Enterprise AI Is Learning To Charge For Work, And

What happened

Enterprise AI pricing is shifting from tokens, seats and usage towards successful tasks and outcomes, as vendors increasingly charge for the work their agents complete. As agents become embedded in enterprise workflows, the companies controlling the agent, workflow and outcome could gain greater pricing power, potentially moving from charging for AI usage to claiming a share of the value created.

One account of the launch put it plainly, that customers may never need the Salesforce app UI again. 5 Bn, up more than 240% year on year, and its CEO using the earnings stage to say the talk of a software apocalypse should stop. What founders should take from this is how durable billing moves now.

The harness converts a customer’s accumulated institutional knowledge into portable skills that travel across the enterprise, and that encoded knowledge, as valuable as the model weights, is what becomes hard to replace. Even the way today’s agents remember, through scribbled notes rather than elegant databases, makes the point, since it is that messy accumulated context that carries institutional memory forward.

The fantasy of one harness that holds the whole company runs straight into a problem four decades old, because centralising an organisation’s knowledge in a single place has never worked at enterprise scale. Each great system of record conquered one domain and stopped there. Salesforce owned sales, SAP owned the procure to build cycle, Workday owned HR, and none of them became the company’s apex brain.

The agents inherit exactly those boundaries. Claudeforce is the agentic expression of the CRM, not of the enterprise, and the same will most likely be true of every system of record that ships or partners for its own agent. The reason the boundary holds is organisational as much as technical.

Companies ship their org charts, so no one owns the process that runs end to end, and AI gets bolted onto each silo instead of rewiring the workflows that cross them.

Why it matters

Salesforce, Sierra and others are already experimenting with outcome-linked pricing, while Indian AI companies have an opportunity to build model-agnostic, domain-specific AI stacks that run on customer-controlled infrastructure. The most consequential change in enterprise AI this year is not a model release, it is a change in what buyers agree to pay for.

OpenAI’s CFO has recast the buyer’s question away from cost per token and towards cost per successful task, proposing “useful intelligence per dollar” as the scorecard and arguing that AI should be measured by work accomplished rather than usage.

By August 2026, the same leadership was telling investors that the age of “tokenmaxxing” had passed, as enterprises routed routine work to the cheapest capable model and reserved frontier models for the tasks that justified them. The deeper reason this shift was inevitable is that effort and value were never tightly coupled.

Industry observers building agentic products point out that an agent which closes a tenth of the deals while consuming a hundredth of the tokens is still the one worth paying for. Token count measures exhaust, not distance travelled. Sierra, the customer service agent company, already prices on this basis, charging a set rate for a resolved conversation and nothing when the case escalates to a human.

Salesforce has begun reporting Agentic Work Units in the billions, a unit defined by work performed rather than seats occupied. Value based pricing is a seller’s construct as much as a buyer’s, and analysts at Constellation Research have noted it has historically served vendors more than customers. The question is more than whether the meter is moving, it is whether the outcome lands when it does.

Once work is priced by outcome, the product stops being the screen and becomes the agent that produces the outcome, and the interface starts to disappear behind it.

AI founders describe the valuable object not as the API, the actions initiated when a button is clicked, but as the harness around it, the skills, documentation and rules that encode how the best expert runs workflows on and extracts value from a system. On this reading, the application of the future is not the form and the dashboard, it is that harness.

Claudeforce is that argument arriving as a product. Announced on 26 August 2026, it makes Anthropic’s Claude the default reasoning engine across Salesforce and brings the platform’s data, actions and governed workflows into Claude through a layer Salesforce itself calls a harness, AIforce, with 37 prebuilt sales skills. A seller can run a pipeline review or prepare a deal without opening Salesforce at all.

What to watch

The universal enterprise harness therefore waits on a reorganisation that most companies have not attempted, which is why the near term winner is not the universal agent but the team that owns the harness and the outcome unit inside one bounded, high value domain. That is a narrower and more winnable problem, and it is where new entrants take share before incumbents finish reorganising around agents.

The counter case deserves a hearing. Frontier agents can already reason across disconnected systems, plucking data from many places without a unifying database, which argues against hard boundaries.