Every product has a moment when its makers realize they have been solving the wrong problem, and solving it very well.
Ours began with a conversation that looked completely ordinary.
A customer messages a business on WhatsApp. They ask about a jacket. Then whether it comes in a larger size. Then what it costs with shipping. Then, almost as an afterthought, where the order they placed last week has gone. Finally, with a little less patience, they ask to speak to a person.
Five questions. Four different jobs. One thread.
For a long time, our answer to that thread was a single AI assistant. We trained it, tuned it, fed it more knowledge, and watched it get better. It did get better. But something stubborn remained. The more we tried to make one assistant excellent at everything, the more clearly we saw that “everything” was the problem.
This is the story of how we came to that conclusion, what we built because of it, and what the process taught us.
The ceiling
Think about the people who work at a good business. The salesperson is warm, persuasive, and fluent in the catalog. The support specialist is patient, methodical, and calm under pressure. The person who handles orders is precise and lives in the system of record. The receptionist knows when to schedule and when to escalate.
Nobody expects one human to be all four at once, at every hour, in every language, across every channel. We would call that a staffing failure, not a talent.
Yet that is exactly what we ask of a single AI assistant. One set of instructions. One pool of knowledge. One personality that must be persuasive in one sentence and procedural in the next.
We kept running into the same ceiling from different directions. Make the instructions broader, and the assistant became vague. Make them stricter, and it became rigid. Add more knowledge, and relevant answers had to compete with irrelevant ones. Every improvement in one direction taxed another.
A ceiling like that is not a bug to be patched. It is a signal that the architecture is asking for a different shape.
A year of asking better questions
2026 turned out to be a year of that kind of question for us.
110+
updates shipped this year
85%
fewer tokens used by AI Training, roughly
We redesigned the interface so the product felt calmer and more coherent. We rebuilt the Flow Builder, so that creating an automation felt less like operating a machine and more like drawing an idea. We rethought AI Training and cut its token consumption dramatically. We strengthened security, deepened the API, sped up the platform, and shaped the Shared Inbox into a place where a conversation and a customer’s full context live side by side.
Each of those began as a small, almost impolite question. Why does this take five steps? Why can’t I see this here? Why is this slow? Why does this cost so much?
Asking those questions gets less comfortable over time, because the easy answers run out. And somewhere in the middle of the year, the questions stopped being about features and started being about foundations.
What is an AI assistant actually for?
Back to first principles
We tried to answer from the conversation itself. We looked at that jacket, that order, that request for a person, and asked three things: what does a customer actually need, what does a business actually need, and what would a great human team do?
The customer
needs the right answer, quickly, from someone who already knows what was said a minute ago. They do not want to repeat themselves. They do not want to be passed around. And when the situation is delicate, they want a human, immediately and without a fight.
The business
needs those same outcomes, plus control. It needs to decide what the AI may say, what it may do, what it must never do, and when it must step aside.
A great human team
specializes. It hands work over cleanly. It shares context. It knows its limits.
Put those together and the shape of the answer becomes almost obvious. Not a smarter assistant. A team.
Designing the team
So we built Multi-Agent AI around a simple idea: give every job its own agent.
- A Sales Agent that answers product questions, recommends from your catalog, and recognizes buying intent.
- A Lead Qualification Agent that asks for requirements one at a time, saves each answer, and starts the right follow-up.
- A Support Agent that resolves issues from your own knowledge and knows when a case is beyond it.
- An Order Agent that checks live order status and explains it in plain language.
- An Appointment Agent that captures the service, the date, and the time.
- A Human Handover Agent whose entire purpose is to recognize the moment a person is needed.
None of these is a template we impose. They are examples of what you can create, in any number and with any responsibilities. The point is that each agent is small enough to be excellent.
The brain: a system prompt
Every agent is governed by its own system prompt. This is where you tell it who it is, how it should speak, what it must never invent, and, crucially, what should happen when a condition is met.
A prompt in BotSailor does not only shape replies. It shapes behavior. Type ## and the available actions appear, and you can write, in plain language, that when a customer confirms they want to buy, the agent should tag them as an interested buyer and start your sales follow-up sequence. The words become operations.
The hands: real actions
An agent that can only talk is an advisor. We wanted workers.
Agents can add and remove labels, start and stop sequences, and save what they learn into custom fields. They can trigger the flows you have already built, call any HTTP API you have connected, and look up live Shopify and WooCommerce orders. They can set follow-up reminders, leave internal notes, mark a conversation solved, and stay silent when silence is the right reply.
The memory: one conversation, many specialists
The detail we are proudest of is the least visible. When a customer moves from a sales question to an order problem, the next agent does not begin from zero. Names, requirements, selected products, saved fields, order numbers, and lead status travel with the conversation. The customer experiences one continuous exchange, not a relay race in which every runner forgets the baton.
The knowledge: focused, not flooded
Each agent draws on its own knowledge, and that knowledge can come from many places together: written content, web pages, files such as PDFs and spreadsheets, product images with descriptions, Google Sheets, and connected APIs. A Sales Agent learns your catalog and pricing. A Support Agent learns your policies and troubleshooting. Neither is distracted by the other’s world.
The routing: the right specialist, automatically
You define rules for which kinds of messages belong to which agent. BotSailor reads the customer’s message, assigns the right specialist, and keeps that specialist in the conversation until the work is done, a transfer is needed, or a human takes over. If the topic changes mid-chat, the agent can transfer the conversation to another agent directly.
The handoff: knowing when to stop
When someone asks for a person, an agent can assign the conversation to a specific team member, or to a role, distributing the work across whoever is available and respecting your business hours. The AI is available at every hour. People are not. The handoff honors both facts.
The discipline of restraint
One of the hardest design questions was not what agents should do. It was what they should never do carelessly.
Capability is easy to add and hard to constrain. A team that can do everything with no judgment is not a team, it is a liability. So a great deal of the design is restraint.
- Each agent receives only the actions it truly needs.
- An agent transfers only when another specialist is genuinely required, not because a customer said “hi,” “thanks,” or “okay.”
- An agent never transfers to itself, and never chains transfers within a single message.
- It never invents a price, a stock level, or a discount.
- Cancelling an order requires a clear request.
- Blocking a subscriber is for genuine spam, never for a customer who is simply frustrated.
- And when the system prompt and the knowledge disagree, the prompt wins, because behavior must outrank information.
Each of these rules sounds small. Together they are the difference between a demo that impresses and a system a business can trust with its customers.
A product’s character is defined less by what it can do than by what it refuses to do carelessly.
Why we kept your flows
There was a tempting path: declare structured automation obsolete and let AI handle everything. We chose not to.
Predictable processes deserve predictable tools. A checkout confirmation, a booking form, a verification sequence: these should run the same way every time. Judgment and conversation, on the other hand, are where AI belongs. So agents work alongside your existing flows and trigger them whenever structure is what the moment calls for.
The best systems are not those that replace what already works, but those that know where each tool belongs.
What it feels like
Return to the customer with the jacket.
They ask about the product, and a Sales Agent answers, drawing on your live catalog. They ask for the larger size, and the same agent confirms without losing the thread. Then they mention their earlier order. The conversation moves, quietly, to a Support Agent that already knows what they were shopping for, asks for the order number, checks the live status, and explains where the parcel is. When the customer finally asks for a person, a teammate receives the chat with the context intact.
The customer experienced one conversation. Your business experienced a team that worked in the right order. Nobody repeated themselves, and nobody waited.
That is the entire ambition: for the complexity to live inside the system, so that the simplicity can live in the experience.
What we learned
Looking back, a few lessons stand out, and they apply well beyond software.
Specialization scales better than size.
Our instinct was to build one more capable generalist. The breakthrough was to stop. Focus beats breadth, in assistants as in organizations.
Memory is what makes a team feel like a team.
Skills are visible. Shared context is not, yet it is the thing customers feel most.
The most valuable feature is knowing when to stop.
An AI that hands over gracefully earns more trust than one that insists on answering.
Simplicity has to be earned.
“Type ## and choose an action” sounds effortless. It took a great deal of work to make something deeply capable feel obvious.
And above all, listen.
Almost nothing we shipped this year began in a meeting room. It began with a message from one of you: a question, a complaint, a half-formed idea at the end of a support ticket. The best roadmap we have ever had is our community.
What comes next
We will not turn a reflection into a roadmap. But we can describe the direction, because it has not changed since that first conversation: fewer steps between a customer’s need and the outcome, more judgment where judgment helps, more structure where it matters, and a person always one request away.
Multi-Agent AI is not the end of that journey. It is the first moment it felt like the right shape.
