Air Canada got sued in 2024 because its chatbot hallucinated a bereavement-fare policy that didn't exist. The tribunal made Air Canada honour what the bot said, not what the actual policy was. The precedent is now baked in: if your AI says it to a customer, legally, you said it.
Now imagine that same dynamic for a wedding venue. The AI, acting autonomously, quotes a couple £9,500 for a Saturday in peak season — when the real price is £14,500. The couple books based on the quote. You find out three days later when the contract arrives. Do you honour the AI's price? Do you lose the booking? Do you absorb a £5k hit to revenue? There is no good outcome here.
This is why we've been unambiguous from day one: at MarsAI, the AI drafts. A human approves. Always. The small label that reads "Human approves every outbound. AI never fires blind" on our homepage isn't marketing copy — it's the design constraint the entire product is built around.
The myth of "fully autonomous"
The tech press loves the word "autonomous." It sounds futuristic. It suggests magic. Every investor deck in 2025 was variants of "our AI handles the entire customer flow, no humans needed."
For some workflows — customer support triage, internal scheduling, data entry — autonomy is fine. For anything that sends outbound communication with revenue, brand, or contract implications, autonomy is reckless. There are four specific reasons.
1. Hallucinations aren't rare — they're occasional
Even the best models hallucinate roughly 1-3% of the time on factual questions. For 100 enquiries a week, that's 1-3 materially wrong statements going out the door, unchecked. Over a year, that's 50-150 potentially embarrassing, costly, or legally problematic replies.
2. Tone is a weapon
Venue replies aren't customer service tickets. They're sales artifacts. A slightly-off tone — too formal for a laidback DM, too casual for a corporate enquiry, too pushy in a bereavement follow-up — kills conversion. Humans catch tone mismatches in under a second. AI doesn't.
3. Context beyond the enquiry
A good venue manager replies to an enquiry knowing that this couple's friends booked last year, that this planner sends regular business, that your calendar has a quiet week worth discounting, that there's a local event that'll affect parking. The AI doesn't know any of this. The human operator does.
4. Liability
UK contract law doesn't care that an AI sent the email. Your venue is on the hook for anything quoted, promised, or agreed to in a customer-facing message. One click of human approval is also one click of legal diligence. That click is not optional.
"But won't human approval slow us down?"
This is the real question, and the reason so many vendors push autonomous AI. They assume human-in-the-loop = slow. It's not — if you design the approval flow properly.
The time math works out like this. Typing a reply from scratch: 8-15 minutes per enquiry. AI drafts + human approves: 20-40 seconds. That's 10-20x faster than fully manual, while being safer than fully autonomous.
The question isn't "should AI send replies?" — it's "who should do the first draft?" The answer is always AI. The question of "who clicks send" should always be a human.
Done well, the approval UI should let an operator review + approve a draft in under 15 seconds. Here's the breakdown we've optimised toward:
- Glance (3-5 seconds): Does the draft reply to the right enquiry? Is the tone right? Does it reference the correct details (date, pax, channel)?
- Skim (5-8 seconds): Are the price points correct? Does it include the attachment we'd normally send? Are there any phrases that feel "off"?
- Action (2 seconds): Approve + send. Or tap to edit inline if a small change is needed.
For venues we've deployed this with, the average approve-to-send time is 22 seconds. Half of approvals are instant — operator reads the draft, sees it's good, taps. The other half involve a small edit (usually adding one sentence of personal context the AI couldn't know).
What to automate without approval
There's nuance here. Not every action needs human approval. The line we draw is roughly: any outbound communication to a customer requires human approval. Any internal action doesn't.
So:
- Auto-approved (internal only): moving enquiries between pipeline stages, calculating SLA breach risk, updating lead scores, surfacing priority enquiries, scheduling internal reminders, drafting replies.
- Human-approved (customer-facing): sending any email or DM, sending pricing, confirming availability, making commitments on price/date/inclusions.
- Never automated: contract issuance, deposit requests, cancellations, refund communications. These are human all the way.
This mental model — automate the back office, approve the front office — scales well. It also mirrors how your team already thinks about delegation. The AI is a very fast junior sales coordinator who drafts replies for a senior operator to sign off. It's not a replacement for the senior. It's what makes the senior capable of running ten times more volume.
How to talk to customers about it
A small but meaningful point. If you use AI to draft replies, should you disclose it? Our answer: you don't need to disclose that AI drafted, because a human approved. The reply is, legally and ethically, yours. If you edited it or approved it, it's your communication.
What you should never do is misrepresent. If a customer asks "am I talking to a bot?", the honest answer (if they're in a messaging thread) is: "You're talking to [operator name]; I use AI tools to help me draft responses faster." That's transparent, true, and in our experience, customers respond well. The "AI is faster" framing is actually a plus signal — it suggests the venue is organised.
The principle that makes it work
If we had to distil MarsAI's design philosophy into one sentence, it would be this:
AI earns speed. Humans earn trust. The system is designed so both get to do the thing they're best at.
Your team doesn't want to type the same wedding enquiry reply for the 400th time this year. AI is great at that. Your team does want to be the human a couple talks to at the viewing, at the contract signing, at the cake tasting. AI is not good at that, and shouldn't try.
Get the division of labour right, and everything else follows. Response times drop, conversion climbs, staff burnout drops, revenue lifts. Get the division wrong — either by keeping the typing work or by handing the trust work to a bot — and you get the worst of both worlds.
This is the bar we hold ourselves to. We think it should be the bar you hold your AI vendors to as well.
Request a 30-minute walkthrough of the MarsAI console. You'll see exactly what the approval flow looks like — and what 22-second review times feel like in real life.
Request an ops audit →