Claudeforce is a genuinely good product.
That is what makes the bill hard to see.

What Claudeforce actually is

On 26 August 2026 Salesforce and Anthropic announced Claudeforce. Two of the things shipping under that name matter here. Claude becomes a reasoning model for Salesforce’s Atlas engine, and the default behind Agentforce Vibes and Agentforce Coworker. And a plugin called Salesforce in Claude puts the CRM inside the assistant, with 37 prebuilt sales skills covering meeting preparation, deal health and pipeline work.^[R1]

It is worth saying plainly, before anything else: this is good engineering, and the part that is easiest to skip past is the part that is best. The plugin does not hand a model a copy of your database. It calls a server that returns only what the person asking already has permission to see, and applies the same business rules before anything is written back.^[R1] Permission-scoped tool access is the right architecture. A great many products shipping this year did the other thing.

The distribution logic is just as sound. One administrator connects the org once, and everybody inherits it at their own permission level. For a company already standardised on Salesforce, that is a genuinely large saving over rolling out yet another interface.

Salesforce also appended its forcesuffix to another company’s product for the first time.^[R1] That is not a detail. Naming conventions are how a platform tells you what it considers part of itself.

You now pay on three meters

Before agents, buying a CRM meant one meter. Seats, times months. It was expensive and it was legible: you could tell a partner what next year would cost without knowing how busy next year would be. An agentic layer adds two more meters, and only one of the three still has that property.

Meter one is the seat licence. Unchanged, predictable, and the one everybody budgets for. It buys the system of record.

Meter two is consumption. The agent layer is metered by usage rather than by headcount, and the unit is derived from tokens. This is where the trouble starts, and it is not that the rate is high. It is that a token is not a business event. It is a property of how a model reads and writes. Two client files that take exactly the same amount of your attention can differ several-fold in what they cost to process, because one client writes long emails and attaches scanned documents and the other does not. Nothing in your business explains the variance. Nothing in your business can be adjusted to reduce it.

Meter three is configuration. Somebody has to decide which skills are turned on, what they may write back, which objects they may touch and what happens when they are unsure. That work is real, it is skilled, and it is billed by the hour or carried as a salary. It is also the meter that appears in no pricing page, which is why it appears in no business case either.

A firm of two hundred with an operations team absorbs all three comfortably. It has the admin already, it has a procurement function that models consumption, and a surprise in one quarter is an annoyance. A firm of twelve has none of that. The same three meters land on one person who also has clients.

Gartner’s cancellation forecast for agentic projects named the causes in this order: escalating cost, unclear value, inadequate risk controls.^[R6] Cost is first, and the reason is usually not the rate card. It is that the meter counts something the buyer cannot forecast.

The specialist does not go away

The most quietly optimistic reading of an agentic layer is that it removes the person in the middle. Describe what you want in plain language, and the configuration problem dissolves. It does not dissolve. It moves.

Consider what is actually true about the platform this runs on. There is an entire profession called Salesforce administrator. It has a certification track, a job title, a salary survey and a global consulting industry attached to it. That is not a criticism of Salesforce. It is the observable consequence of building something configurable enough to model almost any business, and firms that need that depth are right to pay for it.

But a system that rewards configuration that heavily does not stop rewarding it when you add an agent. An agent is a new configuration surface, not a replacement for the old one. Somebody still defines the scope. Somebody still decides what an agent may write back to a record that other processes depend on. Somebody still answers the question of what it does when the file is ambiguous, and that question is harder than any field mapping, because a wrong answer reaches a client.

For a firm that employs one, this is a saving. For a firm that does not, it is a hire, or a retainer, and it is the line that turns a plausible monthly number into a project. The gap between those two firms is the thing a pricing page cannot show you, and it is much wider than the gap between any two rate cards.

Who sets the rate

Here is the part where we have to declare an interest, so we will do it first and plainly. Invice runs on frontier models, Claude among them. We are not about to argue that Anthropic is the problem, because we would be arguing against our own stack. The argument is about a layer, not a vendor.

Salesforce told investors it expects to spend in the region of three hundred million dollars on Anthropic tokens this year, and it holds an equity stake in Anthropic reported at multi-billion scale.^[R3] None of that is a scandal. It is a disclosed supply chain, and disclosure is the good version of this.

It matters to a buyer for one structural reason. When the intelligence is bundled into the system of record, you cannot change either one independently. The model stops being a decision you make each year against price and quality, and becomes a term of your CRM contract. If inference gets dramatically cheaper, you find out what share of that reaches you when your renewal arrives. If a different model turns out to be better at reading your particular documents, switching to it is not a setting. It is a migration.

This is the same lock-in argument the industry has had about databases and clouds for thirty years, arriving one layer up. The reason it is worth restating now is that the layer is new enough that most buyers have not yet decided whether they want to own it. Once it is bundled, that decision has been made on their behalf.

It is worth noticing what Claudeforce is built on, because it cuts the other way. The plugin speaks Model Context Protocol, which is an open specification governed under a Linux Foundation directed fund rather than a private interface.^[R5] The connective tissue is deliberately not proprietary. The commercial arrangement sitting on top of it is a separate choice, and it is the one a buyer signs.

What we do instead

We are not a neutral party, so this section states controls rather than conclusions. Every item below is something you can check, and none of it is a claim about being safer. That word is yours to reach, or not.

The model key is yours. You connect your own provider account, and the AI agent runs on it. Six providers are configured in production. Provider usage is billed to you, by them, at their rate, with nothing added. If a better or cheaper model appears, changing to it is a setting rather than a renegotiation. The honest caveat travels with the claim: if you connect a United States provider, your prompts are processed in the United States. Bringing your own key gives you control over the choice. It does not relocate the inference.

You keep the record you already have. Connecting the CRM you run is the default posture, not a migration path we tolerate. Invice is not built to be bought instead of Salesforce, and a firm standardised on it should stay there. If you have no CRM worth keeping, you can grow into Invice as your primary client-management platform, but that is an opt-in, not the pitch.

Tools are approved one at a time. A connected tool is not admitted because the AI agent asked convincingly, and not because a session happens to be attended. Only an unexpired approval grant crosses the connector boundary. Untrusted content, which means an inbound email or an uploaded document, never carries authority to invoke a tool. That is enforced in code rather than requested in a prompt, because a model instructed to ignore instructions is not a control.

Consequential actions wait for a person. Approval is the default on the send path. The interesting question to ask any vendor is not whether approval is available, since it always is. It is what happens when nobody configures anything.

Your data is not our training corpus. Client data does not train a shared Invice model. There is no analytics or error-tracking SDK wired into the product. Primary records and vault files are configured in a Canadian region.

What we will not tell you is what any of this makes you. Whether these controls meet the obligations you actually carry is a conclusion about your practice, not a property of our software, and a vendor who reaches it on your behalf has told you something about their sales process rather than about their architecture.

Where Claudeforce is right and we are wrong

If you cannot find the paragraph where a vendor loses, you were right not to trust the rest of the page. Five of the six rows below point somewhere other than us, and the first two point straight at the product this piece is arguing with.

The short version: the more of your business already lives inside Salesforce, the better Claudeforce gets and the worse we look. Governance, permissions and audit trails are expensive to build and you have already built them. An agent that runs where they already are inherits all of it. Nothing outside the platform can match that, and we are not going to pretend otherwise.

Our case is the opposite shape. It is for a firm whose client work is spread across tools that were never going to consolidate, where the record is one of five systems rather than the centre of gravity, and where nobody is employed to configure software.

Six questions for any vendor bundling a model

Our field guide on agentic CRMs ended with a checklist, and one line in it has aged into the whole of this article: where does the model cost land, and who sets the rate? Claudeforce is that question becoming concrete. Here it is broken into the six you can actually ask on a call.

• Whose key is it? If the answer is theirs, you are buying intelligence at a margin someone else sets. That can still be the right purchase. You should know you made it.
• What does the meter count, and is it a business event? Resolved cases, closed files and booked appointments are things you can forecast. Tokens are not. Ask them to restate the price in units your business already tracks, and notice how hard that is.
• What happens when volume doubles?Ask for the bill at twice today’s activity, in writing. A vendor who cannot produce it does not know either, and you will find out together.
• Who configures this, and do you employ them? Get the honest number of days for setup and for a typical change six months later. Then ask who does it, by name.
• Can you change the model without changing the platform? If the answer is no, those are one purchase wearing two price tags.
• What sends without you, by default? Ask about the default, never the capability. Everything can be configured to require approval. The question is what ships when nobody configures anything.

If the answers point you at Salesforce, go. It is a serious platform and Claudeforce is a serious addition to it. If your client work is the business and it lives across tools that were never going to become one tool, start with what this looks like on a real desk, or read the longer field guide to agentic CRMs that this piece follows from. We would rather be the right answer for fewer people than the plausible answer for everyone.