How to build a Company OS for an SMB
How to build a Company OS in an SMB: get the rules out of people's heads, write the brain, connect the tools you already pay for and run agents under gates.
To build a Company OS, you first write down the rules, meaning the judgment calls your team makes from memory, then you connect the tools and the agents to them. A Company OS is where you write down the rules your company runs on, connected to the software it already uses, so that software, your AI agents and your team apply the same rules. An AI agent is a program that carries out a sequence of tasks in your tools, for example replying to a lead by text and logging the reply in the CRM. The full definition is in our article on the Company OS. Aktivate, which sells schools software to run their sports teams, followed this method and put its first agent live 5 weeks after the project started.
TL;DR: The build starts with a team workshop, because nobody has written the team's rules down. Then you write the brain, meaning the company's context in a form an agent can act on. You connect the tools the team already pays for. Agents run behind a gate, a checkpoint where a human decides before the agent acts. Each scenario, meaning each type of situation the agent always handles the same way, earns its autonomy by passing set thresholds. Aktivate's reply triage ran on its own in week 13.
Where do you start building a Company OS?
Building a Company OS starts with an on-site workshop with the people who do the work. Your team knows its procedure from memory, exceptions included. The sales rep knows how to qualify a lead and the manager sees when a case needs escalating. The workshop turns that knowledge into written rules.
Nobody wrote these rules down, because no tool could apply them. An agent can apply them, as long as they are written down.
Writing these rules down takes an on-site workshop with the team. You sit with the people who do the work. You watch what they do and listen to how they say it. A questionnaire is not enough, because exceptions come out when someone walks through a real case.
The workshop answers three questions:
- Who decides what. Who approves a discount, who escalates a dispute, who answers for the result.
- Where each thing lands. A hot lead goes to the CRM, an alert goes to Slack, a client reply goes out from the inbox.
- Which tool wins when two disagree. If HubSpot and the inbox say different things, one of them wins and you write down which.
Scoping at Aktivate took two weeks. At Houdz, we sat with the sales reps. We wrote down what they do and how they say it, then turned it into rules. Once the workshop is done, the writing is mechanical.
The five build steps
Step 1, write the brain
The brain holds what an agent needs to know to act like someone on the team: who the clients are, what was promised, which rule applies.
Sort it by how fast things change.
| Part | What it holds | How it is written |
|---|---|---|
| Stable facts | Legal entity, insurance, anything that changes about once a year | Corrected when a fact changes |
| Running record | Deals, quotes sent, decisions taken | New lines added, nothing deleted |
| Working memory | Current priorities, this month's summary | Rewritten, then thrown away |
Keeping the brain current is the hard part. Gyld, a vendor of "company brain" tools (software that stores a company's context), writes that "most implementations break at the maintenance stage". Sorting by speed of change gives each part its own review rhythm.
When two documents disagree, write the conflict down as a question. The question carries the name of the person who can settle it.
A brain on its own does nothing. Many tools sold as a "company brain" stop there. Y Combinator named the company brain in its Summer 2026 Request for Startups. The Company OS vs ERP, Notion and SaaS comparison covers how that differs from an ERP, a Notion template or a software suite such as HubSpot or Microsoft.
Step 2, write the rules
Each rule is a judgment call from the team, written down and approved by a human.
Aktivate's sales reps text high school coaches before the season starts. Before the first text went out, we listed what a coach can reply:
- yes
- send me the link
- a question the agent cannot answer
- the season is over
- wrong number
Each reply is a scenario with its own instruction. When a reply fit none of them, we wrote a new scenario with its instruction. From then on, the agent handled that kind of reply with the new instruction.
Other rules decide who gets which case, which message waits for a human before it goes out and which tool wins when HubSpot and the inbox disagree.
Step 3, connect the tools
The Company OS plugs into the tools the team already pays for. Nobody switches tools. The surfaces, meaning the places the team already works (Slack, Teams, the inbox, the CRM), stay the same. A new login is a failure, because it is one more tool the team has to remember to open.
Matrix OS, a vendor of agent software, describes a Company OS in the same terms in its post on building one: "a shared environment where people and agents can work across the tools a company already uses."
Several tools now ship their own AI: HubSpot has Breeze, Slack has Slack AI. Each one sees only its slice of the company. The agents of a Company OS all read the same brain and the same rules, whichever tool they work in.
Aktivate's agent works inside SalesMessage (the SMS tool), HubSpot, Raise (the tool where Aktivate builds each team's fundraising page) and Slack. These tools were already in place at Aktivate. The agent sends a volume of messages there that no rep could reach by hand. When a coach is interested, the rep sees it where they already work: a task in HubSpot and a message in Slack.
Step 4, run the agents under gates
Every agent stops at every gate. The agent proposes, a human decides.
Aktivate's agent worked this way for seven weeks, from week 5 to week 12 of the project. During the first four (weeks 5 to 8), the reps also reviewed its behavior. Each correction became a rule.
The agent that writes never grades itself. A second agent, which never writes a message, grades every conversation. Hot leads go to a rep, with a task in HubSpot and a message in Slack. The agent writes everything else back to HubSpot.
A Zapier-style automation moves data from one tool to another. It has no context and no judgment. AI agent vs automation covers the difference.
Step 5, earn autonomy one scenario at a time
At Aktivate, a scenario goes on autopilot once the agent has been right 6 times in a row and 85% of the time overall. After one wrong call, it goes back behind the gate.
After seven weeks, every scenario was on autopilot. Triage runs on its own. The reps only see the coaches who need a person.
How it went at Aktivate
Aktivate sells schools software to run their sports teams: registration, team management and fundraising. More than 150 schools use it. Before the agent, a rep contacted about 150 coaches a day.
| Project week | Step |
|---|---|
| 1 to 2 | Scoping: workshop with the reps, rules, tools |
| 5 | First agent live |
| 5 to 8 | Behavior review with the reps |
| 10 | Full volume: from 150 to 1,500 coaches contacted per rep per day |
| 13 | Reply triage on autopilot |
In eight weeks, Aktivate contacted 25% of its contact base without hiring new reps. The reply rate is about 30%.
Chapman Walk, a sales rep at Aktivate, on the agent he tests with Alex Houdayer, the author of this article:
"It definitely sounds like a human. Alex and I have been testing it. It's doing really well."
Core and config: what is yours stays yours
The pieces of a Company OS (the brain, the rules, the agents, the gates and the surfaces) fall into two parts.
The core is the part shared by every company in the same sector: the agent logic, the gates and the rules that hold for the whole sector, such as never texting a contact who has opted out. Each improvement to the core gets a version number. It reaches every company at once. So each new installation starts with much of the work already done.
The config is the part that belongs to you alone: your brain, your access credentials, who gets which case, your tone, your clients. A core update never touches it.
Your rules do not go to your competitors. What holds for the whole sector goes into the core. Anything about your clients or your way of working stays in your config.
Why do Company OS projects fail?
A Company OS project fails when the team makes one of four mistakes: plugging everything in without written rules, skipping the workshop, launching the agent without a gate or recording no numbers before the first agent. In each case, the agent's mistakes reach your clients or nobody can say whether it does better.
- Plug everything in and say "run my business". Without rules, the agent guesses. Its mistakes land with your clients.
- Skip the workshop. You write the procedure as it looks on paper. The exceptions are missing. They are what stop the agent once it is live.
- Launch without a gate. The first mistake goes to a client. The team loses trust in the agent before it has had a chance to learn.
- No numbers before the first agent. Write down today's numbers before the agent starts: how many contacts a rep reaches per day, how many reply. Aktivate's starting point was 150 coaches contacted per rep per day. Without a starting point, nobody can say whether the agent does better.
More traps are listed in the AI integration mistakes SMBs make.
How we do it
At Houdz, the first agent runs in two to four weeks. The agents work inside the tools your team already uses. The team works on its own after we leave. To see whether the method fits your company, book a discovery call.
Frequently asked questions
How long does it take to build a Company OS?
A few weeks for the first agents. At Aktivate, scoping took two weeks, the first agent went live in week 5 of the project and reply triage ran on its own in week 13.
Do we have to change tools?
No. The Company OS plugs into the tools the team already pays for, such as HubSpot, Slack or the inbox. Every new login is one more tool the team has to remember to open.
Do we need a tech team?
No. You need one person who knows the work and approves what the agent proposes during the first weeks.
Do our rules go to our competitors?
No. What holds for your whole sector goes into the core, the part shared by the companies in that sector. Anything about your clients or your way of working stays in your config, the part that belongs to you alone.
When can an agent act without human approval?
One scenario at a time, once it has proved itself. At Aktivate, a scenario goes on autopilot once the agent has been right 6 times in a row and 85% of the time overall. After one wrong call, it goes back behind the gate.
What is the difference between a company brain and a Company OS?
A company brain holds the company's context: its clients, its commitments, its stable facts. A Company OS adds the rules, the agents, the gates and the surfaces, meaning the tools where the team already works.
What to remember
A Company OS starts with a workshop with the people who do the work. The workshop brings out the team's rules, then you write them into the brain.
Agents run inside the tools the team already pays for, behind a gate. Each scenario earns its autonomy by passing set thresholds and loses it at the first wrong call.
At Aktivate, the first agent went live in week 5 of the project and reply triage ran on its own in week 13.
Last updated: 2026-09-24