Product

The Assistant That Asks Before It Acts

AltSlate LabsSeptember 25, 20266 min read

Most software with "AI" in it adds a chat box to the same old screens. We built the other way round for a growing service business: AI agents carry the routine operational work (bookings, schedules, member records, events, invoices, emails) and the team directs them. The result is an assistant that does real work and, for changes to the business, shows you what it will do and waits for your yes. This is what that looks like from the staff side.

~40

Specialist agents, each with one clear job

about half only read; half can propose changes

1 hour

How long a proposed change waits for a decision

then it expires instead of lingering

2 a day

The cap on proactive alerts per business

when switched on; ranked by relevance

01 — the idea

Agents do the work; people make the calls

Running a membership business is hundreds of small tasks a day: who is booked where, which staff member covers which session, which invoices are outstanding, which member asked what. None of it is hard. All of it takes time.

Instead of one general chatbot, the platform has a team of specialist agents. One knows bookings, another staff rotas, another events, another member records, another communications. Ask a question in plain language and a supervisor breaks it into tasks and hands each to the right specialist, several at once when they do not depend on each other. Ask it to do something, and the specialist prepares the change but does not make it.

A team, not a chatbot. One request, split into tasks and handed to the specialists that own them. Read-only specialists answer straight away; specialists that change records prepare a proposal for a person to approve. (Simplified: the real fleet is around forty agents.)

02 — preview, then approve

Every change arrives as a card you can read

When an agent wants to change something, such as moving a session, inviting a member or publishing an event, it produces a preview: what it will change, from what to what, and why. The staff member can approve it, edit it first, send it back with feedback for a better version, or reject it. A previewed change is not written until someone says yes, and most changes to business records work this way.

An approval card (illustrative, sample data). The agent has done the work of finding the slot, checking the staff rota and drafting the notice; the person decides. Edit changes the details before applying; Revise sends feedback so the agent can try again.

Proposals do not pile up silently: an unanswered one expires after an hour, and a change can only be applied once, even if two people press approve at the same moment. Whatever was approved, rejected or edited stays on record: who asked, what the agent proposed and what was decided.

03 — the right amount of oversight

Not everything needs a sign-off

Asking permission for every lookup would make the assistant useless. So the level of oversight follows the kind of work, and it is decided in the design, not left to the AI.

Oversight follows the work. Questions are answered directly. Changes to shared business records are previewed and approved. People acting on their own records, such as signing up for an event, just do it. Messages drafted into the communications queue wait there until staff send or schedule them.

Messages get special care. Newsletters and reminders are drafted into a communications queue grouped by urgency, where nothing sends until staff choose "send now" or "schedule". Even "approve all" only covers what is due today. Incoming customer emails get a suggested reply in the inbox, ready to approve, edit or discard. The assistant drafting it can look up records but never change them, because anyone can send an email.

04 — one assistant, three audiences

Staff, members and visitors each see their own version

The same agent platform serves everyone, but each person only reaches the agents and records their role allows. Staff use it in the web dashboard and the staff mobile app. Members use a member-scoped version in their own app and portal: their bookings, their events, their history. Anonymous visitors on the public website can ask general questions about the business and leave an enquiry for the team.

Same platform, scoped by role. Each audience reaches a different slice of the agents. Visitors can only ask general questions and leave an enquiry; members act on their own records; staff can propose changes to the business, subject to approval and their role.

05 — it speaks up, a little

Two useful alerts a day beat twenty ignored ones

Agents can also work in the background. When a business switches this on, the platform looks twice a day, in its local morning and evening, for things worth a manager's attention: members who have gone quiet, gaps in the booking schedule, recurring themes in feedback. AI judges which of these matter for this business, and it sends at most two. They arrive as a notification on the phone and in the dashboard. If acting on one means changing something, that change goes through the same approval card as everything else.

A morning briefing (illustrative, sample data). Signals are gathered, judged for relevance to this business, ranked, and capped at two per day, so the alerts stay worth reading.

06 — what it adds up to

More capacity without more back office

The agents take the legwork: finding, checking, drafting, preparing. People keep the decisions: what changes, what gets sent, and when.

For the team, the change is less clicking through screens and more reviewing well-prepared proposals. For the business, it is room to grow without hiring at the same pace, plus a record of the changes agents proposed and people decided. The dashboards and apps still work without the assistant, so nothing is hidden behind it.

The engineering behind it, the agent harness of one orchestration graph, permission gates, typed tools and approvals that can land up to an hour later, is the subject of our agent harness article.