AI Is Now a Vendor Risk

Somewhere in a defense contractor office this week, a normal knowledge worker did a normal thing. They opened their browser, reached for the AI tool they have been using to summarize documents and draft emails, and discovered it was suddenly off limits.

Reuters reported that the Pentagon ordered contractors to purge Anthropic's AI tools from their supply chains. That phrase sounds like an IT audit until you picture the real impact. Work stops mid-sentence. Teams scramble. Policies get rewritten. People do the same job, slower, with different tools, and a lot more anxiety.

If you think this is a niche government story, you are missing the point. The mechanism matters more than the headlines. When an AI tool becomes a vendor that touches sensitive data, it becomes a procurement decision. Procurement decisions can flip your workflow overnight.

Here is the quiet shift happening under your feet. For the last two years, most teams treated AI like a personal productivity upgrade. A tab. A shortcut. A private assistant. That mental model fails the moment your organization has to answer three questions. Where does the data go. Who can access the outputs. What happens if the vendor becomes unacceptable.

This is why you should care even if you do not sell to the government. The story is about precedent. A large buyer sets a rule. The rule becomes a template. Then your customer asks you for the same controls. Then your security team asks why you cannot produce an audit trail. Then your procurement team asks why your vendor list includes a tool your staff signed up for with a credit card.

The turn is that AI is graduating from tool to dependency. Dependencies come with obligations. You need continuity planning. You need alternatives. You need the ability to swap a provider without rewriting your entire way of working.

This is not about picking the right model. It is about building an AI stack that can survive a sudden no. No, you cannot use that vendor. No, you cannot send that data. No, you cannot ship that feature in that jurisdiction. No, you cannot keep the same workflow because the rules changed.

So what does a resilient setup look like in a normal company. It starts with something unglamorous. You map your AI touchpoints. Where do people use AI for drafting. Where do they use it for summarizing. Where does it touch customer data. Where does it touch legal documents. Where does it touch product decisions.

Then you separate low-risk work from high-risk work. Low-risk is internal drafting with no sensitive inputs. High-risk is anything that includes customer data, contracts, financial information, or regulated content. Treat those two categories like different systems. Because they are.

For high-risk work, you need receipts. What input was provided. What tool was used. What output was produced. Who approved it. Where it went next. This is not bureaucracy for its own sake. It is what makes it possible to answer a compliance question in one hour instead of one month.

You also need an exit ramp. If a tool disappears tomorrow, what is your fallback. A second provider. A self-hosted model for specific tasks. A set of templates that let humans do the work without the model. The goal is not to be paranoid. The goal is to avoid being fragile.

There is an easy way to pressure-test your setup this week. Pick one workflow that uses AI and pretend the vendor is banned tomorrow. What breaks. What cannot be reproduced. What requires access you do not have. What decisions are trapped in a chat window that nobody can export.

If that exercise makes you uncomfortable, good. Discomfort is useful when it turns into design. Build the smallest guardrail that would have made the scenario survivable. A policy. A checklist. A standard prompt library. A second tool approved by security. A shared place to store outputs with context.

The headline is not the Pentagon. The headline is the shape of the future. AI is moving into the same category as payments, identity, and email. Once it is infrastructure, it stops being optional. Your job is to build a workflow that still works when the world adds constraints.

Do not wait for a ban to learn this lesson. Make your AI stack swappable. Make your outputs auditable. Make your team faster even when the rules change. That is what separates leverage from dependence.