Let’s talk about the week two of the biggest buyers in enterprise software made opposite bets on the exact same question — and why the answer that’s right for each of them is a warning for you.
Start with the scene, because the scene is where you’ll recognise yourself.
A sales floor, Tuesday morning. Twenty reps open the CRM they’ve complained about for years. Instead of clicking through a pipeline, they type a sentence, and the answer comes back. No training deck. No forty-tab cockpit. It genuinely feels better.
Picture a person who has spent a decade learning where everything lives in that screen — and then one morning the screen learns to talk back.
That moment feels like an upgrade. It’s also the moment the software stopped being a tool and became a room.
Now the two bets.
At the end of August, Salesforce and Anthropic launched Claudeforce. Not a plugin. Not a connector. One model wired in as the default reasoning engine at the centre of the product, with the pilot running in August, open beta in September, and production waves landing through October. Salesforce plans to spend roughly three hundred million dollars on Anthropic tokens in 2026 — sitting on top of a three-hundred-million-dollar equity stake it already holds.
Follow that structure for a second. The company buying the model is also part owner of the company selling it.
Five days earlier, on the thirty-first of August, the Pentagon went the other way. It added ChatGPT Mil and Grok for Government to GenAI.mil — the portal its three million military and civilian people use, where Google’s Gemini was already sitting. Around 1.7 million people are already on the platform. The department said out loud that pairing models side by side was a deliberate hedge against vendor lock-in.
Four models. One door. On purpose.
Same question. Opposite answers. And both are rational, because the two buyers carry completely different costs.
Here’s the pattern worth detecting.
The lock-in fight is over. Not because anyone won it — because it moved. Out of the model layer, out of the pricing layer, into the interface. The front door.
Model prices collapsed. Open weights closed the cost gap. Every serious platform now advertises model choice, because choice is cheap when the model is a commodity sitting in the middle of the stack. So the dependence relocated to the last layer nobody wanted to audit: the place where a human forms intent.
And that layer is different, because the switching cost there is not data. Data export is a technical problem. You can pay engineers to solve a technical problem.
The front door is a habit problem. When the model becomes the interface, the cost of leaving stops being measured in migration scripts and starts being measured in people — every rep, every onboarding doc, every workflow that only exists as a sentence somebody types from memory. You can’t buy your way out of that. You have to retrain it, out of humans, at human speed.
That’s the moat now. Not the weights. The muscle memory.
And notice the asymmetry in who still holds the old posture. “Model-agnostic” was never a principle. It was a negotiating position. It stayed cheap to hold for exactly as long as the model sat in the middle. The buyers who bear the switching cost are still holding it. The vendor who collects it just put one model at the centre of the product and signed a nine-figure token commitment to keep it there.
Which side of that ledger you sit on decides your doctrine. That’s the whole RADAR.
Now, what people are missing.
First: a choice is not a control.
The Pentagon did the hard, correct thing. It designed out dependence before it mattered, and it has the receipts — four models, one portal, 1.7 million users. And that menu was still a decision somebody made. Claude is absent, after the company insisted on terms barring mass surveillance of Americans and lethal autonomous weapons that the department wouldn’t accept. And Grok went live over an internal finding that its own engineers couldn’t reliably fix its worst failure mode.
A longer menu is not a standard. Optionality stops being a shield the moment you have no floor of your own — you just get more vendors to be disappointed by. Choice without standards is a longer list of the same risk.
Second: the three hundred plus three hundred.
When your software vendor is also a shareholder in the model lab, the roadmap, the pricing and the renewal get set at one table you’re not at. Nobody has to be malicious about it. The incentives simply line up in a direction, and the direction is not yours.
So here’s the operator lesson. Three moves, and they’re cheap this week.
One. Write your workflow down in plain words you own. Not a prompt sitting inside a vendor’s product. If the definition of how your business runs only exists inside the thing you’re renting, you don’t own the process — you own the subscription. Prose travels between tools. JSON exports from a dead platform don’t.
Two. At renewal, ask for two numbers. Token cost is the one they’ll give you. Ask for the retraining cost — if swapping the model tomorrow means two hundred people relearn their day, that’s the actual renewal price, and it lives nowhere on the invoice.
Three. Own the door. The layer where intent gets formed should be something you can point at. Your interface, your data layer, your descriptions. One test question: if this vendor changed the model tomorrow, would my business notice? If the answer is no, you’ve bought a tool. If the answer is that the sales floor stops, you’ve bought a landlord.
The front door is where sovereignty goes. Quietly, one habit at a time. It’s the last place you notice, because it feels like convenience — nobody ever signed their judgment away. They just started typing a sentence instead of clicking five buttons. Six months later, the sentence was the business.
Gartner reckons about forty percent of enterprise apps will carry a task-specific agent by the end of the year. Every one of those is somebody deciding where your front door is going to be.
The model will change again next quarter. It always does. The only question that matters is whether that lands as an upgrade — or as a migration somebody else scheduled for you.