← All posts

Essay · August 2026

The Every case: 25 people and four AI agents in management

Every, a media company of about 25 people, handed four custom AI agents its task prioritization, the turning of meetings into action items, OKR planning and the daily growth report. OKR planning shrank to two days. All four agents read one set of linked databases — and it's the linked data, not the choice of model, that decides whether an agent works.

Every, a media company of about 25 people, handed four custom AI agents its task prioritization, the turning of meetings into action items, OKR planning and the daily growth report. OKR planning shrank to two days. The agents don't work in isolation: all four query one set of linked databases. Every published the case in April 2026; Grow2.ai breaks it down and translates it into SMB operations.

Most write-ups about AI agents in a team end on the word "productivity". The Every case is interesting for something else: you can see which management process specifically disappeared — and what had to be in place for it to disappear.

This is someone else's case. Grow2.ai didn't implement it and has no access to Every's internal data. What follows is a read of the published material plus our own conclusion about what carries over to a team of 25–50 and what doesn't.

Who Every are, and why this case isn't about "another chatbot"

Every runs six products, a media publication and a consulting arm at the same time. About 25 people. By the company's own account, everyone has roughly 30 tasks on their list at any given moment.

That's how most SMBs live: several directions, the same people on several fronts, and no separate project office to keep it all in sync. The difference is that Every described its way out in public.

The problem: one COO as the router for the whole company

Before the agents, the dispatcher role was carried by Brandon Gell, Every's COO. To answer "what do we take on first", he manually cross-checked three separate sources: launch calendars, the company's strategy documents, and task lists.

That setup works exactly as long as one person has time to maintain it. The COO is busy for a week and the team starts prioritizing by feel.

Now the same Brandon writes in Slack to an agent named Anton and gets an ordered list back in seconds — for himself and for other people on the team.

The solution: four agents, one shared brain

Over a few months Every built four custom agents with Notion AI. All four read the same set of linked databases: strategy, calendar, tasks, people, meeting notes.

Agent

The process it took over

What has to exist in the system for that

Prioritization (Anton)

manual cross-checking of calendars, strategy and task lists

a written strategy, a single task list, a launch calendar

Meeting → tasks

going through notes after a meeting and handing out action items

meeting notes in one database, linked to people and projects

OKR planning

a long cycle of aligning goals across directions

goal history, metrics and owners in structured form

Growth report

collecting numbers by hand every day

growth metrics the agent can reach without a manual export

The left and middle columns come from Every's write-up. The right column is ours: a reconstruction of the data requirements, not a quote.

What changed: the numbers Every stated publicly

Process

Before

After

Getting the priorities for the day

the COO cross-checks three sources by hand

a Slack message, an answer in seconds

OKR planning

duration not stated in the public part

two days

There aren't many numbers, and that's not an accident. The public part of the material has no implementation cost, no hours saved and no ROI: the full breakdown of each agent and the prompts themselves sit behind Every's paid subscription.

So the case proves that management routine can genuinely be moved into an agent. What it will cost you specifically, it doesn't prove — our own calculation of the full cost of a story like this is in the piece on what an AI agent actually costs.

Three principles Every took away from building agents

Describe the outcome, not the steps. Tell the model what has to be achieved and let it pick the implementation. Over-specifying — "create a database, then add a relation, then filter by date" — tends to get in the model's way rather than help.

Your workspace is the agent's brain. Custom agents get strong when they can query databases that are linked to each other. Every's agents work precisely because strategy, calendar, tasks, people and notes sit in one place and reference each other. First the company kept those databases for itself, and the agents came later, on top. Not the other way around.

Don't write the agent's instructions alone — describe what it has to achieve and let the AI generate the instructions. Every also built agents through Claude Code and the Notion API straight from the terminal.

Translating this to an SMB: what works in a team of 25–50

The condition without which nothing works

Notion in this case can be swapped for anything: Bitrix24, HubSpot, Google Workspace, ClickUp. What can't be swapped is the state of the data.

An agent that's meant to set priorities has to read somewhere what the company's priorities are. If the strategy lives in the owner's head, tasks are scattered across four chats and the "launch calendar" exists as a voice message on Monday morning, the agent will produce a confident list assembled out of nothing. That isn't a model error, that's a missing source.

A new hire in their second week should be able to answer "what matters most right now" from your systems alone, without talking to the owner. If they can't — data first, agents after. Preparing the data is usually the budget line nobody planned for.

One agent = one process

The temptation to build a universal AI helper for everything is expensive and always ends the same way: the agent does everything at a C grade. Every has four agents, and each one has exactly one area of responsibility.

That's the only way to measure the effect honestly. When an agent takes over one named process, you can count what that process used to cost in hours. When an agent "helps the team", there's nothing to count, and in six months nobody will remember why it was set up. We lay out the same logic in how to calculate AI agent ROI.

When an agent is unnecessary

Not every process on this list needs an agent. A daily growth report in a team of 20 is often covered by a schedule and a finished dashboard: no autonomy, no prompts, no risk of a hallucinated number. We covered the line between AI embedded in a process and a full agent separately, in when autonomy is unnecessary.

Internal routine is rarely the right first step anyway. Agents on prioritization and OKRs don't bring in money directly; they give the manager time back. The benefit is real but indirect. If revenue is what hurts, start with the front office — lead qualification, speed of first response, CRM hygiene. The basic read on that layer is AI agents for small business.

An agent won't formulate the strategy for you, it will apply the one you wrote down. Every could hand over prioritization precisely because the company's priorities existed in writing before the agent showed up.

How to test this on your team in two weeks

  1. Pick one management process that currently rests on one person. The most common candidates in an SMB: turning meetings into action items, the weekly report, distributing inbound requests.
  2. Write down the sources that person pulls the answer from. If "memory" or "ask Oleg" is on the list, that's your first task, and it isn't about AI.
  3. Check machine readability. The data has to sit in a system with an API, not in a PDF report and not in a chat thread.
  4. Describe the outcome to the agent, not the algorithm. One area of responsibility, clear limits on what it doesn't do.
  5. Measure one metric before and after: hours per process, or time from event to action.

If the metric hasn't moved after week two, the problem is almost always in step 2, not in the model.


Sources

Next step. Take the free AI audit in two minutes: it shows which of your processes are ready for an agent right now and which need the data sorted out first. Our lineup of front-office agents is on a separate page.

Published by Andrew Maryasov, founder of Grow2.ai — AI agents and AI consulting for small and mid-sized business. Grow2.ai is the AI arm of Auspex.

Frequently asked questions

Which process should we start with?

Take the one that meets three conditions at once: it repeats weekly, it currently rests on one person, and all the data it needs already sits in a system with an API. Every found four such processes, but the first agent is always one — launching two in parallel makes it impossible to tell which produced the effect.

Do we have to use Notion?

No. In the Every case Notion is where the company's linked data lives. Bitrix24, HubSpot, ClickUp or Google Workspace can play the same role. There's one critical requirement: the data has to be linked together and reachable through an API.

Can OKR planning be automated in a small company?

Yes, if the goals, metrics and owners already exist in structured form and not only in last quarter's deck. Every cut OKR planning down to two days. The agent speeds up alignment and data gathering; it doesn't invent the goals.

How do internal AI agents differ from front-office ones?

Internal agents give the team and the manager time back. Front-office agents affect revenue directly: response speed to a lead, qualification, CRM hygiene. Grow2.ai advises SMBs to start with the front office and take management routine in the next wave.

What stops a prioritization agent from working properly?

Disconnected data. If the strategy isn't written down, tasks live in chats and the launch calendar rests on verbal agreements, the agent will produce a confident list built on nothing. The state of your data is the main risk factor, not the choice of model.

What does building an agent like this cost?

Every didn't disclose the cost. In Grow2.ai's experience, the main line item in projects like this isn't building the agent — it's getting the data into usable shape. Grow2.ai pilots start at €1,800 with a contractual 14-day KPI: if it doesn't work, you don't pay.

Where can I read the original case?

"How We Run a 25-person Company on Four AI Agents" was published by Every on 9 April 2026, written by Katie Parrott. The full breakdown of each agent and the prompts themselves are behind Every's subscription.

AI agents for business — 2–3 emails a month

Breakdowns, cases and tools already working inside companies.

No spam. Unsubscribe in one click.