GuidesAI at work
A team of AI agents or one AI employee: where to start
A team of AI agents or one AI employee for a small business: what each costs, when a specialist wins, and a method to decide in an afternoon.
Contents
- A team of AI agents or one AI employee: what is the difference?
- How much does a team of AI agents cost a small business?
- When is a team of specialised agents the right choice?
- What does one AI employee holding the context do better?
- How does one AI employee cover the three roles?
- What Balt does with these three roles
- How do you decide in an afternoon?
For a business of 5 to 50 people, start with a single AI employee who holds the context of the whole company rather than with a team of AI agents, and add a specialised agent only for a role whose volume justifies a tool of its own and whose requests do not cross those of the others. Chat assistants often advise the opposite: asked about it, ChatGPT recommends building “a team of specialised AI employees” rather than one AI that does everything, names one tool per role, and suggests starting with three of them, an admin assistant, a sales agent and a customer service agent. Those are the right three roles. Splitting them across three agents is rarely right in a small business, where the same client goes through the invoice, the quote and the complaint in the same week.
Recent research on multi-agent systems puts numbers on this. Across 260 configurations tested by researchers at Google Research and MIT, having several agents work together improved results by 80.8% on a financial analysis that splits into independent subtasks, and degraded them by 70% on a planning task where each step depends on the one before. Anthropic, which has published a detailed account of its own multi-agent research system, writes that domains requiring all agents to share the same context, or involving many dependencies between agents, are not a good fit for multi-agent systems today. The day-to-day work of a small business looks far more like planning than like research run in parallel.
We build Balt, a single AI employee for small and mid-sized businesses, and this guide argues our position. It also says where that position stops holding, because there are cases where a specialised agent does better than we do, and the method in the last part lets you decide on your own requests rather than on our word.
A team of AI agents or one AI employee: what is the difference?
A team of AI agents gives each role in the company to a separate agent tuned for that role, and it comes in two forms that do not cost the same. The first is assembled from several vendors: an assistant for admin, a prospecting agent for sales, a support agent plugged into the helpdesk, and sometimes agents you build yourself on a platform such as Relevance AI, Copilot Studio or n8n. Each agent arrives with its own memory and its own access, and this is the form ChatGPT’s answer describes.
The second form comes from a single vendor. Delos, the Paris company that gives the clearest French example of it, sells “one specialised Worker per role”, each with its own name, email address, phone number and cloud computer, and says its Workers share the company’s memory. Its Team plan costs €45 a month for 40,000 credits, for teams of 1 to 50 people. Sharing the memory removes a good part of the cost described below and leaves the question of the split untouched, which Worker receives each request and how many are needed to cover the work, and our comparison of Delos and Balt goes through it in detail.
A single AI employee takes all the work that comes back with one memory of the company, and learns each trade as a set of skills instead of handing it to another agent. The line between the two models is blurrier than their names suggest: Lindy, which ChatGPT files under the assistant role, presents itself as an AI employee and describes seventeen roles it can fill, as our comparison with Lindy shows. The products on the market are reviewed in our selection of the best AI employees for a small business, and this guide is about choosing the form.
How much does a team of AI agents cost a small business?
A team of agents is paid for role by role, and the subscription is the only part of the cost that appears on a price list. Here is what the tools ChatGPT names for each of the three roles showed in early November 2026, along with the platforms used to build your own agents.
| Role | Example | What is billed | Entry price found |
|---|---|---|---|
| Admin assistant | Lindy | a seat for each person who uses it | $29.99 per person per month |
| Sales agent | 11x (Alice) | an annual subscription, by volume of prospects | $3,750 a month billed annually, for 2,000 new prospects a month |
| Sales agent | HubSpot Prospecting Agent | each recommended lead, in HubSpot credits | $1 per lead, from HubSpot’s Starter edition according to its product page |
| Customer service | Intercom Fin | each resolved conversation | $0.99 per resolution, minimum 50 a month on your own helpdesk |
| Customer service | HubSpot Customer Agent | each resolved conversation, plus a Service Hub Professional seat | $0.50 per resolution, seat from $90 a month billed annually, mandatory $1,500 onboarding |
| Agents to build yourself | Relevance AI, Copilot Studio | actions or credits, depending on the platform | $19 a month at Relevance (Pro billed annually, per its documentation), $200 a month for 25,000 Microsoft credits |
| Automations | n8n | each execution of a workflow | €50 a month for 10,000 executions (Pro, billed annually) |
The gaps are wide, and they say mostly who each tool was built for. The 11x Growth plan, $45,000 a year, covers five users at most, with managed sending mailboxes and their warm-up: it is a sales team’s budget, which our comparisons with 11x, with Fin and with HubSpot’s Breeze agents go through tool by tool. What these lines do not show comes down to four items, paid for in time rather than in dollars.
The company’s context, re-explained to each agent
Each agent from a different vendor keeps its own memory, so what you taught one, the other does not know. The grace period granted to a loyal client or the dispute under way with a supplier has to be written into every agent that needs it, then corrected in each of them on the day it changes. Companies already run twelve agents on average, half of them working in complete isolation, and we explained with those twelve agents that do not talk to each other why the right criterion is the number of places where your context is copied. Each copy ages at its own pace, and the drift between two copies never announces itself.
HubSpot makes the same argument in defence of its own agents. Announcing their outcome-based pricing, the vendor writes that generic AI agents know a lot about the world but not about yours, and that its agents are reliable because they live in the CRM with each client’s data and history. The argument holds, and it holds in proportion to what really lives in the CRM: when part of the client sits in the accounting software, in a partner’s inbox or in the support tool, each agent knows only a piece of it.
Access to keep track of for each agent
Each agent receives its own access, the mailbox for the assistant, the CRM and sending mailboxes for prospecting, the helpdesk and the knowledge base for support. Each of these sets of rights has to be reviewed when someone changes role and closed on the day the tool is dropped, as many times as there are agents. The rule we apply, an agent that acts with the rights of the person asking rather than with an account of its own, is harder to hold when each agent was installed by a different team with whatever admin account it had to hand.
Hand-offs between agents
Real work crosses roles, so a team of agents spends part of its time passing files to one another, and these seams are where multi-agent systems fail most often. A team at UC Berkeley annotated more than 1,600 execution traces from seven open-source multi-agent systems, on programming, maths and general assistance tasks, and found failure rates of 41% to 86.7% depending on the system. Nearly a third of the failures they recorded come from misalignment between agents, such as an agent that carries on with a wrong assumption instead of asking for clarification, or crucial information that one agent keeps to itself.
The Google Research and MIT researchers measure the same fragility from another angle. Agents that each work on their own, with no central check, amplify errors by a factor of 17.2 compared with a single agent, against 4.4 when a coordinator verifies their work, and tasks that call on many tools, such as a sixteen-tool business workflow, pay a coordination overhead. Their explanation is useful to a manager: a single agent keeps one continuous memory stream, whereas several agents have to compress the overall context into the messages they exchange, and that compression loses information.
Cognition, the company behind the software agent Devin, drew two principles from this that it considers important enough to rule out by default any architecture that breaks them: share the full context rather than a few messages, and remember that every action carries implicit decisions, which two agents that cannot see each other will make in contradictory ways. This research concerns agents within a single system, designed to cooperate. Between two agents from two different vendors, which share neither memory nor a format for exchanging work, there is no reason to think the seam is any narrower.
The upkeep of each agent
A specialised agent is set up at installation and then maintained every time the company changes, and someone at your company has to answer for it. 11x counts two weeks of onboarding on its Growth and Pro plans, HubSpot charges $1,500 for the mandatory onboarding of Service Hub Professional, and Relevance AI describes two weeks to identify the right use cases before building a first team of agents with its deployment team, as our comparison with Relevance AI details. On Copilot Studio, at $200 a month for 25,000 credits on top of Microsoft 365 Copilot licences at $30 per user, the real cost is the person who writes each agent’s business rules and keeps them current, which our comparison with Copilot puts into figures. In a fifteen-person business, that person already has another job, and every extra agent makes their list longer.
When is a team of specialised agents the right choice?
A specialised agent wins when a single job carries enough volume that a tool built for it, with its own billing and infrastructure, returns more than it costs in seams, and when that job’s requests stay in their lane. This is what the research measures, since the gains of multi-agent systems are concentrated on tasks that split into independent pieces. Four situations meet these conditions in a business, and in each of them a specialised tool does work that Balt does not try to do.
Customer support at scale
A support agent built for the job answers live on chat, email or WhatsApp, plugs into a helpdesk and is billed per resolved conversation, which ties its cost to volume. Anthropic’s head of AI support says, on Intercom’s pricing page, that Fin resolves around 560,000 conversations a month there, at a resolution rate of about 79%, and HubSpot states that its Customer Agent resolves 65% of conversations across more than 8,000 customers, two figures published by the vendors themselves. At that volume, the questions come from thousands of customers and find their answer in a public knowledge base, so their context overlaps little with the rest of the company.
An agent that answers live cannot be reviewed sentence by sentence. What a person approves in advance is its scope, the questions it handles alone and what it never promises, as we describe for an agent that answers the phone, and since 2 August 2026 the European AI Act requires that the customer knows they are talking to an AI.
High-volume outbound prospecting
Prospecting at scale needs infrastructure that a general AI employee does not try to provide: contact data, managed sending mailboxes, their warm-up and their rotation so that messages land. The 11x Growth plan includes 2,000 new prospects a month with those mailboxes, two-way sync with your CRM and meeting booking. At that pace, approval no longer happens message by message, it covers the targeting and the sequences, and that is a decision to take with your eyes open.
Until the prospect replies, their context is separate from that of your clients. The seam appears with the first reply, when the file passes to the salesperson who will follow it, and that is the moment by which you need to have decided which memory is authoritative on them.
Deterministic automations
A flow that has to give the same output on every run, copying an order from a form into the management software or alerting a channel when an invoice is paid, belongs in an automation tool. n8n runs the workflow someone built, node by node, and its Pro plan includes 10,000 executions for €50 a month billed annually; on that work it does better and costs less than an agent, as our comparison with n8n acknowledges. A workflow remembers nothing about your clients and therefore adds no copy of your context, and Balt calls your n8n workflows like any other tool he has wired.
Research that splits apart
Research that explores many independent leads at once, such as finding the board members of every technology company in the S&P 500, benefits from being spread across several agents working in parallel. Anthropic’s multi-agent research system did 90.2% better than a single agent on its internal evaluation, at a matching cost: a multi-agent system uses about fifteen times more tokens than a chat. Those subagents live inside a single tool, which starts and stops them itself, and they leave you no team to manage.
What does one AI employee holding the context do better?
In a business of 5 to 50 people, admin, sales and customer service concern the same clients, and an agent’s value lies first in what it knows about each of them at the moment it acts. A few ordinary situations show this better than an argument.
The follow-up that depends on the last admin exchange. A client due a follow-up on their quote disputed an invoice by email the day before. A sales agent that cannot see the accounts inbox chases as if nothing had happened, and the client receives the quote follow-up and the reply to their complaint on the same day. An employee who holds both sees the dispute and holds back the follow-up, or offers it to you with a sentence that takes the dispute into account.
The quote that needs both the CRM and the accounts. Pricing a new order means knowing what the client has bought before, the discount granted last time and how regularly they pay. The first fact lives in the CRM, the last in the accounting software, and a sales agent connected to the CRM alone proposes payment terms the bookkeeper would never have accepted.
One place to approve what goes out. Each specialised agent has its own approval rules, set by the person who installed it, and they differ from one vendor to the next. With a single employee, the rule is written once and holds for everything that leaves the company, whether a follow-up or a reply to a client, and it follows the action rather than the agent, as we explain in describing how two agents that know you can coexist.
Research on context loss points the same way, with a caveat we owe you. A study published in May 2025 on more than 200,000 simulated conversations shows that the best models lose 39% of their performance on average when information reaches them piece by piece over a conversation rather than all at once, and that a model that takes a wrong turn does not recover. On sales, service and quoting tasks, Salesforce’s CRMArena-Pro benchmark measures success of around 58% when the request arrives in one go, and around 35% when it builds up over several exchanges.
Both studies concern a conversation between a person and an agent. We draw an inference from them, and present it as one: a hand-off between two agents is a request restated by a third party, and there is no reason to think the second agent receives it more completely than a user who gives it piece by piece.
A single employee has its own limits too, and we have written them down before. It is harder to improve, because an improvement has to hold across every case at once, and it is less impressive in a demo than an agent that is very sharp on a narrow task.
How does one AI employee cover the three roles?
Take the three agents ChatGPT proposes and look, role by role, at what a single employee takes on and at the sign that a specialist is worth adding.
| Role | What a single AI employee takes on | The sign that a specialist is worth adding |
|---|---|---|
| Admin assistant | sorting the inbox, draft replies, invoice follow-ups, the calendar, data entry, meeting notes | rarely, because this role touches every client and every person, and its context is the most shared |
| Sales agent | quote follow-ups, meeting preparation, CRM updates, following existing clients | outbound prospecting to strangers, at volume, that needs its own sending mailboxes |
| Customer service agent | replies to clients you know, drafted then approved, and following complaints with their history | a volume of conversations too large to review, answered from a public knowledge base |
The first role is where separation costs the most, since admin sees everything that reaches the company. The other two separate from a certain volume onwards, and the order matters: a single employee first, on the requests that cross roles, then a specialist where the volume has shown itself, with a hand-off rule written before it is installed.
What Balt does with these three roles
Balt is an AI employee for small and mid-sized businesses, and he takes on all three roles with a single memory. He lives in Teams, Slack, WhatsApp and his web app, has his own email address, writes by SMS, and on the phone he calls, answers and transcribes every call, handling alone the calls you have entrusted to him. He works in the company’s tools with the access he is given, among more than 3,200 integrations, from the mailbox to the CRM and the accounting software.
What he learns with one person he writes into a skill the whole team shares, which you can read in the console, so a rule given by the bookkeeper holds for the follow-up prepared the next day at a salesperson’s request. Recurring work becomes a routine requested in one sentence, and a routine that fails five times in a row stops instead of retrying endlessly. Nothing he prepares leaves the company until a person has approved it, and no setting can relax that rule. He is billed in credits shared by the whole team, with no seats and no cap on users, from €99 a month for 40,000 credits, with 20,000 credits offered at sign-up and no card required, and his data is hosted in France and the European Union.
He does not take on what the specialists described above do better. Balt is not a helpdesk billed per resolution for thousands of conversations a month, and he provides neither a contact database nor managed sending mailboxes for mass prospecting. His SOC 2 Type 2 process is under way and the certification has not been obtained to date, whereas Lindy, 11x, Relevance AI and Delos display one. He works with what stays separate: he calls your automation workflows, and he can consult an expert agent one question at a time as long as it keeps nothing of the exchange, which is the rule set out in our article on agents that call other agents.
How do you decide in an afternoon?
The method starts from your real requests rather than from vendor catalogues, and it takes four hours, with a notebook open next to the inbox, the CRM and the helpdesk.
- List thirty requests, in an hour. Take the last thirty requests that touched a client: a quote, an invoice reminder, a complaint, preparing a meeting, a question about a delivery. Next to each one, write the right answer, the one you would have given, because this is also the set of cases you will use to test the tools, as in our method for testing an agent before you sign.
- Mark the roles each one crosses, in half an hour. For each request, tick admin, sales or customer service according to what you need to know to handle it properly, regardless of who received it. Count those that tick at least two.
- Count the volume of each role, in half an hour. Pull from your tools the number of support conversations last month, the number of new prospects you would like to contact each month and the flows that repeat identically, such as copying an order from one piece of software to another.
- Do the sums for each role, in an hour. A role deserves its own agent when most of its requests stay in their lane and its volume pays for the tool at the published price: 300 resolved support conversations a month would cost $297 at Fin’s rate, for example, to set against the time your team spends on them today. You also need a person at your company who agrees to answer for it, or the agent ages unwatched.
- Write the hand-off rule, in half an hour. For each specialist you keep, note which memory is authoritative on a client, at what point a file passes from one to the other, and who approves what goes out, since approval follows the action.
- Choose the first task, in the half hour that remains. The requests that tick two roles or more go to a single AI employee, who takes them on for three to four weeks, the time it takes the team to get into the habit of handing them over, before you count again.
Every request that ticks two roles would become a hand-off in a team of agents, and if there are many of them in your list, a single employee is the right starting point. If nearly all of them stay within one role and that role carries a heavy volume, the specialist is justified for that role, and the single employee takes the rest. The next question is the product, and it is settled by having the candidates work on those same thirty requests, whose right answers you wrote down before the first trial.
Frequently asked questions
Which AI team should a small business build first?
No team to begin with: one AI employee who takes the admin work, the sales follow-up and the replies to clients, because in a business of 5 to 50 people those three roles concern the same clients. A specialised agent comes next for the role whose volume justifies it, most often support at scale or high-volume outbound prospecting, once you have decided which memory is authoritative on each client.
How much does a team of AI agents cost?
In early November 2026, the 11x sales agent started at $3,750 a month billed annually, Fin cost $0.99 per resolved conversation with a minimum of 50 a month, the Lindy assistant $29.99 per person per month, and HubSpot’s Customer Agent $0.50 per resolution plus a Service Hub Professional seat from $90 a month. On top of those subscriptions comes a cost no price list shows: the context to re-explain, the access to keep track of and the upkeep of each agent.
Can one AI employee handle admin, sales and customer service?
Yes for the work that comes back in a small business: sorting email, chasing invoices and quotes, preparing meetings, updating the CRM, and replies to clients drafted then approved. It does less well than a specialised tool on a volume too large for anyone to review, such as thousands of support conversations a month or outbound prospecting that needs its own sending mailboxes.
When should you add a specialised AI agent?
When most of a role’s requests stay within that role without touching the others, when its volume pays for the tool at the published price, and when a person at the company agrees to answer for it. Before installing it, write down which memory is authoritative on each client and who approves what goes out, since approval follows the action and not the agent that prepared it.
Can an AI employee work with specialised agents?
Yes, as long as their contexts do not overlap. An automation workflow remembers nothing about your clients and can be called like a tool, and an outside expert agent can be consulted one question at a time if it keeps nothing of the exchange. Two agents following the same clients would hold two versions of them, and the gap between the two never announces itself.
Sources
- Kim et al. (Google Research, MIT), Towards a Science of Scaling Agent Systems, arXiv 2512.08296, version 3 (April 2026)arxiv.org
- Anthropic, How we built our multi-agent research system (13 June 2025)anthropic.com
- Cemri et al. (UC Berkeley), Why Do Multi-Agent LLM Systems Fail?, arXiv 2503.13657, version 3 (October 2025)arxiv.org
- Cognition, Don’t Build Multi-Agents (12 June 2025)cognition.ai
- Laban et al., LLMs Get Lost In Multi-Turn Conversation, arXiv 2505.06120 (May 2025)arxiv.org
- Huang et al. (Salesforce AI Research), CRMArena-Pro, arXiv 2505.18878 (May 2025)arxiv.org
- Belitsoft, 2026 AI agent trends: enterprises run 12 AI agents on average but half work alonebarchart.com
- 11x, Alice pricing page (read 4 November 2026)11x.ai
- Fin, pricing page (read 4 November 2026)fin.ai
- Intercom, pricing page (read 4 November 2026)intercom.com
- HubSpot, Customer Agent and Prospecting Agent: now you pay when the task is complete (13 April 2026)hubspot.com
- HubSpot, Service Hub pricing (read 4 November 2026)hubspot.com
- HubSpot, Prospecting Agent (read 4 November 2026)hubspot.com
- Lindy, pricing page (read 1 November 2026)lindy.ai
- Lindy, roles it can fill (read 1 November 2026)lindy.ai
- Relevance AI, documentation: pricing (updated 28 September 2026)relevanceai.com
- Microsoft, Copilot Studio pricing (read 4 November 2026)microsoft.com
- n8n, pricing page (read 1 November 2026)n8n.io
- Delos, pricing page (read 4 November 2026)delos.so
- Delos, AI Workers catalogue (read 4 November 2026)delos.so

