The US Department of Defence told the BBC on Monday that it has stopped using Anthropic’s AI products. The order to phase them out came in late February with a six-month deadline, and people who’ve worked with the department told the BBC that Anthropic’s Claude model was still in use as recently as last week.
For the agencies we work with, the interesting part is why it took so long.
What happened between February and now
Earlier this year the Pentagon asked Anthropic to lift its restrictions on how the military could use Claude. Anthropic dropped most of them but held firm on two, covering mass domestic surveillance and autonomous weapons. Defence Secretary Pete Hegseth’s position was that the military couldn’t let a private company dictate how it uses the tools it buys, and in February he designated Anthropic a national security supply chain risk, a label previously used against foreign firms.
Anthropic sued. In late August, a federal judge in California ruled the designation unlawful, while noting that the department is free to choose whichever AI vendor it likes. The government can appeal, and a separate case is still running in Washington DC. In the meantime, the Pentagon has signed AI contracts with Google, OpenAI and xAI.
Why removing it took longer than planned
At the department, Claude sat inside Maven Smart System, the Palantir-operated platform the Pentagon uses to organise intelligence and other data. Taking it out meant working through somebody else’s product, on somebody else’s schedule.
Lauren Kahn, a senior research analyst at Georgetown’s Centre for Security and Emerging Technology and a former US defence official, summed it up for the BBC: “Once they become integrated, it can be painful to remove them.”
What this looks like inside an agency
For most of the agencies we look after, AI arrived inside software they already pay for: Microsoft 365, Google Workspace, Adobe, Canva, Notion, Slack, the CRM. Each of those features runs on a model from one AI company or another, and the vendor can change which one without asking you.
Then there’s what staff have added themselves. When we map AI use with a client, the real list is nearly always longer than the approved one, usually because someone signed up for a free trial during a busy pitch week and kept using it.
Picture a client asking you to stop using a particular AI model on their account, or a provider changing its terms in a way your legal team doesn’t like. You’d want to know within a day where that model touches your work. For a lot of agencies, that would mean a week of emails, plus some guesswork about the free accounts.
What we’d do
None of this needs a big programme. We’d start with four things:
- List every tool with an AI feature switched on, including the ones built into software you already use, and note which AI provider sits behind each where the vendor says.
- Find each vendor’s sub-processor list. Most business tools publish one, and it shows who handles your data and how you’ll hear about changes.
- Keep your own working material, such as prompts, brand voice notes and templates, in documents you own, so moving to another tool is mostly a case of copying and pasting.
- Write down which tools are approved for client work and who decides when that changes.
If that list already looks longer than you’d like, we’re happy to work through it with you as part of an IT strategy review.
In Dr Logic’s experience, a couple of hours with the right people in the room gets most agencies a clear picture. That’s a good deal cheaper than discovering mid-pitch that the AI behind your favourite tool has quietly changed.
