Most founders who’ve been experimenting with AI for the past year or two have the same collection of things.
A ChatGPT or Claude subscription someone uses for drafting. Copilot in the codebase. Maybe a chatbot on the website that was set up and then quietly forgotten. A few team members who’ve found prompts they like and use them for specific tasks. A Slack thread somewhere with “useful AI prompts” that nobody’s looked at in three months.
That’s not an operating system. That’s a drawer full of tools.
And the difference between a drawer full of tools and an operating system is roughly the difference between a 1x business and a 5x business.
Tim Barker, former CMO at Salesforce and CEO of a 600-person healthcare company, now running a five-person AI-native startup, spent seven months building the latter from scratch. He came to Business of Software Europe 2026 to explain what it actually involves.
The answer is not more tools. It’s not better prompts. It’s a system. And most businesses don’t have one yet.
What an AI Operating System Actually Is
The analogy Tim reaches for is a new hire.
Imagine bringing on the most brilliant person you’ve ever hired. Extraordinary capability. Across every function. Available instantly, at any hour, at a cost that would have seemed impossible five years ago.
There’s one problem: they wake up every morning with no memory of what happened yesterday. No institutional knowledge. No recollection of the decision made in last Tuesday’s leadership meeting, or the pricing change from three weeks ago, or the brand positioning you’ve been refining for six months.
So how do you manage that person effectively?
You’d do what you do with every high-performing hire. You’d onboard them properly. You’d give them your strategy, your values, your brand standards, your products, the markets you serve and those you don’t. You’d give them a clear role with explicit boundaries. Not vague, not broad, but specific enough that they know exactly what authority they have and when a decision needs to come back to you. You’d give them work to do, review their output, and build a feedback loop that makes their next output better than their last.
“You do all of this today,” Tim told the BoS Europe audience. “The great thing about building an AI-first system is you just need to institute all of these things in a way that AI can use them.”
That’s an AI operating system. It’s not software you buy. It’s organisational infrastructure you build.
The Architecture: Five Layers
Tim’s operating system lives in a folder structure. Five layers, each serving a specific purpose, each feeding into the others.
Strategy is the foundation. Your brand, your values, your products, your priorities for the year, the markets you serve. Everything that probably exists somewhere on a shared drive, probably stale, that every employee should know but most don’t fully. This is your first task: pull it together, refresh what’s out of date, and format it so AI can read it.
“It’s a good opportunity to refresh these things before you start feeding them in,” Tim noted. Most organisations discover that articulating their strategy clearly enough for AI to use it reveals gaps that existed long before AI was in the picture.
Agents is where the job specs live. Each agent has a defined role, clear constraints, and an explicit understanding of what decisions it can make independently and what requires human sign-off. Tim has 14 agents operating today. Some do work: writing, analysis, monitoring. Others are guardians, whose job is to check the work of the agents that produce it.
“Build guardians that check the work of agents,” he said. “You don’t want an agent having growth at any cost. You’ve got to balance it against trust and economics.”
Missions are time-bounded projects: the things you’re actively working on right now, with defined outcomes and deliverables. The output of one mission can become the input to another. Keeping missions discrete keeps token costs manageable and keeps agents focused.
State is your current reality: where the business is today, what decisions have been made, what’s in progress, what’s blocked. This is the layer that most businesses building with AI completely skip. That’s why their AI experience feels like Groundhog Day.
“Context rot is real,” Tim said. “Token windows are only so long. You need to make sure you are always compounding everything back into assets that AI can read every morning.”
Artifacts is the library of approved outputs (content, analysis, frameworks) that feeds back into the system as inputs for future work.
The Layer Most Businesses Are Missing: State
If there’s one thing Tim’s talk made clear, it’s that state management is the difference between AI that learns and AI that resets.
Every leadership meeting at Attain IP is recorded. A few hours after it ends, an agent processes the recording, extracts decisions, and presents Tim with a structured list to review. The ones he approves go into a decision log: a running timeline of every significant call the leadership team has made. That log gets fed back into the context that every agent reads when it starts a new session.
So when an agent wakes up tomorrow, it already knows about the pricing decision from last week, the hiring choice from three weeks ago, the new market focus from last month. It doesn’t need to be told. It reads the state, updates its context, and starts from an informed position rather than a blank slate.
Without this, you’re not building organisational intelligence. You’re re-explaining your business to a very capable but perpetually amnesiac contractor, every single day, from scratch.
Feedback Loops: Where the Compounding Happens
Tools produce outputs. Systems compound learning. The gap between them is feedback loops.
Tim gave a specific example. His agents write content in his voice and in the voices of his co-founders. Every time an agent produces a draft and a revised version gets approved, the system logs the gap between the two, capturing what changed and why. Over time, the gap narrows. The output gets closer to what Tim would have written himself, without Tim having to explain his preferences from scratch each time.
“There is a compounding effect if you just do this on all the assets you produce or draft over time.”
Another example: his LinkedIn monitoring agent delivers ten engagement recommendations every morning at 8am. His co-founder, a patent attorney of 30 years and not particularly active on LinkedIn, reads them, then replies to the agent’s daily email with feedback: which ones she’ll use, which ones missed the mark and why. The next morning, the agent has incorporated that feedback. The recommendations improve. The process costs $0.20 a day.
That’s not a magic prompt. That’s a system that gets better because it has a mechanism for getting better.
The Hard Constraints Your System Needs From Day One
Tim learned this one the expensive way.
Early in building Attain IP, he asked an AI agent with browser access to stub out some LinkedIn ad campaign skeletons, just placeholders he’d optimise and launch properly later. He didn’t specify that the agent wasn’t authorised to spend money. The agent set everything up, switched the campaigns on, and spent $2,000 on traffic to a staging website before Tim noticed three weeks later.
The AI did exactly what it was asked to do. There was no bug. The problem was the absence of a constraint.
“Make it super crisp,” he said of agent instructions. “That’s what’s going to make the whole system efficient and effective.”
Every AI operating system needs hard constraints embedded from the start: what agents can spend, what they can publish, what they can send, what they can change. Not as a brake on capability, but as the governance layer that makes expanding capability trustworthy over time.
The Difference Between a Copilot and an Autopilot
Tim draws a distinction that matters a great deal in practice.
A copilot AI works with you: it produces something, you review it, you approve it, it ships. An autopilot runs without you: it produces, decides, and acts. Both have their place. The question is whether you’ve built enough trust in the system, through enough real-world iteration, to remove yourself from a particular loop.
“Move at the speed of trust,” he said. “I’m not yet at a point where I want agents interacting with prospects and customers directly – until I’ve got their tone of voice to a point where I’m happy with it.”
The path from copilot to autopilot runs through the feedback loop. You review, you feed back, the system improves, you review less frequently, you eventually hand off the loop entirely. That progression is only possible if you built the feedback mechanism in the first place.
Where to Start
Tim’s advice for founders who want to build this is practical and unambiguous.
First: do it yourself. “All of you as leaders cannot delegate this.” Build something: a personal project, a single-function prototype, that forces you to encounter the real problems directly. Not to evaluate AI. To understand it from the inside.
Second: start with your strategy layer. Pull together the documents that define your organisation: brand, values, positioning, priorities. Ask honestly whether they’re current enough to be useful. This is foundational. Everything else builds on it.
Third: define one agent’s role precisely. Not broadly (“help with marketing”) but specifically (“monitor LinkedIn conversations in this domain, score them against these criteria, surface ten recommendations each morning, send to this email address”). Crisp role definition is the hardest part and the most important part.
Fourth: build the state layer before you need it. The decision log, the session summaries, the current priorities. Set up the infrastructure to capture this before there’s urgent content to capture. The habit is the thing.
Build Your AI Company Operating System
Tim Barker walks through the full architecture in the Business of Software workshop specifically designed for this: strategy layer, agents, guardians, missions, state management, and feedback loops.
→ Build Your AI Company Operating System
Tim Barker spoke at Business of Software Europe 2026. His talk drew on seven months of building Attain IP, an AI-native legal tech business, after leaving the CEO role at a 600-person digital healthcare company.