is repetitive, the answer lives in a database, and the customer wants a fact rather than a judgment. Deflection at that layer works, it has worked since before anyone called it AI, and it removes seats. If your day is mostly tier-zero lookups, the exposure is real and it is not new.
The second is restructure-first. This is the logic that says you cannot layer new technology on top of fragmented processes, so the reorganization comes before the capability. Under this reasoning, the cut is a precondition, not a consequence. The company is not replacing you with a model. It is thinning the org so a future model has somewhere to sit. The distinction matters because it explains something people find confusing: the announcement arrives before the technology is good, and the technology being bad does not bring the jobs back.
The third is narrative. Eight hundred people cut for weak demand reads as distress. The same eight hundred cut for AI reads as investment. One version invites questions about the business. The other invites questions about the roadmap. I am not going to speculate on any individual executive's intent, because I do not know it and neither does anyone writing about it. But the incentive is legible without speculation, and any explanation that flatters the person giving it deserves a slower read.
Most real announcements are a blend. The trouble is that the blend is presented as if it were entirely the first thing.
What the public record actually shows
A few reference points, each attributed to its moment rather than presented as the current state of anything.
In 2023, IBM's chief executive told Bloomberg the company expected to pause hiring for roles it thought AI could take over, describing something in the range of 7,800 back-office positions over about five years. That number got quoted for years afterward as though it had already happened. It was a projection about hiring, not a completed reduction.
In 2024, Klarna said its AI assistant was handling the workload of roughly 700 full-time agents. It was the most quoted number in the category. By 2025 the company had publicly acknowledged that quality had suffered and said it was recruiting human agents again. Both statements came from the same company about the same system, roughly a year apart.
In 2025, Salesforce's chief executive said publicly that support headcount had gone from around 9,000 to around 5,000 as AI took on a large share of conversations. Duolingo's AI-first memo the same year produced enough public backlash that the company spent weeks clarifying what it had meant. Snap, Block, and others made cuts in the same period with similar framing.
Read as a set, these do not describe a clean substitution of machines for people. They describe companies experimenting in public, over-claiming, correcting, and reducing headcount throughout — which is a different phenomenon from the one the phrase automation displacement usually conjures.
How I'd actually read your own exposure
This is the part I wish someone had given me while I still had the badge. Not a prediction about the industry. A set of named criteria you can apply to your own seat this week.
| Signal | What it looks like | What it means for you |
|---|---|---|
| Repetition | The top five ticket reasons cover more than half your volume | High exposure. Repetition is the raw material of deflection. |
| Channel | Chat and email vs. phone vs. in-person | Text queues automate first. Voice lags, mostly for handoff and liability reasons. |
| Liability | Disputes, medical, financial adjustments, anything with a regulator | Lower exposure. Someone has to own the decision, and companies want that someone to be human. |
| Training data | You write macros, tag transcripts, correct bot answers | Highest short-term value, lowest long-term safety. You are the input. |
| Instrumentation | Handle time, CSAT, resolution rate all already tracked per person | High exposure. A job that is fully measured is a job that can be modeled, and modeled work gets targeted first. |
| Revenue attachment | Retention, saves, upsell, churn prevention | Lower exposure. Cost centers get cut. Desks with a dollar figure attached get argued about. |
| Employment status | BPO vendor, contractor, agency, FTE | Vendor contracts get cut first because they do not require a layoff. |
If you scored high on repetition, text channel, and instrumentation, and you are on a vendor contract at a cost center, your seat is exposed regardless of how good any model currently is. That combination was true before the technology existed. It is what got offshored in 2011 and consolidated in 2019.
If you scored the other way — voice, liability, revenue-attached, on payroll — you are not safe, but you are not the first call either. The teams that go first are the ones that are cheapest to remove, not the ones that are easiest to replace.
What the reversals actually buy you
The rehiring stories get passed around as reassurance, and they should not be. When a company walks back an automation-first posture, the jobs tend to come back in a different shape.
The honest version of the Klarna arc is not that humans won. It is that a company discovered a quality floor, and hired back toward it on terms that suited the company. When roles return after an automation push, they frequently return as contract rather than payroll, at a lower band, and with a new title that contains a word like quality, escalation, or oversight. The work is real work. It is often more interesting work. It also frequently pays less than the seat that was removed, and it comes with an implicit expiry: you are being hired to cover a gap that the company is actively spending money to close.
So the reversal is not a reprieve. It is a repricing. Both things can be true at once — the technology is worse than claimed, and your job is still worth less than it was.
Three moves, and what is wrong with each
I am not going to hand you a plan that works. I will hand you three that are defensible, with the objection attached to each, because a recommendation with no downside listed is marketing.
Move toward the work that generates escalations rather than absorbs them. Save desks, disputes, trust and safety, anything where a wrong answer costs the company money it can name. These teams get defended in budget meetings because someone can point at a number. The objection: this work is harder, the burnout is worse, and the triage layer above it is being automated too. You are moving from the first row of the cut to the third row.
Move toward the tooling side. The person who designs the macros, tunes the deflection flows, and audits the model's answers is on the other side of the ledger. In practice this means learning the admin layer of whatever platform your company runs, volunteering for the bot QA rotation nobody wants, and being in the room where the deflection targets get set. The objection is the obvious one and I am not going to dress it up: you are building the thing that shrinks the team, possibly including your own seat, and these projects have a pattern where the team gets smaller once the thing ships. What you get in exchange is a transferable skill and about eighteen months of leverage. That is a trade, not a rescue.
Keep your own ledger, starting now. This is the one I did too late. The memo will describe you in aggregate — as part of a function, a percentage, a cost line. Your defense against being read that way is a private record of what you specifically did, in numbers, updated monthly. Tickets is the wrong unit. Nobody hiring cares that you handled 60 a day; that is a commodity claim and every applicant makes it.
What travels is the second kind of line. Compare these two:
- Handled 60+ customer support tickets daily with 94% CSAT
- Owned the refund-exception queue; rewrote 40 macros and cut escalation rate from 18% to 7% over two quarters; the deflection flow shipped on my documentation
The first describes a seat. The second describes an operator, and it survives the sentence "we are simplifying operations." The objection here is honest too: this takes months to accumulate, and the market for support roles is soft, so having the better resume line does not make the search fast. It made mine less humiliating, not shorter.
Who this applies to, and who it doesn't
This is for you if you work a queue, or manage people who do, at a company that has started using the word AI in all-hands meetings. It is for you if you are watching hiring patterns to figure out which industries are actually contracting versus which are rebranding a contraction. And it is for you if you have noticed that opportunity language and headcount reduction keep arriving in the same email, and you wanted someone to say plainly that the pairing is not a coincidence.
It is not for you if your role carries regulatory sign-off, if you handle the accounts large enough to have a name rather than a number, or if you work in field or on-site service. Your exposure is real on a five-year horizon and close to irrelevant on a one-year one, and planning around the wrong horizon costs you money.
It is also not for you if you are looking for a reason to relax. I do not have one. The honest read is that this category of work is being repriced, that the technology is a genuine part of that and a smaller part than the memos claim, and that the repricing was underway before any of it shipped.
What I would take from all of it is narrower than a prediction and more useful. Stop asking whether a model can do your job. Start asking whether your job has been made legible enough to be modeled — fully measured, fully repetitive, fully documented by you — because that is the condition that precedes the cut, and it is the one you can still act on.
The myth is that AI is coming for customer service jobs, and that your seat disappears the day the model gets good enough. The more accurate version is that the seat was priced out first, and the model's real job is to make a number on a spreadsheet look like a strategy.
