Insight / signal
The new AI divide is ownership, not usage
A prompt is not an asset. A repeatable agent workflow might be.
I think we are asking the wrong AI question.
For the last two years the question has been: “Are you using AI?”
That was useful for about five minutes. Maybe ten if we are being generous.
Now it is too blunt. Everyone is going to use AI. Your staff. Your competitors. Your accountant. Your kids will probably be better at it than half the boardroom.
Usage is not the advantage.
The better question is: what part of the AI system do you actually own?
That sounds abstract, so let me make it blunt.
A prompt is not an asset. A ChatGPT tab is not an operating model. A few saved replies in someone’s browser history are not company infrastructure. They might help with a task, but they do not compound.
What compounds is the stuff around the model.
The workflow. The source material. The customer context. The approvals. The logs. The repeatable skill. The named owner. The measurement. The distribution. The habit of turning a messy useful job into a system that can run again tomorrow.
That is ownership.
And that is where the next serious AI divide opens up.
This came through hard in the latest Diary Of A CEO debate between Daniel Priestley and Nick Hanauer. Strip away the left-right economic argument and the useful point is simple: if AI accelerates the hollowing out of routine work, the split gets nastier between people who own leverage and people who only sell hours.
Ownership can mean equity, property, audience, distribution, intellectual property, or a business. Increasingly, it also means owned AI workflows.
Not “I know a few prompts”.
Proper operating assets.
A sales follow-up system that knows your offer, checks CRM context, drafts the next move, and asks before sending anything risky.
A content system that takes a podcast, call or screen recording and turns it into drafts, clips, posts, email angles and source-backed notes.
A research system that scans the market every morning, ignores most of the noise, and gives you one useful commercial angle.
An ops agent that checks services, reads logs, spots broken workflows, and leaves a note a human can actually trust.
That is a different thing from “using AI”.
It is also where a lot of agencies are going to get caught out.
If your offer is still basically “we can make more content because AI”, good luck. That market is already sliding into mush. More output is not the same as more value. In many businesses, more AI content just means more beige noise moving faster.
The stronger offer is operational: take one repeated commercial process and turn it into a controlled AI workflow.
That is not as flashy as saying “fully autonomous AI employees”. It is also much less stupid.
The Hermes walkthrough from Alex Finn is a good example of where this is going. The interesting parts were not the shiny demo bits. They were sessions, profiles, artefacts, skills, cron jobs, sub-agents and context management.
Boring words. Useful words.
Because that is how AI turns from a conversation into infrastructure.
A profile gives the work a role and a model choice. A skill turns a repeated method into something reusable. Artefacts become memory. Cron makes the work show up without someone remembering to ask. Sub-agents handle bounded tasks in parallel. Context management stops the whole thing getting expensive and messy.
None of that is guru nonsense. It is ops.
The Matt Van Horn Agentic Engineering note points in the same direction. Fresh research first. Plan for the agent. Execute in a bounded way. Review. Then convert the repeated lesson into a skill.
That last bit matters.
Most people use AI and lose the benefit the moment the chat ends. Operators turn the useful bit into a reusable asset.
That is the ownership gap.
There is another signal worth watching: agents are starting to get real-world identities.
Saperly is early-stage, so I would not treat it as client-production infrastructure yet. But the direction is obvious. Agents can have phone numbers, SMS routing, audit trails and compliance boundaries. Suddenly the agent is not just drafting in a browser. It can talk to customers. It can be called. It can send a message. It can sit inside a support or booking flow.
That is exciting. It is also a liability if you are casual about it.
The moment an agent touches customers, phones, inboxes, CRM records or payments, you are out of toy-land. You need consent, logs, escalation, QA, permissions and someone responsible when it gets weird. And it will get weird, because software always does. AI just gives the weirdness better grammar.
This is why Anthropic’s partner announcement is worth paying attention to, even if you ignore the inevitable enterprise gloss. Their line is basically: pilots are not the same as systems a business can run on. The real work is integration, evaluation and how people’s work changes.
That is the market.
Not prompts. Not AI vibes. Not another agency deck with a robot on the cover.
Integration. Evaluation. Workflow redesign. Governance that does not strangle the useful parts. Human judgement where it matters. Automation where it removes drag.
For business owners, the question becomes very practical.
Where are you renting AI, and where are you building something you own?
If all your AI use lives in individual chat histories, you are renting. If one person leaves and the workflow disappears with them, you are renting. If nobody can explain why the agent did something, you are renting chaos with a nicer interface.
Owning the system looks different.
You have a small number of important workflows. Each one has a clear job. The agent has limited access. The source material is known. The output is checked. The logs exist. There is a human owner. The workflow improves because the team reviews it, not because everyone hopes the model gets smarter next month.
Less glamorous than the usual AI keynote. Also how businesses actually get leverage.
This is the post-agency shift I keep coming back to. The next serious agencies will not win because they can generate more assets. Everyone can generate more assets. They will win because they can build the operating layer around the work.
Research to positioning. Positioning to campaign. Campaign to assets. Assets to publishing. Publishing to follow-up. Follow-up to measurement. Measurement back to the next decision.
AI can speed up every part of that loop, but only if the loop exists.
If the business is already chaotic, AI mostly gives the chaos a faster engine.
So no, the question is not “should we use AI?” anymore.
The question is: what AI leverage are you building that still exists next week, next month, and after the person who made the original prompt has forgotten what they typed?
Because that is the dividing line.
Prompt dabblers will get a productivity bump.
Operators will build assets.
And in a market where routine work gets cheaper every month, I know which side I would rather be on.
A practical checklist before buying another AI tool
- What repeated workflow are we improving?
- Who owns it?
- What source material does the agent use?
- What can it do without approval?
- What must a human approve?
- Where are the logs?
- How do we know the output is good?
- What happens when it fails?
- Does this workflow still exist if the original prompt-writer leaves?
If you cannot answer those, you are probably not building AI infrastructure yet. You are just using AI.
That is fine as a starting point.
It is not a moat.