One AI Assistant or a Team of AI Agents? How to Tell Which Your Business Actually Needs
Every AI vendor now sells "agents," and lately the pitch has escalated. You're no longer offered one assistant, but a whole team of them: a sales agent, a support agent, a scheduling agent, an operations agent, all supposedly working together.
Some businesses need that. Many don't, and pay for complexity they never use. This guide is for owners and operations leaders who want to make that call without a computer science degree.
First, what is an "agent"?
An AI agent is software that doesn't just answer questions. It takes actions. It can book an appointment, update a customer record, send an email or raise a ticket, usually by connecting to the tools your business already uses.
A single-agent setup is one assistant with one set of instructions and access to a set of tools.
A multi-agent setup is several specialised assistants, each with a narrower job, passing work between them or coordinated by a "manager" agent.
The honest debate among the people who build these
What vendors rarely mention is that the AI companies themselves disagree about this.
In mid-2025, Anthropic, the company behind Claude, published how it built its multi-agent research system. One lead agent hands sub-questions to several helper agents that research in parallel. On Anthropic's internal research tests, it beat a single agent by about 90%. The catch: it used roughly 15 times more computing than a normal chat. Anthropic was clear that it pays off for complex tasks worth that cost, and is wasteful for simple ones.
Around the same time, the team behind the AI coding product Devin published an essay titled "Don't Build Multi-Agents." Their argument: when agents hand work to each other, context gets lost. Each agent makes reasonable decisions based on a partial picture, and the combined result contradicts itself.
Both are right, about different problems. That's the useful insight for a business owner: the answer depends on the shape of your work, not on which approach sounds more advanced.
When one assistant is enough
A single agent is usually the right choice when:
The job is one connected conversation. Answering customer questions, qualifying a lead, booking an appointment.
It touches a small number of systems. Your CRM and calendar, say, rather than eight different platforms.
The steps follow one after another. First this, then that, rather than lots of things happening at once.
A mistake is easy to spot and fix.
For most small and mid-sized businesses starting out, this covers the first project. It is how most AI automation for small business should begin. A single, well-built agent is cheaper to run, easier to test and much easier to fix when something goes wrong.
When you genuinely need more than one
Multiple connected agents start to make sense when:
Different parts of the work need different rules and permissions. The agent taking payments shouldn't have the same access as the one answering general questions.
Work happens in parallel, such as researching many sources at once, or handling many customer conversations that all feed one system.
The work spans departments. A customer call creates a booking, which triggers a payment, which updates records, which might need a human follow-up.
You're planning your third, fourth and fifth AI use case, not just your first.
That last point is the one most businesses miss.
The real question: point solutions or a shared foundation?
Here's a pattern I see often in our client work. A company builds a support agent. Later, a lead qualification bot. Each works well on its own. The company considers AI "done."
Then they try to add a third, and find each agent was built as a standalone island: its own connection to the data, its own copy of customer information, its own rules. Nothing is shared, so the third agent costs as much as the first, and the same customer is now recorded slightly differently in two places.
The alternative is to build agents on a shared foundation: one place where customer and business data lives, one set of safety rules, one way of logging what every agent does. In our experience, the cost picture flips around the third or fourth use case. With a foundation in place, each new agent costs a fraction of the first. Without one, every new agent starts from scratch.
So the better question isn't "one agent or many?" It's: "Are we building one-off tools, or something future agents can build on?" One agent is a fine way to start. The second one is the moment to decide.
What this looks like in practice
We built an AI system called Veda for a UK healthcare administration company that runs the operations of 40 doctors. On the surface it's one thing: an AI voice assistant for clinics that answers patient calls for all 40 doctors at any hour.
Underneath, it's several connected parts sharing one database: the voice agent, a practice management system holding doctors, patients and diaries, schedule controls, patient forms, payments, and a ticketing system where staff pick up anything the agent can't finish. Every part reads from and writes to the same records. Nobody copies information between screens.
The results: 85% of routine bookings need no human involvement, booking accuracy is 99.2% (up from a 12% error rate under manual scheduling), and administrative labour costs fell 40%.
The voice agent was the easy part. The foundation underneath made it work.
Three warnings before you commit
1. Agents don't give the same answer every time. That's how the technology works, not a fault. With several agents, small variations can add up. Each agent needs clear limits, checks on its output before it acts, and a log of what it did.
2. Test somewhere safe first. Agents act in the real world: emails get sent, records get changed, payments get triggered. We've been called in to clean up after agents that wrote bad data into a live CRM during what was meant to be a test. Always run new agents against a copy of your systems before the real thing.
3. Expect months, not weeks, for a connected system. A single focused agent can go live in a few weeks. A properly connected multi-agent system usually takes two to three months, because most of the work is testing against real situations and tuning how agents hand off to each other and to your staff. Anyone promising a full multi-agent system in two weeks is showing you a demo.
A simple way to decide
Answer these honestly:
How many separate systems would the AI need to work with?
Does the work happen in one conversation, or across several departments?
Do different steps need different levels of access?
Do you already know your second and third AI use case?
Mostly "few, one conversation, no, not yet": start with one well-built agent. Mostly "many, several departments, yes, yes": plan for connected agents on a shared foundation from the start.
