On 11 September, Salesforce added seven new AI agents to Agentforce, each with a first name and a bounded job description. Casey resolves customer service issues across voice, SMS, WhatsApp and web chat. Piper qualifies inbound leads. Marshall runs back-office processes and leaves an audit record of every action. Six of them are generally available now; one is still in pilot.
The list is precise, and it is worth reading for what it leaves out. Among the things Salesforce says Casey handles is escalation to a human. That belongs there – it is a capability any customer service agent needs. But the description stops at that point. Who that human is, what kind of calls they are left with, and what their job looks like after Casey has been running for three months is not addressed. It rarely is, by any vendor, ourselves included.
The claim in this piece is simple, and it is uncomfortable for everyone selling this: a voice agent does not make the working day easier for the people who answer the phone. It makes every single call they take harder. That is not a weakness in the agent. It follows directly from what the agent is good at.
What is left is not the old queue in miniature
A voice agent owns the calls where the right answer sits somewhere you can look it up: in a calendar, on an order line, in a rule someone wrote down. Booking, rebooking, opening hours, the status of a delivery, a change of address. Which calls it takes is not arbitrary. It is the definition of which calls it can take.
The rest follows. What is left over is precisely the complement: the calls where the answer is not written down anywhere. The exception. The complaint. The case that has already been through twice. The customer who is annoyed before she has said what it is about. The question that requires someone to make a decision on the company's behalf.
Before the agent, that was perhaps one call in five, spread across the day with easier calls in between. The easy calls were not just volume. They were the breathing space between the heavy ones, they were how new people learned the job, and they were what gave a day on the phones a rhythm. Take them away and what remains is a concentrate.
This is not a forecast. It follows from which calls the agent takes. If it handles 60 per cent, and that 60 per cent is the easiest, then the remaining 40 consists entirely of the hardest.
Three things change for the person who answers
None of them shows up on a report of answer times.
The first is pace. A mixed queue has natural gaps in it; a queue of difficult calls does not. Average handling time goes up, and that is usually read as a sign that someone has slowed down. The opposite is true.
The second is emotional load. The share of calls that open with an unhappy customer rises – not because customers have got angrier, but because the satisfied ones got their answer without speaking to anybody. The person answering experiences the first of those anyway. It is worth saying out loud to the team, because the alternative is that people work it out for themselves and conclude the job has got worse for reasons nobody has explained.
The third is that ambiguity moves upward. An agent set up properly hands over when it is uncertain. It therefore systematically passes on what is unclear and keeps what is clear. The human ends up not merely with more difficult cases, but with a higher proportion of cases that have no single right answer at all.
The counterargument is a good one, and it deserves to be put properly
There is a solid answer to all of this, and it is not an evasion.
Volume falls. If the queue shrinks from a hundred calls to forty, the person answering has more time per call – and difficult calls are exactly the ones that need time. A short, rushed conversation about a complaint is worse for both parties than a long, calm one.
The handover can carry context. The agent has already identified the customer, found the order and heard what the issue is. Done right, the human does not start at zero but at minute three. That is precisely where most set-ups fail, but it is a failure that can be fixed, not a law of nature.
And the industry is investing in the human side. TELUS Digital reports that as of June this year it had run more than 90,000 simulated training calls in its own agent-training tool, that onboarding time for new hires is down 20 per cent, and that it is seeing early indications of lower turnover. Centrical reports improvements of up to 10 per cent among customers using its coaching capabilities. Newo.ai counts a transfer to a human as a successful outcome rather than a failure.
The strongest objection, though, is a different one, and it is worth meeting head on. If headcount stays where it is while volume falls, each person gets fewer and heavier calls – but fewer of them. Load per person need not rise at all; it can fall, even as load per call rises. That is a real possibility, and it does happen in organisations that keep their people.
The problem is that this assumption is the opposite of how the purchase is usually justified. In the Nordics, where labour cost is the most expensive input in almost any customer conversation, the business case for a voice agent is written strikingly often as a headcount case. If headcount falls in step with volume, calls per person are unchanged – and every one of them is harder than before, with nothing offsetting it.
All of this is correct. The question is what it presupposes.
