Why We Built a Team of AI Avatars Instead of One Chatbot

Short answer: One chatbot that tries to serve everyone ends up serving no one well. A team of AI avatars gives each audience and channel its own specialist, with a clear role, a consistent voice, a narrow scope and a defined handoff to a human. At Home Alliance we call it One Molly, four faces: Molly for business and strategy, Milli for the office, Max for contractors and field work, and Lumi for homeowners and Passport members. Max and Milli are in pilot, and Lumi is planned. None of them decides anything a person should decide, and sensitive actions require human approval.

Yesterday I wrote about the AI C-suite: thirteen AI executive seats, each owning one function inside the company. The rule that held that structure together was simple: one seat, one function. Today's article is about what happens when you apply the same rule on the outside of the company, where customers, technicians, partners and the public meet us.

The common move in our industry is to put one chatbot on the website, connect it to a phone line, and call it "AI." I understand the appeal. It is one vendor, one integration and one line in the budget. But after watching how people actually talk to a home-services company, I am convinced that the single-chatbot approach is the wrong model. Here is why, and what we are doing instead.

The problem with one chatbot for everyone

A home-services company does not have one audience. It has several, and they want very different things.

• A homeowner with a broken air conditioner wants to know when someone can come, what it might cost and whether the company can be trusted inside their house.

• A technician standing in front of an unfamiliar unit wants model history, likely causes and the next diagnostic step, fast, with dirty hands.

• A partner or independent owner wants to understand how a shared platform works, what it costs them and what they keep.

• A reader, candidate or journalist wants to understand who we are and what we believe.

Put one chatbot in front of all four and you get a familiar set of failures.

The voice is wrong for most of them. A tone that reassures a worried homeowner sounds patronizing to a technician. A tone that works for technical troubleshooting sounds cold to someone whose basement is flooding.

The scope becomes infinite. A general bot is asked about pricing, diagnostics, contracts, hiring and company history, often in the same week. Every new topic adds instructions, exceptions and data sources. Eventually the instructions contradict each other, and nobody can say with confidence what the bot will answer.

The risk profile is mixed. Some questions are low-stakes ("What are your hours?"). Some are not ("Is it safe to keep running this unit?"). When one system handles both, you have to set its guardrails for the riskiest case, which makes it useless for the easy ones, or for the easiest case, which makes it dangerous for the hard ones.

Accountability blurs. When a single bot gives a bad answer, who owns the fix? Marketing? Operations? IT? In practice, often no one.

This is the same failure I described for AI executives: the problem with capable AI is rarely laziness. It is overreach. A general assistant will try to answer everything, and confident wrong answers are worse than honest "I don't know, let me connect you."

What an AI avatar is, and what it is not

By "avatar" I mean more than a face on a screen. An AI avatar, in the way we use the word, is a named AI role with a defined audience, a defined channel, a consistent persona and a written scope.

It has:

• A name and a recognizable identity, so people know who they are talking to and can come back to the same "someone."

• One primary audience. It is built for that audience's questions, vocabulary and level of urgency.

• A home channel, or a small set of channels, where it lives.

• A scope and a "never" list: what it handles, what it routes elsewhere, what it may never do.

• A human owner inside the company who answers for its behavior and approves changes to it.

• A handoff path to a person, always visible, always working.

It is not:

• A person pretending to be a person. Our avatars are identified as AI. Trust that starts with a disguise does not last.

• A decision-maker. An avatar can explain, guide, collect information and prepare. Pricing exceptions, safety calls, refunds and anything legal go to accountable humans.

• A replacement for our people. Avatars take the repetitive, the after-hours and the "first five minutes" load, so that our people spend their time where judgment and empathy matter.

Meet the team: Molly, Max, Milli and Lumi

We are rolling out four avatars across Home Alliance channels. Each one has a single job.

We describe this on our site as One Molly, four faces. Underneath there is one Molly: one set of approved knowledge, one set of rules, one company behind it. On the surface there are four faces, each built for a different group of people.

Molly handles business and strategy. She is the face that works closest to the leadership of the company: preparing analysis, keeping context and connecting the work of the other faces to the company's priorities.

Milli is for the office. She answers the questions the office team gets every day and helps coordinate the work that keeps jobs moving: who is going where, what is waiting, what needs a follow-up. Milli is in pilot and subject to validation.

Max is for contractors and field work: technicians, contractors, academy students and partners. Max is the technicians' copilot. His job is to help someone in the field get ready for a job and work through it: what was reported, what the equipment history shows, what the likely causes are and what to check next. Max does not replace a technician's judgment. He shortens the distance between a question and a useful answer. Max is also in pilot and subject to validation.

Lumi is for homeowners and Passport members. She will be the face a homeowner meets: clear, calm and focused on what they need to know about their home and their service. Lumi is planned, not live.

I want to be precise about status because it matters. Two faces are in pilot, which means we are testing them with real work and checking whether they actually help before we expand them. One is planned. Nothing here is a finished product, and in every face, sensitive actions require human approval.

What matters more than the roster is the design rule behind it. We did not start with four characters and look for work for them. We started with audiences and channels, asked where the questions were most repetitive and the wait was most painful, and then gave each of those places a dedicated specialist.

Why a team works better than a single bot

1. Scope stays small enough to be trustworthy

An avatar with one audience has a short instruction set and a limited set of data sources. That makes it testable. We can write down the questions it should handle, the ones it should route, and check its answers against both lists. A single bot covering every audience can never be tested this cleanly, because its surface area keeps growing.

2. The voice fits the person

People respond to tone before they respond to content. A technician wants short, precise, technical answers. A homeowner wants clarity and reassurance. A partner wants plain business language. Separate avatars let each voice be right instead of averaged.

3. Risk can be set per role

Each avatar gets guardrails that fit its job. An avatar that explains company services can be relatively open. An avatar that supports diagnostic work must be strict about safety, about what it does not know and about escalating to a senior person. You cannot set those dials independently if everything runs through one bot.

4. Every avatar has an owner

Because each avatar maps to one audience, it maps to one team and one accountable person. When an answer is wrong, the fix has a home. This is the same logic as the AI C-suite: every AI role has a human who answers for it.

5. Continuity builds relationships

A named avatar that people meet again and again becomes familiar. A homeowner who had a good experience remembers the name. A technician learns how Max thinks and how to ask her for what he needs. Familiarity is the beginning of trust, and trust is the product in home services.

6. The system scales by adding roles, not complexity

When a new audience or channel appears, we can add a role instead of stretching an existing one. Adding a specialist is a contained project. Expanding a general bot is a risk to everything it already does.

How the avatars connect behind the scenes

A team of avatars only works if they do not become four separate islands. Behind the faces there is one company, one set of facts and one set of rules.

One source of truth. Avatars draw from the same approved knowledge: services, policies, schedules, technical references. If a policy changes, it changes once, and every avatar inherits it. Two avatars giving two different answers to the same factual question is a bug, not a personality trait.

Clean handoffs between avatars. If a homeowner asks something that belongs to another role, the avatar does not improvise. It routes, the same way our AI executive seats route out-of-scope work to the seat that owns it.

Handoffs to humans are a feature, not a failure. Every avatar knows when to stop and bring in a person: when the question involves safety, money, a complaint, a contract, or simply when the customer asks for a human. We measure an avatar partly by how well it knows its limits.

Shared memory with permissions. Context from one conversation can help the next one, but each avatar sees only what its role needs. A homeowner-facing avatar does not need internal operational data; a field copilot like Max does not need marketing plans.

Coordination by the AI executive layer. The avatars are the outside face; the AI C-suite seats are the inside brain. Customer-experience, operations, technology and media seats each have a stake in how avatars behave, and changes go through the same proposal, challenge and founder-decision path we use for everything else.

The rules every avatar follows

These rules matter more than any avatar's personality.

1. Say you are AI. Always.

2. Never invent. If the avatar does not know a price, a time or a technical fact, it says so and finds out or hands off.

3. No promises about money or outcomes that a human has not approved.

4. Safety first. Anything that could hurt a person or a home goes to a human immediately.

5. Sensitive actions require human approval. Money, commitments, complaints and anything irreversible wait for a person.

6. A person is always one step away. The customer can ask for a human at any moment.

7. Stay in scope. Out-of-scope questions are routed, not improvised.

8. Every change is reviewed. New instructions or knowledge go through the owner and are tested before they go live.

9. Privacy by design. Each avatar collects only what its job needs.

What we expect to learn

We are at the beginning of this rollout, and I would rather describe expectations honestly than claim results we do not yet have.

We expect the avatars to shorten waiting time for simple questions, especially outside business hours.

We expect technicians to arrive better prepared, because Max helps them review the job before they knock on the door.

We expect our human team to spend more time on the conversations that need a human: complaints, complex jobs, sensitive situations.

We also expect problems. Avatars will misread internal terms. They will meet questions nobody anticipated. Some customers will not want to talk to AI at all, and they should not have to. Our plan is to treat every miss as a correction to the shared knowledge, written once, so every avatar inherits the fix.

I will report back on what we learn, with real observations, not marketing numbers.

How to design AI avatars for your own service business

If you run a service company and want to try this, here is the order I would follow.

1. Map your audiences. List every group that talks to your company: customers, technicians, partners, candidates, vendors.

2. Map the channels. Where does each group reach you? Phone, website, text, field app, social media.

3. Find the pain. Where are questions most repetitive, and where is waiting most costly? Start there.

4. Define one avatar for one audience. Write its role, its voice, its scope, its "never" list and its handoff rules before choosing any technology.

5. Assign a human owner. No owner, no avatar.

6. Connect it to one source of truth. Do not let avatars keep their own private versions of your policies.

7. Test against real questions. Use real (anonymized) conversations, including the hard and strange ones.

8. Launch narrow, then widen. One channel first. Add channels and roles only when the first one is stable.

9. Review weekly. Read conversations, correct mistakes, update the knowledge, repeat.

The point is not to have AI everywhere. The point is to have the right specialist in the right place, with a person right behind it.

FAQ

Is it better to use one AI chatbot or several specialized AI assistants? For most service businesses with several audiences, several specialized assistants work better. Each one has a smaller scope, a fitting voice, its own guardrails and a clear human owner, which makes it easier to test and trust than one bot that answers everything.

What is an AI avatar in a business context? An AI avatar is a named AI role with a defined audience, a home channel, a consistent persona and a written scope. It explains, guides and prepares, identifies itself as AI, and hands off to a human whenever money, safety or judgment is involved.

Who are Molly, Max, Milli and Lumi? They are the four faces of one AI at Home Alliance: Molly for business and strategy, Milli for the office, Max for contractors and field work (technicians, contractors, academy students, partners), and Lumi for homeowners and Passport members. Max and Milli are in pilot; Lumi is planned. Details: molly.homealliance.com.

Do AI avatars replace customer service staff or technicians? No. Avatars take repetitive questions, after-hours requests and preparation work. People handle complaints, complex jobs, safety issues and decisions, and a customer can always ask for a human.

How do you keep several AI avatars consistent? All avatars draw from one approved source of truth, follow the same standing rules, route out-of-scope questions instead of improvising, and go through the same review process for any change.

Want more essays like this? Subscribe to The Operator's Edge.