The busywork of corporate travel is going to get very cheap. The judgment underneath it will not
A vendor predicts AI will cut the cost of a travel task by 75 percent in two years. The number is the least interesting part.
Every few months the corporate travel industry produces a number so clean and so dramatic that everyone quotes it before anyone checks what it actually measures. The latest one is a good one.
A senior product executive at HRS, the corporate travel and hotel sourcing company, told Business Travel News that artificial intelligence could cut the cost of a typical corporate travel task by about 75 percent within roughly two years, a projection he was careful to label as exactly that.
I read the full interview twice, the first time as a skeptic and the second time as someone who builds this kind of thing for a living, and I came away convinced that the headline percentage is the least important sentence in it. The interesting part is everything the number quietly assumes.
I want to take the claim seriously, because I think the direction is right even if the precision is theater. But taking it seriously means pulling it apart, and when you do, the story stops being about a falling cost and becomes something more useful, which is a question about what work actually is once a machine can do the boring half of it.
The number everyone will quote, and should hold loosely
Start with the arithmetic he lays out, because it is worth seeing before you decide how much to trust it. The model runs roughly like this. The fully loaded cost of an average human task across procurement, travel management and accounting sits near one hundred dollars.
Hand the repetitive parts of that task to AI, and the human share drops to about twenty dollars, while the machine share, built on a workflow consuming around a million tokens, lands near five. Add it up and you get about twenty five dollars, which is the advertised 75 percent cut. He even sketches the tailwind underneath it, with average inference costs around fifteen dollars per million tokens today, expected to fall toward six, and a best case of one to three.
It is a tidy model, and tidy is the word I want you to be suspicious of. The same interview describes the metric in at least three different ways, as the cost of completing a task, then as the cost of saving one percent on a travel program, then as the overall cost of delivering savings. Those are not the same thing, and the slippage matters, because a 75 percent cut in the first is a productivity story while a 75 percent cut in the third is a margin story. “This is a prediction,” he says plainly, and will vary with the size of your program.
I appreciate the honesty, and I would add only that a prediction built on an inference cost that still has to fall by more than half, a token budget nobody has standardized, and a definition of “task” that moves around inside a single conversation is a direction of travel, not a line item. Quote it as a hypothesis. It is a good one. It is not a measurement.
Adoption is already high, and the payoff is strangely absent
Here is the fact that should temper everyone’s excitement, and it comes from the hotel side of the same industry. A recent State of Distribution report from NYU, RateGain and HEDNA found that more than half of hotels are already using or procuring generative AI, which sounds like a revolution in progress.
Then it found that more than 80 percent of commercial teams still spend one to two days a week on manual reporting, and that fewer than one in ten report cutting manual work by more than 30 percent. Read those three numbers together and you get the real state of play. The tools are everywhere. The results are not.
That gap is the most honest data point in this whole conversation, and it is the one the 75 percent projection skips over. Buying AI is not the same as capturing its value, and the distance between the two is not technical, it is organizational. I have made a version of this argument before about corporate travel’s real bottleneck, which was never a shortage of software. A company can adopt every agent on the market and still have its analysts spending Thursday afternoons rebuilding the same spreadsheet, because nobody redesigned the work around the tool.
The promise is a 75 percent cut. The lived reality, so far, is a majority of teams doing the same manual chores with more expensive software watching them do it.
The real move is cutting the job into tasks
So what separates the companies that will actually get the cut from the ones that will just buy the subscription? It is a single, unglamorous discipline, and it is the thing I took most seriously from an executive program on AI and organizations I did at Stanford in August, in its first cohort, alongside seventy leaders from seventeen countries.
The central move was deceptively simple. The right unit of analysis is the task, not the job. AI does not replace a travel manager or an accountant. It eats specific tasks out of their week, and which tasks you hand over, and which you deliberately keep, is the entire game.
Once you look at corporate travel through that lens, the executive’s list of candidates is obviously correct. Spreadsheet work, email administration, data ingestion, reconciliation, the mechanical execution of a hotel sourcing round that traditionally takes about five months.
These are high volume, low judgment, easy to check tasks, which is exactly the profile AI handles well. The useful framing from that program was a continuum of autonomy, from a tool that merely advises to one that acts on its own, and a simple test for where a task belongs on it, which is the cost of being wrong. A rate audit that flags a discrepancy for a human to confirm is cheap to get wrong and safe to automate aggressively. A payment released to the wrong supplier is not. Maturity is not automating everything. It is deciding, task by task, where the machine acts, where it only advises, and where you keep it out entirely.
Machines predict, people judge
There is a deeper idea underneath all of this, and it is the one that reorganized how I think about my own company. The tidiest way I heard it put was that machines predict and people judge, and the competitive advantage of the next decade will belong to whoever separates those two functions cleanly inside their own operation.
Prediction is forecasting the fare, scoring the anomaly, drafting the RFP analysis. Judgment is deciding what the program is for, which supplier relationship is worth protecting, when a rule should bend for a traveler having a terrible day. The first is getting very cheap very fast. The second is not getting cheaper at all.
This is why I find the 75 percent framing slightly backward, even where it is right. As the cost of generating an analysis collapses, the scarce thing is no longer the analysis, it is the judgment to know which analysis matters and when it is wrong. There is a quiet trap in here that the program named well, something like an architect’s dilemma.
If you let the machine do all the generating, you can slowly lose the very expertise you need in order to judge its output, until one day you are approving work you no longer fully understand. The companies that win will not be the ones that automate the most. They will be the ones that automate the prediction and invest the savings back into human judgment, which is the opposite of treating the 75 percent as money you simply pocket. I have argued before that the enduring value in this business tends to hide in the human read of a situation that no algorithm has yet matched, and cheaper prediction makes that human read more valuable, not less.
Trust is the adoption cost nobody prices in
The projection also leaves out a cost that turns out to be the real gate, and it is not computational. In a separate conversation, the chief product and strategy officer at American Express Global Business Travel argued that the thing deciding AI adoption in business travel is trust, not technology. His line was blunt and correct. “If you don’t trust it, you’re not going to use it.” Unlike the early internet, which handed you more information to weigh yourself, an AI recommendation asks you to accept its logic on faith, which means trust has to be designed into the product rather than bolted on as reassurance afterward.
That is a cost, and it is missing from the one hundred dollars. Building trust means explainability, so a traveler can see why an option was excluded and push back. It means permission based handling of personal data, because people share their preferences only when the benefit is obvious. The same executive at HRS, to his credit, flags cybersecurity as the single biggest additional risk in an agent heavy world, with guardrails, monitoring, logging and incident response all carrying real expense. None of that shows up in a token count.
So the true cost of a task after AI is not just the twenty dollars of human time plus the five of inference. It is those two plus the ongoing, unglamorous investment in making the system trustworthy enough that anyone actually lets it run. The cheap task has an expensive foundation, and the programs that forget this will be the ones still spending two days a week on manual reports while paying for AI that nobody trusts to touch anything important.
Where I sit, and where this actually gets built

I should declare my hand here, because this is precisely the problem my company spends its days on, and I would rather say so than pretend to be a neutral observer. We build these agents, aimed at exactly the repetitive tasks in the executive’s list, the rate audits, the reconciliation, the expense checks that a human could do but should not have to.
So I have an obvious stake in the direction being right, and I think it is. But building them has taught me something the projection does not capture, which is that the hard engineering is almost never the prediction. It is the governance around it.
The platform I co-founded sits on top of work that is sharp edged on one side and soft on the other. The financial tasks, payment, compliance, reconciliation, are unforgiving, where a propagated machine error is far more dangerous than an idiosyncratic human one, because it repeats at scale. The discovery and recommendation tasks are forgiving, where a wrong suggestion costs a shrug. A travel and expense platform that treats those two halves the same way will either move too slowly on the safe tasks or too recklessly on the dangerous ones.
The whole craft is putting the human in the loop at the right point for each, letting an agent flag an anomaly in an expense report while a person decides what it means, letting it hunt for a lower fare while a person owns the supplier relationship that fare sits inside. That is not a 75 percent cut you can put in a slide. It is a redesign of where decisions happen, and it is slow, specific work. I have come to believe the companies that win the next few years will be the ones doing that patient structural redesign, not the ones with the most impressive demo. The value, as usual, hides in the structure underneath rather than the interface on top.
What this means for Latin America
I will close, as I tend to, from my own region, because the cheap task is quietly good news for us. A falling cost of routine work is a leveler, and levelers favor the challenger. The large, mature programs in the older markets carry something Latin America mostly does not, which is twenty years of accumulated process, legacy tools and headcount built around manual work that now has to be unwound before any of this efficiency can land. That unwinding is painful, and it is the real reason adoption is high while payoff is low.
A newer program does not carry that weight. A company designing its travel and expense operation in São Paulo, Bogotá or Mexico City today can build task first from the start, deciding deliberately which work the machine does, which a human judges, and which nobody should automate at all, without first dismantling a decade of habit. That is a genuine structural advantage, and it fits the stubborn optimism I keep returning to about why this is an unusually good moment to be building in this region. The 75 percent, if it ever arrives, will not arrive evenly. It will arrive fastest for whoever has the least to undo, and on that particular measure, being early to the game is worth more than being big in it.
The cheap task and the expensive question
So hold the number loosely. AI almost certainly will make the busywork of corporate travel dramatically cheaper, and the direction matters more than whether it is 75 percent or 50 or 90. But the cheaper the task gets, the more the whole thing turns on the part that stays expensive, which is the judgment about what to do with all that freed up capacity.
The programs that treat the savings as money to pocket will quietly hollow out their own expertise. The ones that treat it as budget to reinvest in the human read of a hard situation, the thing travelers have always quietly refused to give up, will come out the other side both cheaper and smarter. The machine will soon be very good at answering the routine questions. The advantage now belongs to whoever still knows which questions are worth asking.
About me
I am an entrepreneur with over 20 years of experience at the intersection of tourism and technology. I am co-founder and Chief Business Officer of VOLL, the largest mobile-first corporate travel and expense management platform in Latin America, where I lead the commercial strategy behind our AI adoption.
A first-cohort graduate of The AI-Powered Organization at Stanford Graduate School of Business and a Marketing specialist from Fundação Dom Cabral, I serve on the Tourism Council of FecomércioSP and on the Executive Council of the Latin American Association of Corporate Events and Travel Management (Alagev). I write and speak about innovation, digital transformation, entrepreneurial leadership, and the future of corporate travel.








Discussion about this post