No. AI is taking call types, not whole jobs, and the two move at very different speeds. The occupation is shrinking slowly while turnover keeps it hiring: the Bureau of Labor Statistics projects customer service representative employment down 5.3% between 2025 and 2035, with 289,500 openings a year across that decade.
If you run a contact center or a BPO calling program, there is a narrower version of the question you can act on. Which of your call types can software finish on its own, which ones split between software and a person, and which ones stay with your team whatever the technology does. The scorecard below settles that one call type at a time.
Is call center employment already falling?
Yes, and it was falling before generative AI existed. Employment at firms whose main business is operating call centers averaged 356,400 jobs in 2024 and 327,600 in 2025, and stood at 307,200 in August 2026, a preliminary figure (BLS series CES6056142001). The annual average has dropped every full year since 2017.
That series is narrower than it looks. It counts jobs at companies whose main business is running call centers, so an in-house support team inside a bank or a hospital is classified under banking or healthcare and never appears in it. Read on its own, it describes the outsourcing industry, not the occupation.
The occupational projection covers everyone who does the job, wherever they sit. BLS put customer service representatives at 2,666,000 in 2025 and projects 2,524,100 in 2035: a loss of 141,800 positions, or 5.3% over ten years. Median annual pay was $44,770 in 2025 (BLS Employment Projections, occupation 43-4051).
Then the same table gives the number nobody quotes. That shrinking occupation is projected to open 289,500 positions a year.
| What is measured | Latest figure | Source |
|---|---|---|
| Jobs at call center firms, NAICS 56142 | 307,200 in August 2026, preliminary, down from a 356,400 average in 2024 | BLS CES6056142001 |
| Customer service representatives, all industries | 2,666,000 in 2025 | BLS Employment Projections |
| Projected in 2035 | 2,524,100, a 5.3% decline | BLS Employment Projections |
| Openings per year, 2025 to 2035 | 289,500 | BLS Employment Projections |
Both of those are true at once, because openings count the people who leave the occupation as well as the positions being added. An occupation can lose 141,800 jobs over a decade and still need to fill a quarter of a million seats a year, since the people in those seats keep moving on. Anyone who tells you the data shows either a collapse or a steady state is reading one column of it.
For a center, that means planning against slow contraction and heavy churn at the same time. Those call for opposite reflexes, which is why the question has to come down to call types.
Does the evidence say AI replaces agents or helps them?
The strongest field evidence is about helping them. Across 5,172 customer support agents, access to a generative AI assistant raised issues resolved per hour by 15% on average, and the gains went mostly to the least experienced, who improved in both speed and quality (Brynjolfsson, Li and Raymond, arXiv:2304.11771).
The most experienced agents saw small gains in speed and small declines in quality. Three other findings in that study matter more for the replacement question than the headline number does.
The gains were largest on relatively rare problems, where agents have the least training to fall back on. The assistant also improved English fluency, particularly among international agents, which cuts against the assumption that offshore delivery is the first thing automated away. And customers behaved better: more polite, and less likely to ask to speak to a manager.
All of that measures people keeping their jobs and getting better at them. It says nothing about how many calls software can finish without a person, which is the number that decides headcount. Figures for that come from vendors measuring their own deployments, so treat them as marketing until someone shows you the call mix behind them. Your own call mix is the better estimator, and scoring it takes about an hour.
Which of your call types can an AI agent take?
Score each call type on six questions instead of arguing about the center as a whole. Call types with a written procedure, data the agent can reach and a short list of outcomes move to software now. Calls that turn on a judgment nobody has written down do not, however simple they look from a report.
The first question decides more than the other five. Which calls an AI agent can finish comes down to whether a new hire could do that call on their first day from what is written down, so that test is where the scoring starts.
| Question about one call type | 0 points | 1 point | 2 points |
|---|---|---|---|
| Is the procedure written down well enough for a new hire to follow it? | It lives in supervisors' heads | Partly written, with known gaps | Documented and current |
| Can the agent reach the data the call needs? | Only in a system with no API | Reachable, read only | Reachable, and it can write back |
| How many outcomes does the call have? | Open ended | A handful, plus exceptions | A short fixed list such as answer, book, reschedule, cancel, take a message |
| Can the caller be identified from what they can say out loud? | Needs a document or an image | Usually, with a fallback | Phone number, order number, or name and date of birth |
| Does the call ever need a policy exception or a pricing decision? | Routinely | Occasionally | Never |
| What happens if the call goes wrong? | Someone is harmed, or money moves irreversibly | Recoverable, with effort | Recoverable with a callback |
Twelve points is the maximum. The bands below are the planning heuristic we use to pick a first call type, not a validated benchmark.
| Total | What to do with that call type |
|---|---|
| 10 to 12 | Route it. The procedure exists, the systems are reachable, and the downside of a bad call is a callback. Start with one of these and read the transcripts before adding a second. |
| 6 to 9 | Split it. The agent takes the front of the call, identifying the caller, collecting what the decision needs and answering what is written down, then transfers. You get the queue relief without handing over the decision. |
| 0 to 5 | Keep it, and score it again in a quarter. Most calls land here on the first two questions, which are fixable, instead of the last two, which usually are not. |
The same scorecard, filled in for five call types most centers have. These scores show the shape of the judgment and are not a benchmark for your center.
| Call type | Procedure | Data | Outcomes | Identity | Exceptions | Risk | Total | Verdict |
|---|---|---|---|---|---|---|---|---|
| Appointment reschedule | 2 | 2 | 2 | 2 | 2 | 2 | 12 | Route it |
| Order or claim status | 2 | 1 | 2 | 2 | 2 | 2 | 11 | Route it |
| Password or account unlock | 2 | 1 | 2 | 1 | 2 | 1 | 9 | Split it |
| Billing dispute | 1 | 1 | 1 | 2 | 0 | 1 | 6 | Split it |
| Hardship or safety call | 0 | 1 | 0 | 1 | 0 | 0 | 2 | Keep it |
Two things tend to surprise people the first time they run this. Call types that feel hard to a supervisor often score high, because difficulty for a person is mostly volume and repetition. And a low score is usually a documentation problem wearing a technology costume. Writing the procedure down is work you own either way.
This scorecard only answers whether software can do the call at all. Once the answer is yes, who should take it is a separate question about cost and control, and outsourcing against AI agents has the weighted scorecard for that one. The staffing consequences are their own arithmetic again, worked through in what automating a call type does to a roster.
What stays with people
Judgment stays, and so does the work at the seam. Calls that need negotiation, a policy exception or a read on someone's tone belong with your team. So does everything that happens at the moment of a handoff, which is where a badly scoped automation shows up first and where a caller feels it immediately.
On Telvana, a transfer is a plain transfer to the number you name. The person who picks up starts the conversation themselves, and the transcript and summary arrive in the dashboard afterwards rather than ahead of the call. The destination has to be staffed and the rules agreed before anything goes live.
Three more pieces of the work do not move.
Deciding what to do about a call nobody answered. Nothing retries an outbound attempt on its own, so an unanswered call is an outcome in your records and a decision for a person, not a loop the platform runs.
Putting things in writing. The voice agent does not send texts during or after a call, because texting is a separate channel with its own agent. A written confirmation is an SMS campaign somebody sets up deliberately.
Reading the calls. Every call comes back recorded, transcribed, summarized and tagged by outcome, which is only worth anything once a person reads some of them. The transfers are the useful ones, because they tell you what is missing from the knowledge base, and what transcripts can and cannot tell you covers how to query a pile of them without buying anything first.
What the agent job becomes
The role shifts from taking the easy calls to handling what the easy calls were hiding. When software takes the short repeatable contacts, the queue a person works is longer and less predictable, so handle time on the human queue climbs even though nobody got slower.
What happens to average handle time runs that out on a worked call mix, where human handle time nearly doubles while total handle hours fall by half. Four parts of the floor change shape with it.
The queue gets harder, so a target set on the old call mix starts reading like a performance problem. Segment handle time by call reason before you automate anything, or you will go looking for a coaching gap that was never there.
Procedures need an owner. An agent follows what you wrote down, which turns writing and maintaining those procedures into a standing job, not an onboarding task.
Transcripts need a reader. Listening to ten transfers and ten resolved calls in the first week tells you more about an agent than any dashboard will, and it is the same muscle call center quality assurance already uses on human calls.
Scheduling changes shape. Coverage for an automated call type stops being a bodies-per-interval problem, while the remaining queues get spikier, so the forecast and the schedule adherence targets around it both need rebuilding instead of scaling down.
If you work the phones rather than run the floor, the parts of the job that get scarcer are the ones a written procedure already covers. The parts that get more valuable are the exceptions, the escalations, and being the person who can tell a supervisor why forty callers asked the same question last week. That last part is our reading and not something the study measured, but it points the same way as the call mix arithmetic: the biggest gains showed up on rare problems, and rare problems are what is left.
Frequently asked questions
Will AI replace call centers entirely?
No. AI agents take call types where the procedure is written down and the outcome is one of a short list. Centers keep the calls that need judgment, negotiation or an exception, plus every transfer destination and the people who read the transcripts and maintain the procedures the agents follow.
Can AI replace call center agents today?
Partly, and only per call type. An AI agent can answer, identify a caller, work from your knowledge, book, reschedule, cancel, take a message and transfer. Score your own call types on the six questions above before accepting any number about how much of your volume that covers.
Are call center jobs going away?
They are shrinking slowly and still hiring heavily. BLS projects customer service representative employment down 5.3% from 2025 to 2035, a loss of 141,800 positions, while projecting 289,500 openings a year over the same decade, because openings count people leaving the occupation as well as new positions.
Will AI replace offshore and BPO call center jobs first?
The evidence cuts both ways. Employment at firms whose main business is call centers has fallen every full year since 2017, and that series is the outsourcing industry. But the one large field study found AI assistance improved English fluency most among international agents. Cost and control per call type is the better frame: outsourcing against AI agents works through it.
How many of our calls can an AI agent handle?
That depends on your call mix, and anyone quoting a figure before seeing it is guessing. Score your top call reasons by volume, add up the ones scoring 10 or more, and treat the total as a starting estimate. Then test the agent on real calls before launch.
What does it cost to try this on one call type?
Telvana starts from $399 a month based on your call volume, plus a one-time setup fee that covers building, testing and launching the agent. There are no tiers to pick between. See pricing.
Where to start
Pull last month's call reasons, sort them by volume, and score the top five on the six questions. The highest scorer is your first candidate. The exercise usually turns up a second thing worth knowing, which is how many of your procedures only exist in somebody's head.
Telvana builds AI agents that answer calls, texts and web chats for contact centers, and transfers to your team under rules you set. Our team builds the agent from your scripts and transfer rules and tests it on your call types, and you approve it before it takes a live call. Book a demo with your scored call types and you will hear an agent handle one. The rest of this cluster is in the AI agents guides.