Here’s a problem that sounds simple but isn’t.
You own a 40-person company. One of your salespeople, Dana, needs to fly from Chicago to Austin next month to close a deal. You’d like Dana to spend as little of the company’s money as possible getting there. That’s it. That’s the whole goal. Spend as little money as possible.
So you write a travel policy:
- Flights under $400.
- Hotels under $250 a night.
- Book seven days in advance. Economy only.
You feed those rules into your expense management software, and now you have a system that will dutifully flag Dana’s $480 flight and $300/night hotel and route them to a manager for approval. Congratulations, you have automated your travel expense policy!
But watch what happens in the real world.
The $340 flight that fits your policy lands at 11:45 PM and connects through Dallas. The $480 flight is nonstop and gets Dana in by 2 PM, with enough runway to prep for a 9 AM meeting that — let’s be honest — is the entire reason for the trip. The in-policy hotel is a 40-minute drive from the client’s office, which means a rental car, which means Dana white-knuckling an unfamiliar highway at rush hour, which means the $300 hotel two blocks from the client is actually cheaper once you count the car, the gas, the parking, and the non-zero chance that Dana shows up frazzled and blows the meeting.
The cheapest option, in other words, was the most expensive option. And no policy you could have reasonably written and coded into your expense management software would have known that in advance.
People who do this for a living have a name for what you’re actually chasing here. It’s not the lowest fare. It’s the lowest logical fare — the cheapest option that still makes sense once you account for everything the price tag doesn’t capture. Where the client is. When the meeting is. How much Dana’s time is worth (and how annoyed she’ll be if you waste it). Whether a four-hour layover is a reasonable trade for ninety dollars. Whether this particular customer is worth showing up rested for.
The “lowest” part is easy. Software has been able to sort by price since approximately forever.
It’s the “logical” part that has broken every expense management tool ever built, because logic, in this context, is not a rule. It’s a judgment call that depends on a dozen variables that change with every trip, every traveler, every customer, every quarter. You cannot hard-code it, because the moment you write the rule down, reality produces an exception that makes the rule look foolish.
For thirty years, we’ve handled this gap the only way we knew how: with humans. A manager eyeballs the booking. A finance person catches the weird one. An exhausted admin builds a spreadsheet of approved exceptions to the policy that already had exceptions baked in. The software enforced the number, and the people supplied the context-dependent logic. It worked, sort of, in the way that things work when you’ve never seen the alternative.
Now we do. And I would argue that travel — this small, unglamorous, deeply annoying corner of corporate finance — is the single clearest window we have into what the entire next era of financial software is going to be about.
A Framework for Assessing the Value of Agentic AI
Let’s start by getting the obligatory AI throat-clearing out of the way.
You already believe AI agents are going to be a big deal. I don’t need to sell you on it. If you work anywhere near fintech in 2026, you have sat through enough keynotes, read enough thinkpieces, and nodded along to enough “this changes everything” LinkedIn posts that the bigness of it all has become a kind of ambient hum.
Agentic AI is transformative.
Yep. We all agree.
The problem is that “everyone agrees it’s transformative” and “anyone can tell you precisely where it adds value” are two very different things. The honest version of the current moment is that we have a powerful new capability and a strong intuition that it matters, but we are still mostly waving our hands when it comes to developing an actual framework — a consistent method for looking at a given workflow and saying, with confidence, this is where an AI agent can add value, and this is where you’re just lighting tokens on fire to do something that existing software or employees handle just fine.
This is the gap I think business travel fills so beautifully. Because lowest logical fare isn’t just a travel problem. It’s the cleanest example I can find of an entire class of problems — the kind where the goal is obvious and the path is not.
For this class of problems, you have a workflow where:
- The objective is crisp and easy to state. Spend less. Get paid faster. Buy the right thing. Everyone agrees on what “good” looks like in the abstract.
- The constraints are fuzzy, contextual, and resistant to being written down. What counts as a “logical” exception depends on facts that live outside the system, change constantly, and require judgment to weigh against each other.
- The cost of getting it wrong is real but diffuse. It shows up as wasted money, wasted time, a blown meeting, an annoyed employee, which means nobody can quite justify hiring a full-time person to optimize it, but everybody quietly knows it’s leaking value.
When you see a workflow with those three traits, you have found a place where agentic AI does something genuinely new.
Not “faster.” Not “cheaper.”
New.
Because the thing an AI agent can do that rules-based software cannot is hold the messy context and the crisp goal in mind at the same time, and reason about the trade-off the way Dana’s manager would, except at the speed of software and across every booking, every time, without getting tired or going on vacation.
That’s the framework — crisp goal, fuzzy constraints, and diffuse cost.
Travel is the poster child. But once you have the pattern in your head, you start seeing it everywhere.
Why B2B Won’t be a Laggard This Time
Almost all waves of technology disruption in financial services follow this pattern: it shows up in consumer products first, and gets dragged into business products later, often kicking and screaming.
I saw this, up close, with the introduction of modern smartphones. The iPhone landed in 2007 as a consumer device, and for years afterward, IT departments insisted that real companies ran on BlackBerry because BlackBerry was secure and serious and the iPhone was a toy.
We all know how that movie ended, and the reason it ended that way is worth considering carefully: the employees were also consumers, and once they’d tasted the better experience at home, they were never going to accept the worse one at work. Consumer-grade experience eventually flowed uphill into the enterprise, because experience is contagious and people carry their expectations across the boundary between their personal lives and their jobs.
That pattern — B2C first, followed (slowly) by B2B — has held for mobile, for instant payments, for digital onboarding, for basically every “why can’t my bank be as easy as Venmo” complaint of the last fifteen years. New experiences win with consumers first because individual consumers ruthlessly optimize for experience. Businesses are slower, because businesses don’t prioritize how it feels for employees to use the software; they prioritize control, security, and not getting fired for picking the wrong vendor.
I think agentic AI might break this pattern.
Previous waves of technology disruption centered on convenience. A better-feeling experience. Fewer taps, prettier screens, less friction.
But convenience is not what agentic AI is for.
I mean, it’s nice — an intelligent and engaging chat interface is pleasant — but a nicer experience is not agentic AI’s superpower. Its superpower is context. Fuzzy matching. Weighing competing constraints. Dynamically reshaping a workflow as conditions change. Solving the “logical” half of lowest logical fare. And that capability isn’t just a consumer-experience upgrade. It’s also a powerful lever for business productivity.
Agentic AI’s value doesn’t concentrate where the experience is worst. It concentrates where the context is thickest; where there are the most competing variables, the most judgment calls, the most expensive exceptions. And that describes business workflows far more than consumer ones. Your personal travel has, what, two stakeholders and one budget? A company trip has an employee, a manager, a finance team, a policy, a client, a calendar, and a P&L, all pulling in slightly different directions. That’s not just more inconvenient. It’s more contextual. And parsing context is the thing that AI agents are uniquely good at.
Convenience flows from consumer to business.
Context, I’d argue, might flow the other way because the density of contextual, judgment-laden, expensive-to-get-wrong decisions is simply higher inside a business than inside a person’s life.
Which means B2B might not drag its feet this time. B2B might be standing at the front of the line.
The Goldilocks Zone
But what exactly do I mean when I say, “B2B”?
A two-person Etsy shop and a 200,000-person bank are both technically B2B, and they have approximately nothing in common. So let me get more specific about where, within the vast expanse of business, the context-versus-convenience logic bites hardest.
It’s the middle. Small and medium-sized businesses. The Goldilocks zone.
On the one hand, a solopreneur or a tiny side hustle doesn’t really have a lowest-logical-fare problem, because they barely have the problem at all. The level of spend isn’t worth the overhead to optimize it. The context exists, but there’s not enough of it, and not enough money at stake, to matter. Agentic AI is a luxury, not a need.
On the other hand, the Fortune 500 enterprise has the lowest-logical-fare problem in abundance, but it also has something that SMBs will never have: raw capacity. It can negotiate custom rates with counterparties. It can build (or buy, or rip out, or rebuild) a sprawling stack of specialized tools, each tuned within an inch of its life for one narrow category, and then staff those tools with departments full of people whose job is to squeeze the last few points of efficiency out of them. For the giant, throwing bodies and bespoke software at a problem is not just possible; it’s frequently rational, because at that scale, even a 2% improvement on a huge number is a huge number.
SMBs sit in between, and the in-between is exactly where agentic AI stops being a nice-to-have and becomes the only sane answer.
Editor’s Note — I wrote more about the unique perspective that small business owners have when it comes to AI and other forms of automation in a different sponsored article with BILL here, if you’re interested in going deeper on this point.
A 40-person or 400-person company has plenty of context: multiple travelers, real policies, actual clients, a budget that bleeds when you ignore the logic. The problem is fully present. What’s absent is that enterprise capacity to brute-force it. There is no corporate travel department. There is no team of analysts. There’s a finance person — maybe two — who are already doing nine other jobs, and a founder who would very much like to stop personally approving hotel bookings.
For that company, you cannot solve the lowest-logical-fare problem by throwing humans at it, because there are no spare humans. And you cannot solve it with rules, because we already established that rules can’t hold the context-dependent logic. The only thing left that can hold the crisp goal and the fuzzy context at the same time, at a price the SMB can stomach, is an AI agent.
That’s the sweet spot. Enough complexity to need the intelligence. Not enough scale to fake it with headcount. It is, not coincidentally, the slice of the market where the next decade of financial software is going to be won or lost.
Context-dependent Judgement Calls All the Way Down
Travel is the most visceral example of where AI agents can help in the SMB back office because everyone has suffered through at least a couple bad work trips. But the reason it matters for where financial software is heading is that the same opportunity shows up across nearly every workflow an SMB runs through its finance function.
Let’s quickly review a few examples:
Travel & Expense (T&E). As we’ve already covered, there’s crisp goal (spend less), fuzzy constraints (what’s logical), and diffuse cost (blown meetings, wasted hours, etc.) An AI agent that knows where the client is, when the meeting is, what the traveler’s history looks like, and what the budget can bear can reason its way to the lowest logical fare in real time. More importantly, it gets better at a company’s specific definition of “logical” the more trips it sees.
Editor’s Note — As BILL explained in a recent blog post, T&E is a top priority for intelligent automation due to its rapid growth, as a category of expense for small businesses. According to BILL survey data, SMBs spent, on average, ~50% more on travel in 2025 than they did in 2022, even though cost per trip stayed the same. This suggests that small businesses are traveling much more than they used to.
Procurement. Same basic shape as T&E. The crisp goal is “buy the right thing at the right price.” The fuzzy constraints are everything that makes “right” depend on context: Is this the vendor we have a relationship with? Is the cheaper supplier actually cheaper once you count reliability and net terms? Does this purchase need three approvals or zero? A rules engine can handle the dollar thresholds, but it falls apart when the fuzzier constraints get applied. An AI agent can weigh the trade-offs — preferred vendors, payment terms, the fact that the marketing team always under-orders and then expedites at 3x cost — the way an experienced ops person would.
Accounts Payable (AP) and Accounts Receivable (AR). The goal is crisp on both ends: pay what you owe at the optimal time, get paid what you’re owed as fast as possible. The constraints are pure context. When do you pay this bill? Early to capture a discount, or late to preserve cash? Which matters more given where your balance sits this week? Which overdue invoice do you chase first, given that this customer always pays on day 47 and that one is genuinely at risk? This is judgment, performed thousands of times a year, currently done by an overworked human operating off of an aging spreadsheet and intuition. It is tailor-made for an AI agent that can hold the whole picture.
These are just the most obvious examples, but there are plenty more. Categorizing transactions for the books. Flagging the subscription nobody remembers signing up for. Deciding which card to put a purchase on. Forecasting next month’s cash position and adjusting today’s decisions accordingly.
Every one of these has the same fingerprint: a goal everyone agrees on, a set of constraints nobody can fully write down, and a cost of getting it wrong that’s real but too diffuse to justify a dedicated hire. The whole back office of a small business turns out to be one context-dependent judgment call stacked on top of another, all the way down.
Why Best of Breed is No Longer Good Enough
In each of the use cases listed above, the AI agent gets better the more context it can see. The travel agent that can also see the cash position makes better travel decisions. The AP agent that can also see the budget makes better payment decisions.
This matters because every lowest-logical-fare problem that an SMB has is much bigger and more context dependent than you might imagine.
Suppose you build the world’s best agentic travel product. It’s brilliant. In a narrow sense, it nails the lowest logical fare every time. But it only knows about travel. It doesn’t know that the trip Dana’s taking is to close a deal that, if it lands, blows past the quarterly budget and changes what “spend less” even means this month. It doesn’t know that cash is tight this week because a big receivable is late, which should make it more aggressive about the cheap flight. It doesn’t know that this client is the company’s third-largest account and therefore worth the nonstop. All the context that makes the fare logical lives in other systems — in AP, in AR, in the budget, in the CRM — and the brilliant AI travel agent can’t see any of it.
This is the structural problem with point solutions in an agentic world, and it’s a genuinely new problem. In the old world, a best-of-breed travel tool and a separate AP tool and a separate procurement tool was a perfectly defensible architecture, because each tool was just enforcing its own rules within its own four walls, and humans stitched the context together across the seams. The people were the integration layer.
But the entire premise of an AI agent is that it reasons using context. An AI agent that can only see one workflow is an AI agent reasoning with one eye closed. It will confidently optimize travel while being blind to the cash position that should have changed its answer. The narrower the surface it sits on, the dumber it is. Not because the model is weak, but because it’s starved of the very thing that makes it valuable.
In a world of rules, fragmentation was fine and maybe even optimal — each tool minded its own business, and best-of-breed always beat all-in-one.
In a world of AI agents, fragmentation becomes a tax on intelligence. The value of the AI agent is a function of the breadth and quality of the context it can access, so the platform that can put travel, expense, AP, AR, procurement, and the budget on a single substrate — one data layer, one set of connected workflows, one unified picture of the business — isn’t just more convenient. It’s categorically more intelligent, because its AI agents can see the whole board.
This represents a sea change in the world of fintech infrastructure. For twenty years, the smart money was on best-of-breed: pick the best tool for each job and integrate them. Picking a product suite offered by a single provider was the lazy choice, the thing you settled for. Agentic AI inverts that dynamic. The suite — or more precisely, the unified platform with a shared data layer — stops being the compromise and starts being the only architecture that lets the intelligence work at full strength. Context is now the product, and context doesn’t survive being chopped into a dozen disconnected tools.
Where It Stops, Nobody (Yet) Knows
I want to end on a big question I actually can’t (at this moment) answer.
For decades, there’s been a predictable arc to how businesses scale. You start small and scrappy, running everything through one or two general-purpose tools because that’s all you can afford and all you have time for. Then you grow, and as you grow, you specialize. You break the all-in-one system apart. You hire a dedicated travel manager, a procurement team, an AP department. You buy best-of-breed tools for each and staff them up. You do this because at sufficient scale, the specialization pays. The last few percentage points of efficiency, squeezed out of a hyper-tuned siloed system, is worth real money when the numbers get big enough.
The unified platform, in other words, has historically been a phase you grow out of. A starter home. Useful when you’re small, but abandoned when you can afford something custom-built.
However, that arc assumed something we no longer have to assume: That the only way to get the last few points of efficiency was to specialize and throw humans at it.
If an AI agent on a broad, unified platform can hold all the context and make all the judgment calls — and get better at it the more of the business it sees — then the efficiency advantage of breaking the platform apart starts to look less attractive.
If the next generation of SMBs grows up on a unified, agentic financial platform — if they get their first taste of “the software just handles the logic” while they’re small — do they ever have a reason to graduate off it as they get big?
Or do they just keep scaling the intelligence on the same substrate, watching the AI agents get smarter as the business gets more complex, never hitting the point where fragmenting into specialized silos makes sense, because the thing that used to make silos worth it has been automated away?
Put more pointedly: For the entire history of business software, the platforms that served small companies and the platforms that served large ones were different platforms, and the transition between them was a rite of passage (one that was worth a fortune to whoever could capture it). If agentic AI collapses that distinction — if the platform a company starts on is also the platform it never needs to leave — then the most valuable real estate in B2B financial software isn’t a product category at all. It’s the unified ground floor that the AI agents are built on top of.
It’s too early to say if that’s where all of this is heading. We are currently witnessing something unprecedented: Software is learning to do the thing we have always needed humans for; to hold a clear goal and a messy reality in mind at once, and find the solutions that are not just cheapest, but the most logical.
Where that ends up, nobody knows yet. But it’s going to be one hell of a trip.
About Sponsored Deep Dives
Sponsored Deep Dives are essays sponsored by a very-carefully-curated list of companies (selected by me), in which I write about topics of mutual interest to me, the sponsoring company, and (most importantly) you, the audience. If you have any questions or feedback on these sponsored deep dives, please DM me on Twitter or LinkedIn.
Today’s Sponsored Deep Dive was brought to you by BILL.
BILL is the intelligent finance platform trusted by nearly half a million businesses and their accountants to manage, move, and maximize their money. BILL powers businesses ranging from fast-moving startups to growing companies with complex operations. We use AI to deliver strategic finance capabilities in one integrated platform that includes AP, AR, expenses, forecasting, procurement and more. With a member network of more than 8 million, BILL’s platform processes ~1% of US GDP annually. To learn how BILL can help, click here.
