Picking a model was never the hard bit. That decision gets made, and then revisited when the next release lands. A decision you keep remaking isn't the one your strategy compounds on. What compounds is the harness around the model, and how that model plugs into where your people already do the work and where you hold your organisational knowledge and context. Salesforce understood that early, which is why Alforce is built as a layer around the model. 

On 15 September at Dreamforce, Salesforce answered it. AIforce launched as an interface layer that exposes Salesforce data, workflows, business logic, permissions, and security to AI environments outside Salesforce, with zero data retention. It has four parts: Claudeforce, Slackforce, Agentforce Coworker, and a headless toolkit of MCP servers, APIs, and skills. 

Claude sits underneath a lot of that. It is the reasoning model behind the Atlas Reasoning Engine, the default model in Slack, available in Agent Builder, and deployable inside the Salesforce Trust Boundary on Amazon Bedrock. 

So the answer to “which model” is now the one your CRM vendor chose, running inside your CRM vendor's trust boundary, under your own credentials.

Scott White, head of enterprise product at Anthropic, described the permission model plainly:

That is a genuine change to the conversation, and it is also the smallest part of the problem.

The Australian Bureau of Statistics puts business AI use at 12% for the 2024-25 year, rising to 35% among large businesses. Financial and insurance services sit at 24%, up from 1% in 2021-22. A 24-fold move in three years, in the sector with the most governance overhead in the country. 

The model releases in that window do not explain it on their own. In the organisations I have worked with, the thing that moved was never a better model. It was somebody finally producing a governance answer that survived a risk committee. 

Australia has no AI-specific statute of the kind the 2024 mandatory guardrails proposal contemplated. That proposal was not proceeded with, and the National AI Plan of December 2025 relies on existing technology-neutral law instead. So there was never a compliance deadline holding these programmes up. There was a board question about a model nobody could point at, sitting in a risk register nobody would sign. 

Claudeforce gives that question an answer with a vendor's name on it. If your AI programme has been parked on model governance, the parking space is closing. 

The model was never the hard part. It was the excuse. 

If this is starting to sound familiar, it should. I've made this same case for a while now, the model is roughly 20% of the work, and the other 80 is the delivery discipline behind it. Claudeforce does not change that argument. It raises the stakes on it, because the vendor has just made the easy 20% easier still, which makes the temptation to skip the hard 80 stronger, not weaker. 

The Claudeforce beta includes 37 pre-built sales skills covering meeting preparation, deal health, and pipeline review. They are good. They will also not fit your process, because they were not built for it. 

Pre-built skills are a delivery pattern, not a product. They get an organisation to a demo quickly, and the demo is roughly 20% of the job. The other 80% is the part that does not demo well: context, data and integration, evaluation, observability, governance, change management, and the boring discipline of keeping an agent in line after the applause stops. 

An agent is a model plus a harness. The model is the intelligence. The harness is everything else. Salesforce has now made a default model choice for a large part of its customer base, which means more of the difference between the organisations that get value and the ones that do not now sits in the harness. That is a better problem than the one we had, and a harder one. 

The honest test of any of this is whether we use it ourselves. At J4RVIS we call this being Customer Zero: nothing goes to a client until it has run on us first. 

The Salesforce connector for Claude that existed before Claudeforce did not meet our own requirements, so rather than wait for Claudeforce to arrive, we built a custom Salesforce MCP connector for our org and rolled it out to the sales team. Our sellers now ask questions in plain language and get answers from live CRM data, under their own Salesforce credentials, with their own sharing rules applied. No bulk export, no shared service account, no new tool for anyone to learn. 

It took our team a couple of hours to make it work, and a couple of days for our sales team to embed it deeply into their sales processes. That gap is the point. Standing up the connector, the 20%, is always fast. The part that matters is making it fit how we actually work: our processes, our AI standards, our business priorities. That is why we built it integrated into Salesforce and Slack, where our people already do the work, rather than as a separate tool bolted on the side. It is also what we tell clients when they ask what this takes: not how fast the connector went in, but how deliberately we built the 80% around it. 

Being ahead for us at J4RVIS meant having a Claude and Salesforce integrated Sales Command Centre live and running before those examples were showcased at the latest Dreamforce. That kind of speed of experimentation is what we like our clients to strive for. Building it on the Salesforce ecosystem for our internal teams means it can sit inside our sellers' Claude apps, inside our own custom-built sales agents, or headless in any interface of our choice. 

We were Customer Zero so that when a client asks what this actually takes, the answer comes from experience rather than a product page.

Stop running the model evaluation. It is finished, and continuing it has become a way of not starting. 

Then pick one business outcome where the reasoning matters and the data is already clean. Not the hardest one, and not the most visible one. One where you can measure the before and after honestly. 

Then budget for the harness. If the business case only covers getting an agent live, the programme has funded 20% of itself and will stall at exactly the point it starts to matter. 

Claudeforce is in beta today, and the space will keep moving fast. If you want to match the speed of innovation with the speed of your own decision making, that is where we can help you find the best path forward.