Nobody Photographs the Junction Box

On 26 August, Salesforce and Anthropic announced a partnership called Claudeforce. The first product is a plugin with 37 prebuilt sales skills that lets a rep work their pipeline without opening the CRM. Pilot customers now, open beta in September, sales skills first.

Most of the coverage read it as an AI story. Benioff said the UI is the AI. VentureBeat ran a headline about Salesforce telling the world nobody needs its app anymore. Fair enough, that is the dramatic part.

I keep going back to a much less dramatic sentence, which is that every action runs through your existing permissions and business rules. Nothing new to stand up. Nothing to re-audit.

That is the whole thing. That is what got sold.

Because if you strip the announcement down, the model is not the scarce ingredient. Reasoning is available to anyone with an API key and a credit card. What is not available for rent is a system that already knows which of your reps can see which accounts, which stage transitions are legal, which discount needs approval, and what your company means by the word "qualified." Salesforce spent 27 years accumulating that. It cannot be conjured, and a startup cannot rebuild it in a funding round.

The moat was never the Lightning pages. It was the rules.

The layer nobody puts on a slide

I have spent about ten years doing governance work, and I want to be precise about how it has been valued for most of that time.

Naming conventions. Permission sets. Field deprecation. Campaign hierarchy. Lead status definitions that mean the same thing in two systems. Consent state that agrees across the CRM and the MAP. Sync monitoring that tells you when the integration user started silently failing on a subset of records.

This work has a reputation problem. It does not demo. There is no slide where you reveal a beautifully deprecated custom field and the room applauds. In most organisations it is filed under overhead, which is the polite word for a thing you fund last and cut first.

It is the electrical rough-in behind the drywall. Nobody has ever taken a photograph of a junction box. Nobody has ever said "come and look at what we did with the conduit." But you cannot hang anything heavy on a wall that does not have one, and everybody finds out which walls have one on the day they try.

Well. The industry just tried to hang something quite heavy.

Bad rules do not get better under load

Here is the part that makes me nervous about how this lands.

Governance being the feature is only good news if your governance is good. Inheriting your permission model is a genuinely elegant design choice, and it inherits whatever is actually there. If your access model is six years of sediment, that is what gets inherited. If half your picklist values are abandoned, those are the values the agent reasons over.

An agent on top of bad rules is cruise control on a car with the alignment out. Cruise control does not correct the pull to the left. It just commits to it, at speed, with more confidence than you had.

The rules do not have to be sophisticated. They have to be true. A three-stage pipeline that everyone actually follows beats an eleven-stage pipeline that half the team routes around, and it beats it by a wider margin the moment something is reading it literally.

Most companies have compensation, not governance

Here is the distinction I would like people to sit with.

Every functioning revenue org runs on a quiet layer of human correction. The rep who sees a close date three months in the past and knows to ignore it. The ops person who knows that the Amsterdam team logs activity in a different tool, so their deal health always looks worse. The manager who knows that the two accounts with nearly identical names are one company, and knows which of them is the real one.

None of that is written down. None of it is in the system. It lives in the heads of people who have been there long enough, and it works, in the sense that the org keeps functioning.

That is compensation. It is not governance. The difference did not matter much while the only consumers of your data were humans who came pre-loaded with the corrections.

It matters now, because the new consumer has no corrections. It has your fields.

I want to be careful not to oversell the peril here. This is a sales plugin in pilot, not a phase change in how every company operates, and I have been in this industry long enough to have watched a few inevitabilities not happen. But the direction is not really in question, and the readiness work is the same work whether the agent arrives in September or in two years.

What I would actually do

Nothing exotic. The unglamorous audit you have been postponing, with a date attached now.

Find out which fields on your core objects are genuinely maintained and which are populated by a workflow whose author has left the company. Pull every permission set and profile and ask who granted it and whether the reason still exists. Check that your CRM and your MAP agree about the same contact, and check it by sampling records rather than by asking the integration whether it feels fine. Write down the corrections your experienced people make in their heads, because that document is the actual spec for your business rules and it currently does not exist.

Then decide which of those corrections should become rules and which should become deletions.

Marketing skills are on the roadmap, not in the product, which means marketing ops teams have a few months where they can do this without an agent narrating their gaps to a CRO in real time. That is a better position than sales is in, and it will not last.

The bit I did not expect to enjoy

For ten years the honest pitch for this work was that it prevents things from going wrong later, which is a terrible pitch. Nobody has ever been promoted for a disaster that did not occur.

The pitch changed last week. Your governance layer is now the thing that determines whether you can use any of this at all. It stopped being insurance and started being capability.

I am aware that I am a governance person telling you that governance is now the most important thing, which is exactly what a governance person would say either way. Judge the argument, not the source. But I have been writing naming convention documents that no human has ever voluntarily read, including, on at least two occasions, me, and it is a strange feeling to watch that category of work get named as the moat by the largest CRM vendor in the world.

I will take it.