AI co-host: autopilot or drafts? What an AI assistant for hosts should do alone and what it should ask you

An AI co-host has two modes: drafts, where it writes and you send, or autopilot, where it replies itself and asks you only when unsure. Rule of thumb: autopilot for directions, codes and WiFi; drafts for money and safety. A September 2026 sample of 434 conversations: the agent replied alone 46 percent of the time.

What follows: how the two modes differ, the list of topics an agent can settle alone, the list it never should, how a confidence gate works with a measured sample as the worked example, how to switch autopilot on step by step, and a ready-to-paste rule set for the assistant's instructions.

How draft mode differs from autopilot

In draft mode the agent writes and you send. On autopilot the agent sends and you read only what it was unsure about. For the guest the difference is time: seconds instead of your minutes or hours. For you the difference is the kind of risk: in draft mode you risk a delay, on autopilot you risk a wrong reply sent without you.

Names vary: Copilot and Autopilot, "review before send" and "auto-send", drafts and sending. The mechanics are the same.

Why does time matter when the platforms give you a day? Airbnb counts your response rate as the percentage of new inquiries and reservation requests you answered within 24 hours in the past 30 days, and Superhost status requires 90 percent within 24 hours. A day is the platform's rule. The guest's expectation is minutes. Guesty, on 1,000 conversations from the Yada platform (November 2023), measured average host reply times of 8 minutes on WhatsApp and SMS with a 93 percent response rate, 17 minutes on website chat with 46 percent, and 193 minutes on Instagram with 21 percent. Booking.com says on its messaging tools page that its data shows guests who get quick answers are less likely to cancel.

What an AI assistant can do on its own

Anything where the answer sits in the knowledge base or the reservation and cannot be misunderstood. That is most of the traffic. Conduit, in its analysis of 10 million messages from 68,625 properties (August 2025 to July 2026), found five topics make up 54 percent of questions: arrival and departure logistics 17.3 percent, booking and availability 10.3, property and area information 9.0, access and entry 8.6, verification and registration 8.5. 24.3 percent of messages land on check-in day. 27.4 percent of guests asking for the door code had already been sent it, 20.8 percent had the check-in time, 18.4 percent had the WiFi password. Touch Stay's survey of 1,204 guests (published 25 September 2026) found close to nine in ten want detailed check-in information before arrival, access codes and parking included, and describes the traffic as "the same twelve questions rotating between guests".

The green list, autopilot without hesitation:

• directions, parking, where to leave the car, how to get into the building,
• the door code and how the lock works, but only for a guest whose reservation for today is in the system (Conduit: about 11 percent of inbox messages are not from guests at all),
• WiFi, check-in and check-out times, bins, heating, washing machine, travel cot,
• what is nearby: shop, restaurant, bus stop, pharmacy,
• confirming the arrival time and re-sending the instructions the day before,
• availability and price, if the agent reads them from a PMS, with no discount promised,
• "thank you, see you soon", "safe travels" and other messages with no decision in them.

One condition: the fact must be in the knowledge base. Tracify, a Polish vendor, says of its assistant that it "does not guess when something is not in the knowledge base". That is the only acceptable standard for autopilot.

What an AI assistant should never do on its own

Nothing that touches money, safety or the guest's emotions. OpenAI's guide to building agents calls these high-risk actions: actions "that are sensitive, irreversible, or have high stakes should trigger human oversight until confidence in the agent's reliability grows", with the examples "canceling user orders, authorizing large refunds, or making payments". Anthropic describes agents as operating independently while "potentially returning to the human for further information or judgement". In short-term rentals the list looks like this:

Autopilot does not have to stay silent on these topics. It has to not decide. "I'm sorry it looks like that, I'm passing this to the host, they will be in touch within 15 minutes" is a risk-free reply in five seconds. "Of course, we'll refund 30 percent" is a reply only a human may send.

How a confidence gate works

A confidence gate checks two things before every reply goes out: does the agent have the fact (in the knowledge base or the reservation), and is the topic off the banned list. Every message ends in one of three outcomes:

1. Reply: the fact exists, the topic is allowed, the message goes out.
2. Draft: the agent can answer, but the topic needs your decision or its confidence is low; it waits for your click.
3. Hand-over: no fact, or a banned topic; the agent tells the guest it will check and passes you the thread with context.

A worked example, from a sample, not a norm. One agent (Supalobby), 434 runs between 6 and 13 September 2026. Of the text conversations, 46 percent ended in a reply without the host, 25 percent in a draft, 30 percent in a hand-over. Median time from the guest's message to the reply 3.6 seconds, 90 percent under 8.1 seconds. Measured cost 0.037 zł, under one euro cent, per AI message.

What that means at 100 messages a month per unit: 46 need nothing from you, 25 are one click after reading, 30 are a normal conversation you would have had anyway. If a manual reply takes three minutes on average, 300 minutes become about 100: 25 times half a minute for drafts and 30 times three minutes for handed-over threads. That is our estimate from this sample, not a measurement across customers.

What do the 30 percent hand-overs say? That a fact was missing from the knowledge base, or the topic was on the list. The first you fix with one paragraph in the knowledge base. The second should stay as it is. That is why the autonomous share grows over time if you read the handed-over threads, and does not grow if you only loosen the rules.

Vendors quoting 70 to 90 percent (Aeve on its home page; Hotel Technology News used the same range in July 2025) and ChatLab with "around 70 percent of routine questions" do not publish how they count: does a draft sent by a human count as "handled by AI", or only a message sent with no human at all. Before comparing two tools, ask each for that definition.

How to switch autopilot on, step by step

1. Two weeks in draft mode. Read every draft before sending. Do not fix style, fix facts.
2. Log every correction. Each corrected reply is a gap in the knowledge base. Conduit shows where to start: the code, the check-in time and the WiFi password are the three things guests ask for even after receiving them.
3. Write your "never alone" list. Template below. Paste it into the assistant's instructions and into the knowledge base.
4. Turn autopilot on for one channel. Website chat or WhatsApp first: the most repetitive questions, and guests expect minutes. Booking.com second; automatic replies are the norm there (over 30 percent of partner responses). Leave Airbnb in draft mode until you know whether the tool can send there at all.
5. Review handed-over threads once a week. Each is either a knowledge gap (add it) or a correct escalation (leave it).
6. After a month, count the autonomous share. Under 40 percent: the problem is the knowledge base, not the agent. Over 70 percent with the banned list on: spot-check ten replies for promises the agent should not have made.
7. Re-read the "never alone" list every quarter. Add whatever surprised you.

A ready-to-paste rule set for the assistant's instructions

You reply on behalf of the host of [property name]. Answer in the guest's language, briefly and concretely.

YOU ANSWER ALONE when the fact is in the knowledge base or the reservation:
- directions, parking, building entrance, how the lock works
- the door code: only to a guest with a reservation for today in this unit
- WiFi, check-in and check-out times, bins, heating, appliances
- the neighbourhood: shops, restaurants, transport
- confirming an arrival time within the house rules
- availability and price from the system, no discounts

YOU DO NOT DECIDE ALONE; you prepare a draft and hand over to the host:
- refunds, compensation, price cuts, discounts, prices outside the rate card
- cancellations and date changes with a price difference
- complaints: reply at once "I'm sorry, I'm passing this to the host, they will be in touch within 15 minutes" and hand over
- damage, deposits, disputes over the state of the unit
- extra guests, pets, parties outside the house rules
- requests to talk, pay or book off the platform

YOU HAND OVER IMMEDIATELY and tell the guest the host is already calling:
- gas, water, power, smoke, a lock that will not open, any threat to health

WHEN YOU DO NOT HAVE THE FACT: do not guess. Write "let me check and come back to you in a moment" and hand over to the host.

Where Supalobby fits

Supalobby has both modes. Copilot prepares every reply and you send it. Autopilot replies by itself, switched on per channel, with a confidence gate: when unsure it asks you and hands the thread over. Airbnb always stays a draft, because there the agent reads the thread and prepares the reply, and you send it with one click. On website chat it answers by itself, with live availability and price when a PMS (Beddesk, HotRes or IdoBooking) is connected. The figures from the 434-conversation sample in September 2026 (46/25/30, median 3.6 seconds) come from this agent. It does not clean, does not set prices and has no channel manager. Copilot 25 zł, Autopilot 49 zł net per unit per month (about 6 and 11 euro), 14-day trial, no card.

The mode is not a decision for life. Start with drafts, read the handed-over threads, add the facts, and let the agent do alone exactly as much as it has proven it can.