Insight / signal
Your AI agents are not a team until you can control them
AI-native does not mean the robots run the company. It means the company has an operating layer good enough for people and agents to work through.
The AI-native company sounds brilliant in a demo.
You set a goal. Agents research the market, write the proposal, build the microsite, run a usability test, summarise the feedback, and produce version two before lunch.
That is not fantasy. It is starting to look like the normal direction of travel.
But there is a quieter, less glamorous question underneath it:
Who is actually in control?
Not in the sci-fi sense. I mean the boring operational version. Who knows which agents are running? Which data they can see? Which tools they can use? What they cost? What happens when they are wrong? Who signs off the work before it reaches a client?
That is the bit most AI conversations skip because it is not interesting to talk about. It is also where the real value lives.
A recent Startup Ideas episode framed the AI-native organisation usefully: people, agents, and context. People handle judgement, strategy, trust and taste. Agents execute repeatable work through tools. Context is the shared brain that stops every task starting from zero.
That model is right. It maps neatly onto how a serious modern service business should actually run.
But if you stop there, you build a very fast mess.
IBM has just put numbers on this. In a June 2026 study of 2,000 C-level technology executives, two-thirds of CIOs and CTOs said they are accountable for AI systems they do not fully control. Seventy per cent said teams across the business are deploying technology faster than IT can track. Only 11% said they are fully ready for the scale of AI agent deployment expected in the next year.
That is the real story.
Not “AI will replace everyone”. Not “AI will change everything”. We have heard enough of that. The more immediate problem is that companies are creating invisible operational debt. Agents are being added before the business has the habits, checks, permissions and review loops to manage them.
If that sounds like an enterprise-only problem, it is not.
Small businesses will hit the same wall, just with fewer meetings and less paperwork. Someone signs up for three AI tools. Someone connects customer data. Someone automates follow-up messages. Someone builds a support bot. Then six months later nobody quite knows what is live, what data is flowing where, what the monthly bill is, or whether any of it is still aligned with the business.
That is not transformation. That is clutter with API keys.
The better way to think about AI-native work is not “how many agents can we deploy?” It is “what operating layer do we need around them?”
That layer needs a few boring pieces.
Shared context, so the agent knows the customer, offer, tone, constraints, past work, live priorities and commercial goals. Clear workflows, so the agent is not improvising every time. Tool permissions, so it can do useful work without wandering into places it should not. Review points, so humans decide where judgement still matters. Evals, so the business can define what good looks like before it scales the process. Logs and cost visibility, so someone can inspect what happened afterwards.
This is where a lot of AI marketing talk is still stuck in the wrong place. Too many people are selling output. More posts. More emails. More variations. More synthetic faces saying synthetic things on synthetic backgrounds.
Some of that is useful. Most of it is noise.
The better agency opportunity is building the control layer for commercial work.
A proper AI campaign operating system does not just write LinkedIn posts faster. It connects research, positioning, assets, publishing, follow-up, sales feedback, measurement and learning. It keeps the human in the places where taste and trust matter. It lets agents handle the grind without pretending the grind is the strategy.
That distinction matters.
Anthropic published a useful technical example from biology this week. Their team tested scientific agents retrieving sequence data from NCBI Virus. Even strong models did not reliably reach the accuracy needed for dataset construction. When the team added a deterministic retrieval layer, accuracy rose to nearly 100%.
Different domain, same lesson.
The model was not enough. The surrounding infrastructure mattered. The tool layer mattered. The workflow mattered. The agent needed a cleaner road to drive on.
That is exactly what most businesses are missing.
They are buying clever engines and sending them down muddy tracks.
For a business owner, the practical question is simple: where do you want agents inside the company, and what has to be true before you trust them there?
Customer research? Low risk, high value. Let agents gather signal, summarise patterns, draft hypotheses.
Proposal creation? Useful, but needs context and human sign-off.
Outbound follow-up? Possible, but permissions, tone, compliance and brand risk all matter.
Client reporting? Great, if the data is clean and the claims are checked.
Strategy? Agents can help, but they do not own the consequences. You do.
That is the line.
AI-native does not mean the robots run the company. It means the company is redesigned so people and agents work through a system instead of a pile of disconnected tools.
This is also where the post-agency model gets interesting.
The old agency sold deliverables. Campaigns, pages, emails, posts, decks. The AI-era agency cannot rely on that for long, because output is getting cheaper fast and every client can feel it.
The useful agency builds operating layers.
It helps the client decide where AI belongs, where it absolutely does not belong yet, what data and context is needed, which workflows should become agent-assisted, how review works, how to measure quality, and how to stop the whole thing turning into a haunted cupboard of automations nobody understands.
That is less glamorous than a viral AI demo.
It is also a better business.
Because if you build the operating layer, you are not just providing content. You are changing the speed and responsiveness of the commercial system. Research gets into campaigns faster. Customer feedback gets back into positioning faster. Sales objections become usable content faster. The team learns faster.
That is the point.
More agents is not the goal.
A business that can safely aim agents at real work, inspect what they did, improve the workflow, and keep the human judgement where it belongs will beat a business with fifty disconnected tools and a folder full of impressive demos.
The uncomfortable rule:
If you cannot inspect it, you do not control it. And if you do not control it, you probably should not scale it.