For the past year the busiest member of my team hasn't slept once. An AI agent runs a large share of IMPT's front line — answering prospects on WhatsApp and email in their own language, quoting live prices in their own currency, booking meetings across time zones, chasing unfinished payments, and escalating to humans when judgment is needed. Not a demo. Production, revenue-carrying work, every day.
Most of what is written about AI agents is written by people who haven't operated one. Here is what running one in the wild actually teaches you.
The work is mostly unglamorous
The public imagination is agents doing clever reasoning. The economic reality is agents doing coverage: the 3am price question from Jakarta, the eleventh gentle reminder, the currency conversion a human would have got wrong at that hour. Travel is a 24-timezone business; the single biggest thing an agent changes is that nobody anywhere waits until Ireland wakes up.
Five operating rules we learned the hard way
1. Never let the agent fake an action. Early on, any gap between "the agent said it sent the email" and "the email sent" destroyed trust instantly. Every action must be verified against the provider's receipt before the agent may claim it happened.
2. Facts come from tools, not memory. Prices, availability, payment status — the agent must look them up at answer time, every time. A model's memory is a rumour; your database is the truth.
3. Speak the customer's money. We watched buyers abandon payment pages simply because the amount was shown in a foreign currency. A price a customer has to convert is a price they don't trust. The agent now quotes in local currency everywhere, converted at the day's rate.
4. Escalation is a feature, not a failure. The agent's most valuable skill is knowing what it must not decide: refunds, disputes, anything legal, anything it isn't certain of. "Let me confirm and come back to you" outperforms a confident wrong answer every single time.
5. Guard the sends. An autonomous system that can message customers can also over-message them. Hard caps, quiet hours, and opt-out discipline aren't compliance chores — they're what keeps an agent welcome in someone's WhatsApp.
What this means for hotels
Hotels have run on thin front-desk staffing for decades, and the booking platforms above them run on ticket queues. Agent-grade AI collapses that: instant, multilingual, around-the-clock service becomes table stakes, first at the platform layer, then at the property layer. The operators who treat the agent as a colleague to be trained — given clear rules, verified tools, and honest escalation paths — will simply out-serve the ones who treat it as a chatbot to be bolted on.
We built IMPT — 8 million hotels and apartments, bookings that fund verified carbon offsetting — with a team a fraction of the size the industry would consider normal. AI agents are the reason that is possible, and the reason it stays possible as we scale. The $30M behind the platform is accelerating many things; none of them requires the head-count curve the last generation of travel companies needed.
Common questions
What does IMPT's AI agent actually do?
Front-line sales and support: answering prospects in their own language, quoting live prices in local currency, booking meetings, following up unfinished payments, and escalating to humans where judgment is required.
What is the most important rule for a production AI agent?
Never let it claim an action it hasn't verifiably performed — every send, booking or payment claim must be checked against the provider's receipt, and every fact must come from a live tool lookup, not model memory.
Do AI agents replace hotel or platform staff?
They replace coverage work — the 24/7, multilingual, repetitive layer — and raise the bar for service. Humans keep the judgment work: disputes, exceptions, relationships.