# How We Run Our Company With AI Agents on Workflows That Remember

Running a company with AI agents means every job runs on a written workflow, not a chat: its first step recalls what was decided before, a person signs off where it matters, and the job keeps what is worth remembering. We run ours on ConvOps, where our tasks, workflows and memories live.

By David Marsa, Founder & CEO. Updated 2026-10-08.

## What you will be able to do

- Pick the first recurring job in your company to hand to an agent.
- Have your AI write that job as steps and mark where a person must say yes.
- Explain the closed loop: a task follows a workflow, keeps what is worth keeping, and the next job's first step recalls it.
- Decide, for each job, whether it waits for a person or goes live after an automated check.
- Know which loop guide to read next for search or site quality.

## Key takeaways

- We do not run the company from a chat: every job is a task that follows a written workflow with steps, sign-offs and a record.
- It runs as a closed loop: jobs keep what is worth keeping, every sign-off note included, and every job's first step recalls what was decided before.
- Code changes, changes to how a site works, and guides wait for a person's yes; routine content goes live after an automated second check, and every change is logged.
- When a job goes wrong, the fix becomes a written decision that every later run recalls first.
- Start with one boring weekly chore, have your AI write its steps, run it once by hand, then put it on a timer.

To run a company with AI agents, put every job on a written workflow, not in a chat. We run ours that way: each job's first step recalls what was decided before, a person signs off where it matters, and the job keeps what is worth remembering for the next one.

This guide is the hub of Loops We Run, our series on the recurring jobs AI agents do for us, where every run feeds the next. It covers our week, the shared brain, one request from start to finish, who decides what, and how to start your own.

## Not a chat: every job runs on a workflow

We do not run the company from a chat. Every job is a task on a workflow, and that one rule makes everything else possible.

A chat is question, answer, forget. Chats also drift: in [a 2025 study](https://arxiv.org/abs/2505.06120) across six task types, AI models did 39% worse on average when a task arrived over several chat turns instead of as one written instruction.

| A chat | How we work |
|---|---|
| You ask, it answers, it forgets | Every job is a task with its own history |
| No steps: the AI improvises each time | The task follows a workflow: written steps, one at a time |
| Nobody signs off | It stops at the sign-off points we set and waits for a person |
| Nothing is kept | Every step leaves a record, sign-off notes are kept, and the job stores what is worth keeping |
| The next chat starts from zero | The next job's first step recalls what was decided before |

**What is an AI workflow?** An AI workflow is a written list of steps an AI agent follows one at a time, with set points where it stops and waits for a person's yes.

A job starts on a timer or from a request and becomes a task, a to-do with its own history. Our tasks, workflows and memories live in ConvOps.

This is not an autonomous company. A person decides what each job may do, and code changes, changes to how a site works, and guides wait for a person's yes. It is one stage of the [AI operating model journey](/blog/ai-operating-model-journey).

## Meet Copperfern, the made-up shop in our examples

Every example from here on uses Copperfern, a fictional company. Each job runs exactly the way ours do; the company, the people, the messages and every number are invented.

| Who or what | In the examples |
|---|---|
| Copperfern | An online homeware shop selling kitchenware and table linen, in English and Spanish |
| Maya | Runs operations and signs off when a job stops for a person's yes |
| Website and support teams | Two team spaces, both storing their notes in the one shared brain |

## Our week: the same jobs on a timer

Most of what keeps a company running is the same work every week, and that is exactly the work we hand to scheduled AI agents. One in five EU companies with 10 or more staff already uses AI (20.0% in 2025, up from 13.5% a year earlier, according to [Eurostat](https://ec.europa.eu/eurostat/web/products-eurostat-news/w/ddn-20251211-2)), so the question is how it runs, not whether. Each job is one turn of the loop. Times are UTC.

![Every recurring job sits on the calendar: the person marks the ones that wait for a yes, the check marks the ones an automated review lets through.](https://storage.googleapis.com/neomanex-public-assets/neomanex/guides/run-a-company-with-ai-agents-v1-v1.png)

| Job | When | What it does |
|---|---|---|
| Search check | Every day, 05:00 | Reads how the site shows up in Google and AI assistants, changes nothing ([full loop](/learn/guides/seo-experiments-claude-code-search-console-loop)) |
| Search planning | Monday, 06:00 | Plans one batch of site changes; each ships only after a person approves |
| Site check | Tuesday, 05:00 | Reads the site like a reader, files a task per problem ([full loop](/learn/guides/self-healing-website-ai-agent-qa-fix-loop)) |
| Site fixes | Wednesday, 00:30 | Fixes small problems, re-checked live; the rest is held for a person |
| Directory check | Monday 00:00 (AI models), every 2nd Thursday 00:00 (AI tools) | Re-checks facts in our public directory of AI models and tools |
| News | Tuesday and Friday: scan 01:00, publish 03:00 | Live only after a separate agent's review passes |
| Translations | Tuesday and Friday, 03:45 | Reviewed by a separate agent |
| Changelog | Monday, 04:30 | The weekly issue of what changed in the AI tools and models we track, reviewed by a separate agent |
| Engineering changes | On request | Agree what, build, check, approve going live |

Search research and strategy are scheduled monthly, on day 2 and day 3 at 07:00. Engineering work has no clock: it starts from requests.

## The company brain: a closed loop

A chat forgets. Our company does not.

![Every job keeps what is worth keeping in one shared brain, and the next job's first step recalls it.](https://storage.googleapis.com/neomanex-public-assets/neomanex/guides/run-a-company-with-ai-agents-v2-v1.png)

**What is AI agent memory?** AI agent memory is how the AI remembers what we decided: short written notes of decisions, lessons and corrections that one job stores and the next job reads before it starts.

The loop has four parts (callouts 1 to 4):

1. **Task.** Every piece of work is a task that follows its workflow.
2. **Kept.** A closing step stores what is worth keeping: a decision and why, a lesson, a correction, a preference. A job with nothing new stores nothing new. Every sign-off note is kept, on the task and in the brain.
3. **One shared brain.** It is one brain. Search from any team space (support, website, sales) and it looks across all of them, and every AI assistant we connect reads and writes the same brain.
4. **First step recalls.** Every job's first step recalls the decisions and lessons that match it, plus where the relevant things live.

At Copperfern, the support team stored a correction: "Say 30 days, never one month." A website task's first step finds that note, though it lives in the support space.

A job can store its own correction. Copperfern's Monday changelog job sees it was started twice, stops instead of repeating the work, and stores: "The Monday changelog can start twice; check for a run already in progress."

The same loop runs for content, code, sales and infrastructure. Copperfern's task "Move the shop database backups to the new storage" opens by recalling "Backups are kept 15 days, decided in March." The change keeps 15 days, and the re-check confirms it.

The agent knows what was already decided, so it does not ask again, and what went wrong last time, so it does not repeat it. One brain is the opposite of [agent sprawl](/blog/agent-sprawl). For the deeper idea behind written steps, see [graph engineering](/learn/guides/what-is-graph-engineering).

If your AI starts from zero every time you open it, [this is how ConvOps works: tasks, written steps, and a memory your agents check first](https://convops.app).

## One request: recall first, two sign-offs

The interesting part of a request is not the work. It is what the agent recalled before starting, and the two moments a person says yes.

Copperfern's request: "Our free returns window goes from 14 to 30 days. Update everything that mentions it."

![The first step reads what was decided before, the work sits between two sign-offs, and a second check catches the miss first.](https://storage.googleapis.com/neomanex-public-assets/neomanex/guides/run-a-company-with-ai-agents-v3-v2.png)

1. **A request becomes a task.** "Returns window 14 → 30 days", status To do, workflow attached (callout 1).
2. **The first step recalls.** A lesson: "Returns window also appears in the order confirmation email." A decision: "The help page is translated separately into Spanish." (callout 2)
3. **It writes what must change.** 5 places: the returns page, the help page in English and in Spanish, the checkout footer and the order confirmation email.
4. **Sign-off 1.** Maya, who runs operations, approves the list, and the list is frozen. Her note, "5 places, Spanish help page included", is kept (callout 3).
5. **The agent does the work.** It edits the 5 places.
6. **A second agent checks it.** It finds the Spanish help page still says "14 días", and the miss is fixed before Maya is asked again (callout 4).
7. **Sign-off 2.** Maya approves going live.
8. **Live and re-checked.** Each place is re-read live, and all 5 say 30 days (callout 5).
9. **Kept for next time.** A new memory: "Returns window lives in 5 places, including the order confirmation email." The task keeps its full history (callout 6).

Step 2 is why the list has 5 places, not 3.

## Who decides what, and what happens when a job goes wrong

We decide before a job ever runs which of its changes wait for a person and which go live after an automated check. That is human in the loop in plain words: a person signs off where the risk is ([more here](/blog/human-in-the-loop-ai-systems)).

Even specialised AI tools get facts wrong, which is why a second check stands between the work and going live: [Stanford researchers](https://law.stanford.edu/publications/hallucination-free-assessing-the-reliability-of-leading-ai-legal-research-tools) found that leading AI legal research tools each made things up more than 17% of the time (2025).

| Waits for a person's yes | Goes live after an automated check, and is logged |
|---|---|
| Code changes: agree what, then approve going live | News posts and the changelog: a separate agent reviews, and a fail stays a draft |
| Changes to how a site works: each search change ships only after a person approves | New directory entries and translations: a separate agent reviews |
| Guides like this one: plan, images, publishing | Directory facts: checked by the checking agent |
| Anything the small-fix job may not touch: held for a person | Small fixes to published pages, re-checked live |

A person decides each job's column before it runs, and reads the record of what went live (see [AI agent observability](/blog/ai-agent-observability)).

### When a job went wrong

We hit this one ourselves. Retold on Copperfern data, the daily search check raised an alert: "CRITICAL: click rate 0.4%, industry standard 3%", plus 5 recommendations nobody asked for. The benchmark was made up.

![The check invented a benchmark; we decided it may only flag a fixed list, and every run recalls that decision first.](https://storage.googleapis.com/neomanex-public-assets/neomanex/guides/run-a-company-with-ai-agents-v4-v1.png)

Maya and the team made a decision. The daily check may only flag items from a fixed list of 4 checks: search visits drop, AI assistant visits drop, pages under test missing (pages we are trying a change on), home page loads. Each is ok or flag. A script computes every number. No outside benchmarks, no recommendations.

The decision is stored, recalled first by every later run, and written into the check's steps. Next morning, all 4 read ok (callout 2).

## How to start running your business with AI agents

Start with the most boring job you do every week, not the most impressive one, and run it by hand before you put it on a timer.

![Five steps from a weekly chore to your first routine: your AI writes the steps, you decide where it stops for your yes.](https://storage.googleapis.com/neomanex-public-assets/neomanex/guides/run-a-company-with-ai-agents-v5-v1.png)

Pick a chore you can check in a minute: a weekly report, a stock check, a price list. No code, no editor. Your connected AI writes the steps, and you decide where it stops for your yes (the person mark on step 3).

[Create a free ConvOps account and set up your first routine](https://my.convops.app/register). Want us to set up the rhythm with you? [Book a free Discovery Session](/services).

## Steps

1. **Pick one weekly job**: Choose one chore you do the same way every week, such as a weekly report, a stock check or a price list update, and write one sentence on what it produces and for whom. Create a free ConvOps account at https://my.convops.app/register; this job becomes your first task.
2. **Connect the AI you already use**: Connect your AI assistant so it can read and work your tasks. The steps for each assistant are on https://convops.app/connect.
3. **Have your AI write the steps**: Ask your connected AI to write the job's steps in plain sentences, the way you would brief a new colleague, and to mark the step where it must stop for your yes. There is no editor to learn: you describe, it writes the workflow.
4. **Run it once by hand**: Start the task with your assistant and watch it follow the steps one at a time. Correct it where it goes wrong, and tell it what is worth keeping for the next run.
5. **Put it on a timer and keep notes**: Once a hand run goes well, schedule it, every Monday morning for example; ConvOps creates the task on time, every time. Read the record each week and turn every surprise into a written decision. Create a free ConvOps account at https://my.convops.app/register to set up your first routine.

## FAQ

### What does it mean to run a company with AI agents?

Running a company with AI agents means every recurring job runs as a task on a written workflow, not in a chat. In ConvOps, the first step of each job recalls what was decided before, a person signs off where it matters, and the job keeps what is worth remembering. It is not an autonomous company: a person decides what each job may do.

### Who is running a company with AI agents for?

Running a company with AI agents is for founders, operators and managers whose week repeats: reports, site checks, content and updates. ConvOps turns each recurring job into a task on a written workflow, so each job's task is created on time, every step leaves a record, and a person still decides the risky changes: code changes, changes to how a site works, and guides.

### How does the AI remember what we decided?

In ConvOps, jobs store what is worth keeping: a decision and why, a lesson, a correction or a preference. Every sign-off note is kept too. The first step of every job recalls the decisions and lessons that match it, so the agent does not ask again what was already decided and does not repeat what went wrong last time.

### Can AI agents run a company without people?

No. In ConvOps, a person decides what each job may do before it runs. Code changes, changes to how a site works, and guides stop for a person's sign-off. Routine content, such as news and translations, goes live after an automated second check and is logged, and a person reads that record and turns surprises into written decisions.

### Is it safe to let AI agents do company work?

Letting AI agents do company work is safe when the risk is decided before a job runs. In ConvOps, code changes, changes to how a site works, and guides wait for a person's sign-off, and routine content passes an automated second check before it goes live. Small fixes run only where the result can be re-checked, the rest is held for a person, and every step leaves a written record.

### How do I implement AI agents in my business?

Create a free ConvOps account at my.convops.app/register. Then pick one chore you do every week, connect the AI assistant you already use, and ask it to write the job's steps and mark where it stops for your yes. Run it once by hand, correct it, and tell it what to keep. Then put it on a timer.

### How much of my time does running a company with AI agents take?

What stays with you is the judgment: the sign-offs, reading the record of what went live, and turning surprises into written decisions. ConvOps keeps every task's status and full history, so you read the record instead of chasing people. The agents do the steps in between.

Canonical: https://neomanex.com/learn/guides/run-a-company-with-ai-agents