By Piet Baudoin · September 2026
Not necessarily, and nobody can say for sure yet: there is no public measurement for businesses of ten to fifty people. The evidence for coordinators comes from companies with a billion dollars in revenue, and only 29 percent of them say that managing agents is fully part of their work. What the numbers do show is where a loose swarm gets expensive. The makers building for everyone still choose a team: xAI with Grok Bot, OpenAI with Dots. Below, I work through the numbers with Microsoft’s and Salesforce’s prices. What AI agents take over in small businesses is covered in the overview of the agentic economy.
What is a coordinator over several AI agents, and who is it built for?
It is one AI agent that directs the others. It decides, step by step, whose turn it is and when a human needs to step in. The marketing term for this is agentic orchestration. You will run into that phrase.
That is how Aide describes it in September 2026. IBM sells that same layer and recommends locking everything down in advance once predictability matters. At that point it is no longer a swarm, but a fixed route.
UiPath asked 590 companies with more than a billion dollars in revenue about this on 9 September 2026. Of the group that has fully embedded orchestration, 89 percent hits the expected return or more. But only 29 percent of those 590 companies are at that point. UiPath sells orchestration software itself.
Now the other end of the market. Among UK small business owners, 44 percent had heard of agentic AI as of 16 September 2026, and 5 percent used it regularly.
What a single AI agent is, you can read in the piece on what an AI agent actually is.
What does that cost per month with twenty people?
$160 to $500 a month just for the usage meter, depending on who you buy from. The prices come from the vendors; the volumes are my assumption, a worked example.
Microsoft sells Copilot Studio per company, in packs of 25,000 credits for $200 a month; one action costs 5 credits. Salesforce charges 20 Flex Credits per action, and $500 per 100,000 credits.
Twenty employees, each with one task per workday: a supplier invoice comes in, gets matched to the right project, and is ready to book. One AI agent does that in a single pass. Put a coordinator on top, and it becomes four steps that are each billed separately: the coordinator hands the invoice to an agent, that agent reads it, looks up the project, and reports back what it found. Microsoft and Salesforce both call a billed step like that an action. Twenty people, twenty workdays, four actions: 1,600 a month.
At Microsoft, that is 1,600 times 5 credits, so 8,000 credits. You buy in full packs of 25,000, so you pay $200 and use just under a third of it.
At Salesforce, it is 1,600 times 20 credits, so 32,000 credits. Pro rata, that is $160, since 100,000 credits cost $500. Whether Salesforce sells pro rata or only in blocks of 100,000 is not stated on its pricing page. If you have to buy the full block, it is $500.
Microsoft also shuts the agent off at 125 percent of your prepaid balance. Pinning down that kind of bill in advance is covered in what an AI agent costs per month.
Why does the bill grow faster than the work?
Because every agent rereads everything before it does anything. That rereading is billed per piece of text, and a piece like that is called a token. That is how one night of running can produce a bill nobody saw coming.
A builder posted his measurement on OpenAI’s forum on 16 September 2026. Two agents, one night, 158,137,319 tokens. Of that, 95.4 percent was rereading what was already there. The coordinator asked to wait 89 times, and 66 times nothing came back. That night cost $99.80 at list prices.
You are not paying for the thinking, you are paying for the reading. One agent that maintains a file reads it once. A coordinator with three agents has the same file read four times, and all four reads land on your bill.
One AI agent or a swarm, what is the difference?
In September 2026, two vendors said the opposite, each for a different customer. Salesforce wrote on 11 September that its agents “pursue goals across days and weeks instead of completing only a task or interaction.” Anthropic wrote on 15 September about its package for small businesses that it “waits for your approval before anything sends, posts, or pays.”
Of the sign-ups for Anthropic’s tour of small businesses, 80 percent run a business of 5 to 50 people. That group gets one assistant that waits, not a swarm.
So the difference lies not in what is possible but in the customer. Why that approval belongs there is covered in the piece on sending email without approval.
| Aspect | One AI agent with approval | Agents under a coordinator |
|---|---|---|
| You pay | A fixed amount per month | Per action, including every handoff |
| Overview | One list with reasons | A log per agent |
| Responsible | Whoever clicks Approve | Whoever set up the coordinator |
| When it goes wrong | A wrong proposal sits there | Work that has to be redone |
What happens to the bill if this pace continues?
First, the pace, in three data points. OpenAI’s heaviest model went from $4 and $20 per million tokens on 9 July 2026 to $10 and $50 on 3 September 2026, two and a half times higher in two months. That same model does work faster: on OpenAI’s computer-use benchmark, 72.6 percent correct in about 40 minutes per task, against 65.7 percent in about 75 minutes. And on that night of 16 September 2026, 95.4 percent of usage went to rereading.
My hypothesis: through the end of 2028, a swarm of agents will get more expensive per completed task for a business of ten to fifty people, not cheaper, and the real gain sits in a good memory, not in the number of agents.
The reasoning in three steps. The strongest model keeps getting more expensive per piece of text. Working faster brings usage down, but most of the bill does not go to thinking, it goes to rereading and waiting. Every agent you add reads the file one more time, so that line item grows with the number of agents, not with the work.
Two things would break this hypothesis: billing per completed task instead of per piece of text, and a steep discount on text that has already been read once.
For your business, that means the following. What gets expensive is a loose swarm in which every agent rereads everything and hands work on without anyone checking. What stays affordable is a small team that shares one memory of what you have approved and has one approval point with you. An extra agent earns its place where there is a clear handoff in the work, such as from quote to follow-up, with someone on your side signing off on the result. If you cannot point to that handoff, you just add a new place for things to go wrong, and the question of who is liable when an AI agent makes a mistake.
Why do the makers choose several agents anyway?
Because the data does not settle it, and the makers who build for everyone are choosing a team. xAI launched Grok Bot on 11 August 2026 and wrote: “Instead of multiple agents you have to manage, Grok Bot gives you a small team that can work in parallel so you’re not the middleman.” Inside the company, according to xAI, people “often run multiple Bots in parallel, with one to manage the others.” OpenAI launched Dots on 29 September 2026: you start with one Dot, and later OpenAI wants whole teams of Dots to work together. Neither company does that for nothing.
Bombos does not add agents of its own. The cost problem above shrinks when agents share one memory instead of each rereading everything at every handoff, and that is what Bombos gives them: it is the memory layer under the agents you already use, one plugin for Claude, Claude Code, ChatGPT, Copilot, Grok, Codex, Cursor and your own agents. Every agent works from the same memory, and every answer shows what it rests on.
What can this not do today?
There is no public measurement that pits one AI agent against several agents for a business of ten to fifty people. That cuts both ways: it does not prove that a single worker is better either. Everything comes from vendors, or from companies with a billion dollars in revenue. My worked example rests on an assumption: if your number of actions is twice as high, the bill doubles.
The vendors are not there yet either. n8n sells agents and wrote on 5 August 2026 that full orchestration does not exist yet. Zapier automatically pauses a run once it passes 75 tasks, and asks you to take a look before it continues.
A builder of systems like this, the founder of Creatify and so a vendor himself, wrote on 20 September 2026 about tasks that looked completed in his log but had never actually been done. That is the risk at twenty people: not that the swarm stalls, but that it looks finished.
How do you start small with this?
With one task that comes back every day, where you see the result before it goes out. First count how often that happens and how much time it takes, because that number is your yardstick.
Bombos sits underneath that task: one shared memory for every AI and agent on the team, with the team’s knowledge, rules it learns from corrections, and what already sits in email, documents and the software you already use.
That memory is what my hypothesis turns on. What Bombos learns about your business does not sit in the model: your correction becomes a rule for next time, for every agent. When a better model comes along, the memory stays in place.
Bombos runs in the cloud or standalone on your own server. One person can try the plugin first for their own agents; for a team, the pilot costs €3,000 €2,000 for three months, then €1,000 per month, excluding VAT.
One task you can check the math on is worth more than a swarm you have to take on faith.
Sources
Every source was opened on 20 September 2026, the last group on 30 September 2026, and every quote appears in it word for word.
What it is
- Aide, What is Agent Orchestration? (updated September 2026)
- IBM watsonx Orchestrate, Multi-Agent Orchestration (live product page, read on 20 September 2026)
What vendors charge
- Microsoft 365 Copilot Studio, pricing (live pricing page, read on 20 September 2026)
- Microsoft Learn, Billing rates and management (14 September 2026)
- Salesforce Agentforce, pricing (live pricing page, read on 20 September 2026)
- Salesforce, Agentforce: Job-Ready AI Agents (11 September 2026)
- Zapier Help, Migrating from Agents to AI by Zapier (7 September 2026)
- n8n, Introducing n8n Agents (5 August 2026)
- OpenAI, developer pricing (live pricing page, read on 20 September 2026)
What was measured
- UiPath, AI adoption and orchestration survey 2026 (9 September 2026)
- OpenAI, Introducing GPT-5.6 (9 July 2026)
- OpenAI, GPT-6 Astra (3 September 2026)
- OpenAI Community, Two agents, one night: 158 million tokens (16 September 2026)
- X post by @AndreyK09474778 on handoffs without confirmation (20 September 2026)
What is delivered to small businesses
- Business Matters, on Small Business Britain and Alibaba.com (16 September 2026)
- Anthropic, What 1,000 small business owners taught us about AI (10 September 2026)
- Anthropic, Claude for Small Business launches new workflows, integrations and training programs (15 September 2026)
Why the makers choose a team
- xAI, Introducing Grok Bot (11 August 2026)
- CNBC, OpenAI DevDay 2026 live updates (29 September 2026)
- TechCrunch, OpenAI launches Dots (29 September 2026)
Bombos team