Build it for me

The autopilot builds a first draft of an assistant on its own. It reads the business’s website, your notes and, when it can, the business’s public Instagram profile. Then it writes everything into the assistant, asks the draft realistic customer questions and shows you how it did. Nothing reaches customers: the assistant stays in draft until you set it live yourself. It works only on an assistant that isn’t live. To add to a live assistant, use Quick setup.

​

What it builds

Everything is written in the business’s language (the assistant’s main language). It uses only facts from what it read. Anything customers usually ask that it couldn’t find is listed under Not found in the material, for you to ask the client.

​

Settings shared with the client’s other assistants

Opening hours, services and the booking system belong to the client, so all of its assistants share them (see Clients and assistants). The autopilot fills them only when they’re empty. Anything already set is kept as it was, and the results say so.

• The booking system is set to Opening hours (built in) only when the client has none yet. If the client books through YCLIENTS or Altegio, which keep their own list of services, the autopilot doesn’t add services.
• The time zone is never changed. The results show the time zone booking times use, so you can check it in the booking settings.
• If the client’s plan doesn’t include bookings, no booking flow is made and the results say why.
• An assistant that already has a booking flow keeps it, so it never gets two. The booking flow comes before a template’s request form, which asks for a date as text and doesn’t book.

​

Start it

Admins and editors can start the autopilot in three places. A run takes a few minutes. You can leave the page while it works: the steps show Reading, Writing, Building and Testing, and the results wait on the page. Only one run per assistant can go at a time.

​

Instagram

The autopilot reads a public Instagram profile only through Meta’s official business discovery, never by scraping. That needs an Instagram account already connected in your agency. It uses this assistant’s own connection first, then a connected assistant of the same client, then another connected assistant of your agency. It never uses another agency’s account. It reads the profile’s name, bio, website and the captions of the 12 most recent posts. The profile has to be a business or creator account. When Instagram can’t be read, the run goes on with the website and notes, and the results say why: An instagram.com address typed as the Website is treated as the handle and isn’t opened as a website.

​

Cost and budget

The run uses the assistant’s own AI brain settings. Its AI use counts against the assistant’s monthly AI budget and, for assistants on included AI, against the agency’s allowance. Estimate the cost shows the most the run can cost, for reading the material, writing the draft and the whole self-test. It also shows the budget left this month and, for included AI, the allowance left. The estimate itself reads and spends nothing. The real cost is usually lower. Start is turned off, with the reason, when:

• the assistant has no AI service connected;
• the most the run can cost is more than what’s left of the monthly budget;
• the included AI allowance is used up, or the run could go past it.

If the budget runs out while it works, the run stops early. What was built so far is kept, and the questions it didn’t get to are marked Skipped. Raise the budget and click Test again, or apply or discard what’s there. An assistant on the test AI builds for free.

​

The self-test

After building, the autopilot writes about 15 realistic customer questions in the business’s language and asks each one to the draft, the way a new customer would in the playground. Nothing is sent to anyone. The questions cover prices, opening hours, booking, facts about the business, something that isn’t in the knowledge and a request to talk to a person. A booking question, a question it can’t know the answer to and a request for a person are always there. Each answer is checked against what the draft knows. A failed answer shows what went wrong: Click Show the answer to read the reply it gave.

​

Fixes

A failed question can come with a suggested fix:

• An exact answer: a question and answer added to the draft’s “Prices and FAQs” knowledge.
• Phrases that pass the chat to the team: whole phrases such as “alguien del equipo” added to the rule that hands a chat to the team, for customers who ask for a person in words the rule doesn’t know yet. The rule already knows the usual ways to ask in English, Russian, Ukrainian and Spanish, so this fix is rare.

Edit the fix if you like, then click Add this fix. Check that an exact answer holds only facts the client gave you. When you’ve added the fixes you want, click Test again to ask the same questions once more. A retest uses the assistant’s AI too, and is estimated and checked against the budget first.

​

Kept as answer tests

The questions are saved as answer tests, marked From the autopilot. They stay switched off while the draft is waiting, so nothing spends AI on a draft nobody approved. Apply switches them on, and from then on they run again by themselves whenever the knowledge or the set-up changes.

​

Apply or discard

Check what it built and how the self-test went, then choose.

• Apply keeps everything. The test questions become active answer tests, and a version is saved in History. The assistant stays in draft: try it in the playground and set it live once the client is happy.
• Discard removes everything the autopilot added: its knowledge documents, exact answers, booking flow and test questions. The shared settings it filled are emptied again, unless someone changed them since. The assistant’s other settings go back to the version saved in History just before the run started, so other setting changes made since then are undone too. Knowledge you added yourself stays.

A run that couldn’t finish can be discarded too. When a run failed after building part of the draft, the page offers Discard before you start again. A run that hasn’t moved on for half an hour, for example because the server restarted, is shown as stopped and can be discarded the same way. After a failed or discarded run, the form comes back with the same website, handle and notes, ready to start again.

​

With an AI agent

AI agents can run the whole flow with these tools: The agent shows you the estimate before it starts, asks before it adds a fix, applies or discards, and sets the assistant live only when you ask. For example:

Build the Demo Studio assistant from demo-studio.example.com and @demo.studio. Show me the cost first, then the self-test results.