Insight / signal
Your AI problem is not the model. It is the mess inside the business.
There was a time, not very long ago, when the AI story was basically "look at the model".
There was a time, not very long ago, when the AI story was basically “look at the model”.
Bigger context window. Better coding. Better images. Better reasoning. New benchmark. New demo. New panic.
That phase is not over. But it is no longer the useful part of the conversation for most business owners.
The useful question now is much more awkward:
Does your business know enough about itself for AI to be useful?
Because if the answer is no, it does not matter which model you buy.
You can put GPT, Claude, Gemini, or whatever comes next on top of a messy company and you will still get messy output. Faster, prettier, more confident messy output, admittedly. But still messy.
Most companies do not have an AI problem. They have a memory problem.
Their customer knowledge is in old proposals, Slack threads, inboxes, Notion pages, call recordings, half-finished spreadsheets, random PDFs, sales decks, and the heads of three people who are too busy to explain anything twice.
Then they ask an AI tool to “write in our voice” or “improve lead quality” or “build a campaign”.
With what, exactly?
A few scraped web pages? A brand deck from 2022? A prompt that says “act like a senior strategist”?
That is not strategy. That is asking a very expensive autocomplete to guess your business.
The last couple of days of AI news have been interesting precisely because the signal is not a dramatic new model announcement. The signal is boring enterprise plumbing.
OpenAI announced Academy courses for practical AI skills, repeatable workflows, and agents in everyday work. It published a case study on BBVA scaling ChatGPT Enterprise to 100,000 employees. It announced plans to acquire Ona so Codex can run in secure, persistent cloud environments for long-running enterprise workflows. It pushed Oracle Cloud access, existing cloud commitments, security, and governance.
None of that is sexy in the way a model demo is sexy.
But it is probably more commercially important.
It says the market is moving from “try this chatbot” to “how do we put AI inside the operating model without making a complete mess of it?” That is a different game.
The podcast world is saying the same thing from the operator side. The real bottleneck is internal knowledge extraction. Not token spend. Not prompt polish. Not whether the model can write a decent paragraph. The bottleneck is turning company knowledge into usable context and then linking that context to measurable business output.
That sounds dull until you have watched a company try to use AI without it.
The sales team has one version of the customer. The delivery team has another. The founder has a sharper version in their head, but it has never been written down properly. Marketing has the polished version, which is often the least useful one because it has been sanded down until nobody can disagree with it. Customer support has the truth, but nobody reads it. Finance knows which customers are actually profitable, but that rarely makes it into campaign planning.
So the AI tool gets fed a thin public version of the company and everyone acts surprised when it produces thin public work.
This is where the next advantage is going to come from.
Not from having access to a model your competitors cannot use. Most of them will be on the same frontier models soon enough, or close enough that buyers cannot tell the difference.
The advantage will come from what you wrap around the model.
Your customer data. Your offers. Your objections. Your best sales calls. Your failed proposals. Your fulfilment notes. Your tone of voice when a real person is talking, not when the brand guidelines are trying to sound clever. Your proof. Your boundaries. Your compliance rules. Your weird internal knowledge that makes the business actually work.
That is company memory.
Not “upload a few PDFs”. Not a folder full of stale docs. Real company memory means a maintained layer of context that both humans and AI systems can use without making things up every five minutes.
The clever prompt is not the asset. The asset is the structured knowledge behind it.
If the model is judging five campaign angles against a vague customer profile, you have built a more elaborate guessing machine. If it is judging those angles against real objections, actual win/loss notes, margin data, proven hooks, and current sales priorities, now you have something.
This is also where the agency model starts to bend.
For years, agencies sold output. Campaigns, websites, content, ads, decks, reports. Some of it useful. Plenty of it expensive theatre.
AI attacks that model from underneath because output is getting cheaper. The more your agency is paid for hours and artefacts, the more AI looks like a discount conversation. “If the tool made it faster, why am I paying the same?” Fair question, if all you sell is production.
But if the job is to build and run a better commercial operating layer, the value is different.
Someone has to extract the knowledge. Someone has to structure it. Someone has to decide what the AI is allowed to know and do. Someone has to build the workflows. Someone has to test the outputs. Someone has to measure whether any of it actually improves leads, sales, retention, response speed, or margin.
That is not a content mill. That is an operating system.
It is also why trust and governance suddenly matter in very practical ways.
Google DeepMind and partners recently announced up to $10m for multi-agent safety research. That sounds like research-world stuff, but the business version is simple: once you have multiple AI systems acting inside workflows, you need supervision, audit trails, and failure handling.
Anthropic also recently apologised for invisible guardrails on a model that were not clearly disclosed to users. Whatever you think of the specific case, the lesson for operators is obvious: hidden behaviour kills trust. If an AI system refuses, downgrades, changes, or guesses, users need to know.
Clients do not need magical black boxes. They need visible systems.
The companies that win with AI will probably look a bit boring from the outside. They will not be the ones posting the most screenshots. They will be the ones doing the unglamorous internal work:
- cleaning up what the business actually knows;
- making sure that knowledge is current;
- connecting it to real workflows;
- setting permission boundaries;
- testing model output against reality;
- measuring commercial impact instead of counting prompts.
A company can have a hundred people using AI every day and still not have an AI strategy. It may just have a hundred people generating more drafts, more summaries, more half-useful outputs, and more stuff for someone else to check.
That is not transformation. It is admin with fireworks.
A proper AI strategy starts with different questions:
What does the business know that is not written down? Where does that knowledge live? Who owns it? How stale is it? Which parts can safely be used by AI? Which workflows would actually improve if AI had better context? How will we know if it worked? What should the system never do without a human?
These are not glamorous questions. Good. Glamour is usually where the waste hides.
For a business owner, the practical move is not to buy another AI tool this week. It is to run a memory audit.
Start with customers: who buys, why they buy, why they do not buy, what they misunderstand, what they complain about, what makes them stay. Then offers: what you sell, what the promise is, what the proof is, what the delivery reality is, where the margins sit. Then voice: not the brand adjectives — the real voice. The way the founder explains the thing on a good sales call. Then workflows: where does work slow down because people are searching, copying, rewriting, checking, or making the same judgement repeatedly? Then measurement: what would actually improve if the system worked?
Only then does the tool choice become interesting.
This is why I keep coming back to the idea that the next agency is not an output shop. It is an operating layer.
The old agency asked: “What assets do you need?”
The better question now is: “What does your business need to know, remember, and repeat?”
Once you answer that, AI becomes useful in a much more grounded way. It can draft, but with evidence. It can research, but with context. It can support sales, but with the real objections. It can create campaigns, but against a living view of the market and the customer. It can run agent loops, but with checks rather than blind autonomy.
AI does not make messy companies magically coherent. If anything, it exposes the mess faster.
So if your AI experiments feel underwhelming, do not immediately blame the model.
Look at the memory underneath it.
The model might be fine.
It just might not know your business.
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.