Salesforce ran its own agents on its own support desk, and published the numbers

An Authority Report is where we read what people with real operating responsibility said in public, in their own words, and work out what it means for a company that isn’t them.

This one is worth reading closely for a structural reason rather than a dramatic one: Salesforce sells agent software and deployed it on its own support desk at scale. Vendors rarely do that, and when they do they rarely publish numbers that cut both ways.

The numbers

Support went from roughly 9,000 people to about 5,000, with support costs down around 17 % since the start of 2025.

Salesforce’s own figures as reported by Fortune. Self-resolution rates for cases handled without a person are in the mid-eighties.

Benioff’s own framing separated two parts of the business:

In engineering, I’ve achieved higher productivity, so I’m maintaining a flat headcount. When it comes to customer support, I’m reducing headcount because I don’t need as many support agents.

And, more bluntly, on the support side: “I need less heads.”

The asymmetry in that first quote is the part to sit with. Same company, same technology, two different conclusions, flat in one function, down sharply in another. That is a more informative result than a single company-wide number would have been, because it says the effect depends on the shape of the work rather than on the technology.

Why support and engineering came out differently

Support work is high-volume, repetitive, and has a defined correct answer for a large share of contacts. That is the profile a system handles well, and the gains show up as fewer people needed for the same volume.

Engineering work is variable, judgement-heavy and constrained by things other than typing speed: review, coordination, deciding what to build. Productivity gains there do not translate into fewer people; they translate into the same people producing more, which is why headcount went flat rather than down.

The useful generalisation is not “AI replaces support and augments engineering”. It is that the gain shows up as headcount reduction only where the work is repetitive enough that volume is the constraint. Everywhere else it shows up as throughput, and throughput does not reduce a payroll.

The number that is missing

Deflection rates in the mid-eighties are a striking figure and an incomplete one. Deflection counts contacts that never reached a person. It does not count contacts that reached a person angrier, reopened cases, or customers who gave up.

That gap is exactly what Klarna’s CEO went on the record about after cutting support too far; we read that story closely, and the mechanism is worth repeating here: a deflection metric looks better every month that the experience gets worse, because the customers who cannot get through stop trying.

Nothing in the public Salesforce reporting suggests that happened.

The number to ask for alongside deflection, from any vendor and about your own deployment: reopen rate, or the share of escalations that arrive already angry. A deflection figure quoted alone is a figure quoted without its control.

What this means at 9 support people instead of 9,000

Almost every company reading this has a support team in the single digits. Three things transfer, one does not.

The one that does not: the headcount conclusion. At nine people you are not going to five. The gain shows up as the same nine people handling a growing volume without hiring the tenth, which is a real result and a completely different budget conversation.

Split your contacts before you split your team. Salesforce’s result came from a support desk with a large, well-understood repetitive tier. Yours has one too, and you probably know roughly what it is: password, order status, invoice copy, delivery date. That tier is where the return is and where a mistake is cheap.

Watch the second number from day one. Reopen rate within 48 hours. If it climbs while deflection looks good, you are paying for deflection, not resolution. This applies whether you buy or build, and when you buy, note that both major vendors bill per resolution and define “resolved” in their own favourDE (auf Deutsch).

Design the exit before the automation. The question is never “what can the agent answer”. It is “what happens to the ones it cannot, and how fast does a person see them”. If the answer involves a queue nobody owns, the agent is not ready to be in front of customers.

The read

The honest way to use this case is neither as proof that agents replace support teams nor as a vendor claim to discount. It is a large, well-instrumented deployment by a company with every incentive to report the flattering number, and the informative part is that they reported a split result rather than a uniform one.

Flat in engineering, down sharply in support, same technology. If you take one thing from it, take that: the question is not whether AI works, it is which of your functions is shaped like support.

Frequently asked questions