Software has been sold by the head for twenty years. You count the people who need access, multiply by a monthly figure, and the invoice is predictable whether those people do ten things or ten thousand. In the past fortnight Salesforce put a different proposition on the table, and the question a buyer has to answer changed with it.
01 — What happened
On 11 September 2026, Salesforce announced what it calls a portfolio of job-ready agents: seven named agents, each scoped to a role rather than a feature. Casey resolves customer service issues across voice, SMS, WhatsApp and web chat. Piper engages and qualifies inbound leads from websites and inboxes. Hunter runs outbound pipeline from research through outreach. Carter handles product discovery and checkout, Paige takes IT and HR requests, Marshall orchestrates back-office processes, Fin works complex customer workflows. Six are generally available; Hunter is in pilot with general availability stated for November 2026.
The more interesting part of that announcement is the runtime underneath it. Salesforce describes a long-horizon runtime that lets an agent “pursue goals across days and weeks” rather than completing a single task or interaction, with memory preserved across sessions and execution that resumes after interruption. Its own worked example is a seller asking Hunter to rescue at-risk deals before quarter end, with the agent turning that into a plan and working towards it.
Four days later, on 15 September, Salesforce unveiled AIforce at Dreamforce — an interface layer that carries Salesforce data, workflows, permissions and governance into whatever interface someone is already working in, launching with Claude, Slack and Agentforce Coworker. Marc Benioff’s framing was that “AI is creating an interface revolution”. It follows the Claudeforce partnership announced on 26 August, under which Claude became the default reasoning model behind parts of Agentforce.
Now the part that will actually reach your invoice. Salesforce’s published pricing lists three ways to buy the same work: $2 per conversation for customer-facing agents; Flex Credits at $500 per 100,000 credits, where a standard agent action consumes 20 credits and a voice action 30 — roughly $0.10 and $0.15 an action; or a per-user add-on at $125 per user per month for unlimited employee agent usage.
It is not an isolated move. HubSpot repriced two of its Breeze agents in April 2026: its customer agent went from $1.00 per conversation to $0.50 per resolved conversation, and its prospecting agent from a monthly charge per contact enrolled to $1 per lead. No resolution, no charge.
02 — Why it matters
A seat is a licence to use a tool. A resolution is a piece of work.
When a vendor prices by the seat, it is selling you capability and leaving the productivity question with you. When it prices by the resolution, it is selling you output and taking on part of the delivery risk itself. That is a different commercial relationship, and it quietly transfers the hardest question back to you: what is this work worth to us at the unit level?
Most businesses cannot answer that. They know their software bill and they know their payroll. Very few know what one inbound enquiry costs to handle, what one quote costs to produce, or what one support resolution costs today. Without that number, a price of $0.50 or $2 is not cheap or expensive. It is unreadable.
There is a second consequence. Per-seat pricing is a fixed cost that improves as you grow — the same licence covers more work each month. Per-outcome pricing is linear. It scales with volume in both directions, which is protective in a quiet quarter and unforgiving in a good one. Businesses that model it at today’s volume and sign for three years are modelling the wrong year.
03 — The economics
Take an illustrative service business with 1,200 inbound enquiries a month, handled by coordinators alongside their other work. Assume an average of nine minutes per enquiry across triage, first reply and chasing. That is 180 hours a month.
At the published $2 per conversation, automating all 1,200 costs $2,400 a month. Divide by the 180 hours it addresses and you get roughly $13 per hour of released capacity. That is the number to compare with your own fully loaded hourly cost for that work — not with your current software bill. If your loaded cost is well above $13, the case is worth building. If it is below, you are paying a premium for consistency and out-of-hours coverage, which may still be worth it, but it is a different argument and should be made as one.
The two Salesforce models cross over at a specific point. At roughly $0.10 per action, twenty actions cost $2 — the same as one conversation. Below twenty actions, credits are cheaper; above, the flat conversation price is. So the deciding variable is how many steps your work actually takes, which is a property of your processes, not of the vendor’s. A well-defined process with a clean customer record resolves in fewer actions. A messy one burns credits looking for information.
Then there is the number that does the most damage in a business case: the residual. Vendor-reported customer results in Salesforce’s announcement include 50% of chat enquiries fully resolved at one company and 79% of conversations resolved autonomously at another. Treat those as vendor-reported and assume yours will be lower at first. Either way, the work that remains is not a random sample of what you had. The easy cases are the ones that automate, so the residual is the hard residual — longer handling times, more escalation, more judgement. Average handling time on the human queue goes up after automation. Plan the staffing on the residual mix, not on the old average, or the saving evaporates into overtime.
04 — What businesses should do
Establish your cost per unit before you take a meeting. Count the volume of the process for one month, estimate honest handling time including interruptions, multiply by a fully loaded hourly cost. You now have a comparison number. Without it every vendor quote is a coin toss.
Get the billable event defined in writing. When the price is per resolution, “resolution” is a commercial term, not a technical one. Who decides it was resolved — the model, the customer, a rule? What happens when the customer comes back on the same issue two days later? Is an escalation to a human free, as HubSpot’s model implies, or billed as a conversation? Answer these before signing, not in the first invoice dispute.
Model three volumes, not one. Today’s, three times today’s, and your worst peak month. Outcome pricing punishes success. If the three-times figure is uncomfortable, negotiate a ceiling now while you are still a prospect.
Do not pre-purchase credits until you know your action count. Run a metered pilot on one process long enough to learn how many actions your work consumes. That figure is specific to your data and your process, and nobody can tell you it in advance.
Start where volume and repetition are highest, not where the demo was best. If you handle 1,200 enquiries a month, the enquiry queue is the candidate. If you handle 60, it is not — the arithmetic of automating follow-up needs volume before it works at all.
05 — Where technology fits
For most businesses reading this, the honest answer is that an agent platform is not the next purchase.
Salesforce’s own announcements make the point better than any sceptic could. AIforce exists to carry data, workflows, permissions and business logic to the model, because the agent is only as good as the context it can read. If your customer record is split between an inbox, a spreadsheet, a quoting tool and someone’s memory, buying an agent buys you a confident guesser — and, under per-action pricing, a confident guesser that bills you for the search. That is the familiar cost of disconnected systems with a meter attached.
The work that makes any of this viable is unglamorous and comes first: one place where the customer record lives, a defined process with a named owner, and deterministic automation for the routing, reminders and handovers that do not need judgement at all. That is a connected system, and it pays for itself whether or not you ever deploy an agent. It is also the only condition under which per-outcome pricing works in your favour, because clean context resolves in fewer steps and fewer steps cost less.
If you are under a few hundred enquiries a month, you probably do not need an AI agent for this yet. You need to know what the work costs and where it currently leaks — and where AI actually creates leverage is inside a workflow that already exists.
06 — RevenueStack’s take
Our view: the agents are not the development. The pricing is.
A vendor only prices by the outcome when it believes it can predict the outcome well enough to carry the risk. That makes the pricing page the most honest capability map in the industry. Work a vendor will sell you per resolution is work the model reliably completes. Work still sold per seat — per user, per month, unlimited, your problem — is work it does not.
So read the price list before the keynote. When routine service conversations move to per-resolution billing and complex sales work stays on a per-user licence, that gap is telling you exactly which parts of your operation are now commodity labour and which still depend on your people. That is a more useful planning input than any benchmark, and it is published.
Sources
- Salesforce Expands Agentforce With a New Portfolio of AI Agents Built for High-Value Work — Salesforce, 2026-09-11
- Salesforce Unveils AIforce, Bringing the Full Power of Its Platform to Any Interface — Salesforce, 2026-09-15
- Agentforce Pricing — Salesforce, 2026-09-21
- Salesforce and Anthropic Announce Claudeforce: The #1 AI Meets the #1 AI CRM — Salesforce, 2026-08-26
- HubSpot moves to outcome-based pricing for some Breeze AI agents — MarTech, 2026-04-02
