Insight / signal
Your AI stack is now a supply-chain risk
The most useful AI story this week is not another model leaderboard.
The most useful AI story this week is not another model leaderboard.
It is access.
Who gets it. Who loses it. What breaks when it changes. What happens when the workflow you quietly started relying on sits on top of one provider, one login, one policy decision, one cloud account, one region, one price change, one terms-of-service update.
That sounds dull compared with the usual AI noise. Good. Dull is where the real business risk tends to live.
A recent episode framed the problem neatly: one government letter and a powerful Anthropic model was suddenly off the table. The episode’s answer was not “panic.” It was more useful than that. Learn local runtimes, understand quantisation, know which model sizes fit your hardware, build routing, and stop assuming the hosted frontier model will always be there when you need it.
TechCrunch is now reporting the same broad issue as a live market story, with Anthropic’s access disruption sparking a debate about India’s AI future. Different angle, same lesson.
AI access is not a neutral utility yet. It is political, commercial, technical, and fragile.
That matters because businesses are no longer using AI for party tricks.
A couple of years ago, if ChatGPT went down, you lost a drafting tool. Annoying, survivable. Now companies are wiring AI into sales follow-up, customer support, content production, reporting, coding, research, proposal writing, internal search, analytics, and lead triage.
The dependency has moved closer to the actual work.
So the question changes.
It is no longer: “Which model is best?”
It is: “What happens to the business process if this model, account, API, agent platform or cloud environment stops behaving the way we expect?”
Most companies do not have a good answer. They have a favourite chatbot. Maybe a few prompts. Maybe a Zapier chain. Maybe someone has built a clever workflow that works brilliantly on a good day.
But if it dies when one vendor sneezes, it is not an AI operating system. It is a dependency with branding.
This is where the local AI conversation gets interesting. Not because every business should rush off and pretend a small local model can replace the best frontier systems. Frontier models are still much better for hard reasoning, long-context analysis, and messy open-ended work.
But local models do not need to replace everything to be useful.
They need to cover the work that should not depend on someone else’s availability: summarising internal notes, classifying tickets, drafting first-pass responses, extracting structured data, running private analysis, keeping basic workflows alive when hosted models are unavailable or too expensive for the job.
That is not glamour. That is resilience.
The better frame is model routing.
Use the expensive frontier model when the job needs it. Use cheaper hosted models for routine work. Use local models for private, repetitive, or resilience-sensitive jobs. Keep a fallback path. Log what happened. Measure quality. Give the system a human owner.
The infrastructure direction is obvious if you look at where the serious money is going.
OpenAI’s planned acquisition of Ona is explicitly framed as expanding Codex with secure, persistent cloud environments for long-running AI agents across enterprise workflows. That is not “write me a LinkedIn post” territory. That is operational infrastructure. OpenAI is also scaling workplace courses around practical AI skills and repeatable workflows. BBVA has rolled ChatGPT Enterprise to 100,000 employees. The EU is moving on content transparency and provenance standards.
AI is becoming part of the operating layer of companies. Which means the operating layer has to be designed properly.
For a business owner, this does not mean building some ridiculous sci-fi command centre. It means asking better, more boring questions before you automate anything important.
What happens if the model is unavailable? What happens if the price jumps? What data is allowed into the system? Which tasks need frontier quality and which do not? Which workflows can run locally? Where does the human approve the output? Where are the logs? How do we know it worked? What is the fallback? Who owns the thing?
That last question is usually the one people dodge.
If nobody owns it, the AI workflow becomes another orphaned system in the business. It half works. Nobody trusts it fully. People work around it. Costs creep up. Prompts drift. Access breaks. The one person who understood the setup leaves. Then everyone decides AI is overhyped.
AI is not overhyped because a workflow breaks. The workflow was badly owned.
This is also where the agency opportunity gets clarified.
“We build AI agents” is already too broad to mean much.
The better offer is: we find the repeated work, clean the context, design the workflow, choose the right model mix, add approvals, measure the output, and make sure the business is not trapped inside one vendor’s walled garden.
Less flashy. Much closer to what companies actually need.
A simple example: a sales follow-up system.
The brittle version takes a lead, sends it to one hosted model, drafts a reply, pushes it straight into email, and hopes nothing stupid happens.
The grown-up version is different. It checks CRM context. It classifies urgency. It uses a cheaper model for simple tagging. It uses a stronger model for high-value drafting. It keeps private notes out of external models where needed. It asks for approval before sending. It logs the source context. It falls back if the main model is down. It measures reply quality over time.
Same basic promise. Completely different risk profile.
I think this is where a lot of AI adoption splits.
Some companies will keep collecting tools. Twelve subscriptions, three demos, a shared prompt doc, and no operating model.
Others will build a thinner but stronger layer: company context, model routing, permission boundaries, approval points, measurement, local capability where it makes sense, and clear ownership.
The second group will not look as exciting on LinkedIn.
They will just get more useful work done with less drama.
That is probably the better trade.
Because the model is rented.
The operating layer is yours.
And if AI is going to sit anywhere near real business work, that distinction matters more than most people want to admit.
Jason Sibley is the founder of Cleo, a post-agency marketing and AI company. JasonVsTheNoise is where he writes about what is actually happening with AI, marketing, and how businesses should be thinking about both.