Deflect routine intake and scheduling without unsupervised AI

Who this is for

This playbook is for owner-operated companies where the same questions hit the phone, the website chat, and the shared inbox every week: hours, address, how to book, how to send files, “are you taking new work,” “can I move my appointment.” A person still answers most of them by hunting. Leadership has seen a receptionist demo and wants the door covered without handing the relationship to a model.

If the pain is a prospect who already had a good conversation and then went quiet, read pipeline follow-up first. If the pain is the chase after a signed yes, read client onboarding. First-line work is the door, not the chase and not the kickoff.

Symptoms

If several of these are true in the same month, this is a live operational leak — not a tooling preference.

• The same five questions — hours, booking, files, “are you taking work,” how to pay a deposit — arrive by phone, chat, and email every week.
• Voicemails and after-hours chats sit until someone “gets through the inbox,” so routine scheduling waits on a person who is already in delivery.
• A chatbot or voice agent was turned on and has already promised a time, a price, or a start date nobody approved.
• The team cannot list the intents they actually handle; the vendor list of “use cases” is doing that job.
• There is no review queue — drafts and transcripts vanish into the same inbox that was already overloaded.
• Sales follow-up and post-sale onboarding are being stuffed into the same agent as “also handle the website chat.”
• Someone proposed letting the agent close, discount, or mark a lead qualified without a human.

What done looks like

Done is a working operating change, not a purchased seat or a dashboard nobody opens.

• A top-20 intent list exists, written in the language callers and form-fillers already use.
• Each intent is tagged auto-deflect, draft-for-review, or human-only, with a named do-not-automate list.
• The FAQ and forms those intents need are published and owned. The agent is not inventing answers.
• A review queue exists: a person sees drafts, exceptions, and anything that smells like a close.
• No agent closes, prices, or commits scope. Humans send those.
• Pipeline chase and post-sale onboarding are separate paths, not extra skills on the same receptionist.

Tier-1 is the door, not the closer

Internal name, used once so the catalog stays honest: Tier-1 AI Agents. In this house that means routine intake, scheduling, and support deflection with a human on review. It does not mean an unsupervised closer. It does not mean a cheaper replacement for the coordinator, the setter, or the owner who still takes the hard calls.

The door job is narrow on purpose. Collect the inquiry. Answer what is already published. Offer times that are actually open. File a complete form. Route the rest. Disclose that the voice or chat is not a person when that is true. Stop talking long enough for the human on the other end to answer. Hand a real conversation to a named person. That is coverage. Coverage is allowed. Selling in the dark is not.

If the company does not yet have a shared operating picture for “what slots are actually open” or “are we taking this kind of work,” do not put an agent on the door. A fluent booking against a stale calendar is worse than a missed voicemail. The same is true for answers: if the operational knowledge base does not have an owned FAQ page, the agent will invent one.

Not pipeline chase, not onboarding

Pipeline follow-up is speed-to-lead plus a visible next step after someone is already a prospect. Stages, a next-step field, stale alerts, draft review. Humans send. That work lives in the CRM. A first-line agent that starts “nurturing” a caller it just met is unsupervised selling wearing a receptionist badge.

Client onboarding starts when the commercial yes exists. Intake packets, countersign, deposit, assets, kickoff. That is a sequence with blockers. Stuffing “also collect the W-9 and the brand kit” into the public chat agent mixes a stranger with a signed client and usually drops both.

Keep the three rooms. The door collects and routes. Follow-up owns the live opportunity. Onboarding owns the signed path. If a vendor wants one agent to do all three, that is a platform story. Rank it as three workflows in the operational ranking, then refuse the bundle until each room has its own rules.

Intent taxonomy before any agent product

Write the intents you already handle. Not the vendor’s library. Interview the person who answers the phone and the person who lives in the shared inbox. Twenty is enough. If you cannot name twenty, you are not short on AI. You are short on attention to the door.

Then sort. Three buckets only.

Auto-deflect

Published, low-blast-radius answers the business already stands behind: hours, address, how to send files, how to request a callback, a link to the booking page for a slot type that is actually open. Auto-deflect still needs a log. It does not need a closer. If the answer is not on an owned page, it is not auto-deflect yet.

Draft-for-review

Messages that sound routine but commit the company if they are wrong: “we can take that scope,” “next week is open,” “send over the files and we’ll start.” The agent may draft. A person sends or books. This is the default for anything that touches calendar, fit, or money-adjacent language.

Human-only

Complaints that name harm, anything legal, anything that sounds like a buying negotiation, VIP or existing-client exceptions, credentials, and any request to change a system of record. The agent’s job is to collect enough to route and stop. It does not reassure, diagnose, or promise a fix.

Write the do-not-automate list in the same pass. Closing. Price. Discount. Start dates. Scope. Collecting passwords or account credentials. Marking a CRM stage won or qualified. Sending a proposal. Changing a booking that has a deposit attached. If a demo cannot honor that list, the demo is the wrong product.

DIY before any voice or chat agent product

Most of the value is available before you buy a face or a voice. Publish the FAQ the top twenty already need. Put the booking links on pages humans can find. Make the form ask for the fields the team actually uses — name, company, callback, what they need, urgency — and stop asking a second time in email. Name the review queue: who looks at drafts, how often, and what “escalate now” means.

Run that path for a normal week without an agent. If the FAQ is wrong, you will hear it from people, not from a model repeating the error after hours. If the form is missing a field, the coordinator will still chase. If nobody owns the queue, a future agent will dump transcripts into the same swamp. The agent does not create the operating path. It sits on one.

Adoption belongs here too. A seat that can talk is still a seat. When AI seats don't equal AI usage is the 90-day test: who reviews, what “handled” means, and whether the door actually moved. Demo theater at the receptionist layer is the same waste as demo theater in a copilot — a vendor hour, then the shared inbox again.

No autonomous closing

Hard rule. The agent does not close. It does not accept a yes on scope or price. It does not send a contract. It does not take a card to “get you started” unless a human-owned booking path already does that for a fixed, published offer — and even then, anything custom stops. A first-line system that can close will close the wrong thing. The cost is not a missed optimization. The cost is a commitment the owner never made.

Scheduling is allowed only against a trusted calendar picture. If two calendars disagree, refuse and hand to a person. If the requested work is not on the published service list, refuse and hand to a person. “I can get you in Tuesday” is a commitment. Treat it like one.

The same rule applies to tone. Do not let the agent flatter, diagnose, or rank the caller’s business. Collect, answer what is published, route. That is enough door work for a long time. Expanding the agent is a later ranked project, not a slider in a vendor console.

What the week looks like when this is working

Monday, the owner can see the top twenty intents and last week’s exceptions in one queue. Routine hours and booking links were answered from the published pages. A handful of drafts waited for a person: a fit question, a reschedule on a paid slot, a “can you also do X.” Nothing in that list is a surprise close.

Midweek, a caller asks for a price that is not on the card. The agent collects the callback and stops. A chat asks to move a kickoff that belongs to a signed client. That is onboarding, not the door — it routes to the coordinator, not into a new sales sequence. A form arrives complete enough that follow-up can start with a next-step field instead of “who is this?”

When the team is out, the door still discloses what it is, answers the published set, and refuses the rest. Nobody comes back to a weekend of invented commitments. That is first-line coverage. It is still human at the edge that matters.

What people usually get wrong

They buy the voice before the intents. A warm voice reading an empty FAQ is a new way to be wrong after hours.

They let one agent be receptionist, closer, and onboarder. Those are three blast radii. Mixing them is how a stranger gets a start date and a signed client gets a sales sequence.

They skip the review queue because “it’s only FAQ.” FAQ drifts. A published page with an owner can be corrected. A transcript nobody reads cannot.

They treat deflection as a percentage target. This playbook will not give you one. The job is a clean handoff and a refused close, not a score that trains the agent to keep talking.

They hide that the voice or chat is automated. Callers notice. Disclosure is part of the door job, the same way a human receptionist would not pretend to be the owner.

DIY vs hire

DIY the top-20 intents, the three-bucket taxonomy, the do-not-automate list, the FAQ and forms, and a named review queue. Most teams can do this with the inbox, booking link, and docs they already pay for. The discipline is marking human-only and refusing a closer.

Hire for voice or chat coverage across several channels, calendar write-back, or a receptionist that must sit on a real phone number with logging. Those are later builds. The $997 assessment is a consult and a written analysis that can put first-line coverage in the DIY now list or on the roadmap. Implementation is quoted only if you want it.

Controls before automation

Recommendations that touch customers, money, contracts, private data, or systems of record need human review, limited permissions, a log, and a fallback. Do not automate a broken path because the tool is ready.

• No autonomous closing. No unsupervised price, scope, discount, or start-date commitments.
• Humans send anything in draft-for-review. Auto-deflect is limited to owned, published answers.
• Do-not-automate includes credentials, legal, harm, VIP exceptions, and CRM stage changes.
• Scheduling writes only against a trusted calendar picture. Conflict or stale slots refuse.
• A review queue is required before a voice or chat product goes live. Transcripts are not “the inbox.”
• The agent inherits the access a receptionist already had — not a wider key to the CRM or billing system.

How this sits next to the other playbooks

First-line coverage is a revenue-motion room in the fourteen mid-market AI asks. It sits at the door. Follow-up sits on the live opportunity. Onboarding sits after the yes. Knowledge and a shared picture decide whether the door is allowed to speak. Ranking decides whether the leak is the door or unused seats pretending to be coverage. Adoption decides whether anyone actually works the queue.

If you want that order written down with you, book the assessment. Bring the last twenty inbound questions. Leave with a taxonomy, a do-not-automate list, and a queue the team can run without unsupervised AI.