By Sagar Shankaran, Founder of CallSphere
Set AI ceilings around wave season instead of the calendar month, give the air desk the strongest model, and judge the whole line item on cost per booking.
Key takeaways
Nine days. That is how long it took, the week the storm turned toward the Gulf, for your team to rewrite 312 itineraries, re-price 84 air segments and send about 1,900 messages to travellers who all wanted the same three answers. Those nine days used more AI than January, February and March put together — and you found out when the invoice arrived in October.
That was the shape of the problem through 2025 and into 2026: AI spend behaved like a phone bill from 1997. You could not see it until it had happened, could not attribute it to a person or department, and could not stop it mid-month without turning the whole thing off. On 2 July 2026 that changed, and the change is boring, administrative and exactly right.
Nobody in this trade spends evenly across twelve months, and neither does the software. Your usage has three humps, and they are the same humps as your revenue.
The first is wave season, 1 January to 31 March, when leisure advisors produce proposals at three times the normal rate and each wants a polished document out of Travefy or Axus by that evening. The second is the pre-departure crunch six to eight weeks before peak departures — final documents, rooming lists, manifests, dietary and mobility notes, and, if you run multilingual product, guide briefings in two or three languages per departure. The third is disruption, which is on no calendar: a hurricane, an airline meltdown, a supplier failure, a State Department advisory. Disruption weeks are the most expensive weeks per person you will ever have, and the weeks you least want anything switched off.
Budgeting AI in a travel business means setting a ceiling that matches your booking curve rather than the calendar month, and deciding in advance who is allowed to blow through it during a disruption. The rest is detail on those two decisions.
The Claude Enterprise governance update on 2 July 2026 added four things that turn AI from a surprise into a line item. A usage and cost dashboard, so you see spend by person and department instead of one number at the bottom of a statement. Spend limits at both the organisation and the individual user level — a real ceiling, not a warning. Alerts at 75 percent and 90 percent of a limit, which is the difference between finding out in time to decide and finding out in October. And model defaults and entitlements: you decide which model a group gets by default and what they may reach for.
There is also an Analytics API and an Admin API, which for a small operator matters because your bookkeeper can pull usage figures into the same spreadsheet as everything else instead of screenshotting a dashboard monthly.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent for hospitality in your browser — 60 seconds, no signup.
Compare 2024, when governance meant one company account, one card and a monthly total nobody could explain. The 2026 version treats AI the way you already treat your Sabre or Amadeus fees and your merchant account — a cost with an owner, a ceiling and a review date.
flowchart TD
A["Set the month's ceiling from last season's booking curve"] --> B["Air desk: cap set high, full model"]
A --> C["Leisure advisors: mid cap each, standard model"]
A --> D["Seasonal guides: low cap, cheaper model only"]
B --> E{"Anyone past 75 percent before the 20th?"}
C --> E
D --> E
E -->|No| F["Leave it alone, review at month end"]
E -->|Yes| G["Owner decides: raise it for the disruption, or find what is looping"]
Here is how I would set this up for a 14-person agency — eight leisure advisors, two on the air desk, a groups coordinator, two in operations and you — plus nine seasonal guides on the books April to October.
Give each role a different ceiling, because their work is different. The air desk does the most expensive work per person: schedule change queues, re-price and re-issue math, fare rules that run to eleven pages. Give them the highest cap and the strongest model by default, because a mistake there becomes an airline debit memo and no cap is cheaper than one of those. Leisure advisors get a middle cap and a standard default — proposal writing and itinerary polish do not need the heavy machinery. Seasonal guides get a low cap and the cheaper model only.
Then set the organisation ceiling by season, not by month. If January to March is triple April to June, a flat monthly cap either strangles wave season or wastes money in the shoulder. Set the wave-season ceiling at what wave season needs, drop it for the shoulder, and put a standing note in your calendar for 20 January, 20 February and 20 March to look at the dashboard for ten minutes.
Entitlements also solve a problem that has nothing to do with money. Your seasonal guides finish in October; their logins do not. The switch that says "cheaper model only, low cap" also says "off in November," which is worth doing on security grounds alone.
Stop looking at the monthly total; it tells you nothing about whether this is working. The number for the wall is cost per booking, next to commission per booking. Illustrative assumptions for the agency above: roughly 1,900 bookings a year at an average $560 in commission and net revenue.
| Period | What is happening | AI spend |
|---|---|---|
| January–March | Wave season proposals, groups quoting | $1,520 |
| April–June | Shoulder, pre-departure documents begin | $780 |
| July–August | Peak departures, guide briefings, on-trip support | $1,140 |
| September | One disruption week | $940 |
| October–December | Quiet, next-season contracting | $620 |
| Year | $5,000 |
That is $2.63 of AI per booking against $560 of commission — a hair under half a percent of revenue. Framed that way, the question is not "can we cut this" but "is half a percent buying us anything." If the air desk is catching one avoidable debit memo a quarter, or proposal turnaround fell from two days to four hours in wave season and your close rate moved a point, argue to raise the cap, not lower it. The governance tooling lets you have that argument with figures instead of instinct.
The one number that should alarm you is a person whose spend is high and flat across all twelve months. That is rarely good work — usually something running in a loop, or an advisor who discovered they can ask for a 40-page destination essay every morning.
A hard per-user limit is a blunt instrument and it will fire during your worst week. Tuesday of the storm, your operations manager hits 100 percent at 4:15 p.m. with 60 travellers still to rebook, and everything stops. That is the most likely way this goes wrong.
Still reading? Stop comparing — try CallSphere live.
See the hospitality AI agent handle a real call — complete, industry-specific, and live in your browser. No signup.
So decide the disruption rule now, in writing. Mine: any 90 percent alert during a declared disruption is auto-approved by the operations manager up to double the normal cap, no owner sign-off, reconciled afterwards. A 75 percent alert in a normal week goes to you and gets a conversation. Set that policy in July, not September.
Caps also cannot tell you whether the work was any good. The dashboard shows spend, not quality. A cheap default on your seasonal guides saves money right up to the point a guest-facing message goes out with the wrong meeting point for the 7 a.m. pickup. Keep a human reading anything that goes out under your name, and treat model defaults as a starting position your people may argue with.
And one category should never be capped: anything touching a supplier contract, a group air deposit deadline or a Seller of Travel disclosure. Low volume, high consequence — they cost almost nothing and should always run on the best thing you have.
Three things. Open the usage dashboard and sort by person for the last 90 days — most owners find one surprise in the top three, rarely who they expected. Set an organisation ceiling roughly 20 percent above your worst month so far, and route the 75 and 90 percent alerts to your own phone, not a shared inbox nobody reads. Then write the disruption exception on one line and send it to your operations manager. Forty-five minutes, and an unpredictable bill becomes a budgeted one.
Yes, but decide first whether AI is a service you provide or a cost you pass through. If contractors pay you a monthly fee, a cap that quietly throttles a top producer in February costs more in goodwill than it saves. A workable pattern: a generous standard cap for everyone, the option to buy up, and alerts going to them rather than you.
On high-volume, low-stakes work, yes — a lot. Proposal polish, review replies, first drafts of departure notes and translations of guide briefings are fine on a cheaper default. A capable model now runs roughly two dollars for the equivalent of several hundred thousand words of reading and writing, and frontier pricing has fallen about tenfold since 2025. Where defaults hurt is fare rules, penalty schedules and contracts, where a subtly wrong cheap answer costs more than the year's bill.
Run one month with no cap but full visibility, then set the ceiling at that number plus 20 percent and adjust after wave season. Guessing before you have a month of real usage produces either a cap nobody notices or one that fires on the 9th and makes everyone hate the tool.
Usually yes, on a low cap and a cheaper default: the questions they field on the ground — meeting points, dietary substitutions, "where is the nearest pharmacy in Reykjavik" — are exactly the work that keeps a guest calm without waking your duty phone. Just make the November switch-off automatic rather than something someone remembers.
One last piece of the same budget: the calls and chats that arrive while everybody is capped, busy or asleep. CallSphere builds AI voice and chat agents that answer the office line and website chat around the clock, book appointments and capture inquiries — and because that spend is per conversation rather than per person, it belongs in this same review. Give it a ceiling and judge it on the same cost-per-booking line.
Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
What an unplanned coach breakdown costs a tour operator, the engine readings that warn you first, and how to test condition-based monitoring on two coaches.
When per-seat AI pricing stops making sense for a host agency or tour operator, what running an open model actually costs in people, and the data case for it.
AI spend in retail follows floor traffic, not the calendar. Set seasonal caps, per-user limits and 75% alerts without killing your December chat agent.
AI crept to $1,900 a month across three rehab clinics. Budget per visit, cap by bucket, alert at 75 and 90 percent, and never throttle these three things.
How a travel agency owner builds a graded test set from last wave season's client threads, scores it bucket by bucket, and widens the agent's authority safely.
Casino surveillance footage cannot leave the property. On-premises AI in 2026 cuts a 90-minute disputed handpay review to nine, with no video sent to a vendor.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco
Watch how CallSphere handles real customer calls, schedules appointments, and processes payments — live.
Try Live DemoBook a DemoCalculate Your ROI