The order of commitments
By Philip Rothaus
Every week the frontier makes a new case for waiting: why build or connect anything of your own if your customer’s agent will soon do the work? But people, markets and regulators move far slower than the models do. Commit first to what pays for itself now, plan in parallel for the world where it matters less, and get the data ready for both.
The text
The text came in on a Saturday morning, from a friend who is an AI leader at a large consumer company.
He’d been planning a big trip with an AI assistant. Not asking it questions, but handing it the job: options, logistics, bookings coordinated across half a dozen places. It had been so good it unsettled him. His question, more or less: if this is where things are going, shouldn’t we stop what we’re doing and put everything into making ourselves easy for other people’s agents to talk to?
It’s a good question. It’s also the question I hear most often right now, in one form or another, from people who have just had their first real experience of an agent doing something for them.
My answer was yes, and.
He was really asking which commitment comes first.
Two ways to get the future wrong
The first is to wait. If the frontier models and personal agents are going to absorb everything, why build anything of your own? Let the platforms fight it out, and plug in when the dust settles. This is the corporate version of every confident, early “SaaS is dead” post. It isn’t wrong about the direction. It’s wrong about the timing, and in business the timing is most of the argument. You give up years of value because you’re afraid of a future that hasn’t arrived yet.
The second is to build as if the frontier isn’t coming. You pour everything into a channel, an app or an experience your customers may stop visiting in person, and you treat the agents knocking at the door as a curiosity, or as an unsettling reality best handled with your head in the sand.
Both mistakes come from looking at one horizon. The first looks only at five years out. The second looks only at this quarter.
The useful question is whether you can operate in the middle: building for the customer in front of you while planning for the one who sends an agent instead. That depends on getting two things right before anything else: how wide to look, and how far out the effects of a decision will land. The order of commitments follows from those.
How fast people actually move
Here’s what my friend already knew, and what was easy to forget in the glow of a well-planned trip: he is not normal. Neither am I.
About half of American adults don’t use AI chatbots at all. Roughly a quarter use one daily, and 4% say they use one “almost constantly” (Pew Research Center, June 2026).
That 4% is the closest public number I’ve found to the people I’m talking about. When I talk with friends and prospective clients, including the ones who tell me they use AI every day, what they describe is almost always the same: drafting and editing documents, one prompt at a time. Nothing happens unless they ask. No code. Nothing scheduled. No multistep process running while they sleep. That’s real value, but it’s a long way from handing an agent a trip and walking away.
So my friend is in the 4%. Nearly everyone else is still well behind, and most of them aren’t in a hurry.
There’s a second pattern hiding in the examples, and it matters more than the adoption numbers. Almost every vivid story about agents acting for people is about discovery: finding, comparing, shortlisting, planning. Very few are about the transaction, the moment money moves and something becomes binding. That’s where the trust still sits, and it’s where the revenue sits too. People will let an agent narrow ten options down to three long before they let it spend on their behalf without looking. Many will still click through to check the details themselves, for years yet.
Trust is only part of it. Some customers don’t know what’s possible yet. Some have tried it and weren’t impressed. Some are resisting AI altogether, on principle or out of unease. The point isn’t where customers might be in five years. It’s where they actually are now.
The order of commitments should start there.
But that’s today’s order, not a permanent one. We don’t yet know how people’s relationship with their agents will change: what they’ll come to hand over, what they’ll insist on keeping for themselves, and whether the parts they keep turn out to be the parts that matter most. That deserves its own memo. For now, treat the line between discovery and transaction as something to watch, not something to build a wall around.
Regulators keep their own pace as well, and it isn’t the models’ pace. Rules about disclosure, consent and who is accountable when an agent acts are still being written. Any plan that assumes they’ll be settled soon is a bet, not a plan.
The order
First: commit to what pays for itself now. What that is depends on who you are. If you have the money, the people and the know-how, it may be your own agent, in your own channel, for the customers who are actually showing up. Where the economics work, my estimate from conversations with people building these systems is that this kind of investment can pay for itself within a few years, and that broad adoption of personal agents is at least that far off. If you don’t have those resources, and many organizations don’t, building your own may be the wrong first commitment. Plugging into what the frontier companies already offer, or simply opening yourself up so their agents can work with you, may pay back faster and cost far less. Either way, the test is the same: does it return value before the future arrives? And either way, you have to be honest about whether you can actually operate in the middle, or whether you’re pretending to.
In parallel: plan for the world where your agent isn’t the front door. Not instead, alongside. Meet customers where they actually are, not where the most excited of them will be. For most, that means letting agents help with finding and comparing before anything binding happens. What happens when agents start negotiating with each other is its own subject, and a later memo.
Underneath both: the data. This is the no-lose commitment, and the oldest lesson in the room, running at a new speed. An agent is only as good as the data and tools you expose to it, and that is true whether the agent is yours today or someone else’s tomorrow. Clean, current, well-described data, with access you can grant and take back, is the one investment that pays off in every version of the future. Companies used to learn this the slow way, building models on data that wasn’t ready. Now the models arrive ready, and the data is what holds things up. Do the work once and it serves both your agent and everyone else’s.
When the tools outrun the plan
On our approach page we wrote about an agent build where the team shipped so much faster than planned that the buy-versus-build analysis we meant to run had to wait. It’s still on the table, but it now gets measured against what’s already built and how close launch is.
I tell that story here because the order of commitments isn’t set once. The inputs keep moving: model prices, coding tools, what customers will tolerate, what regulators expect. A sequence that was right in the spring can be wrong by the fall. Put the order of your commitments on the calendar and revisit it every quarter, the way you would a forecast.
The build is the research
One more thing, which I didn’t say in my reply but should have.
Building the thing is how you learn what’s coming. The team that has put an agent in front of real customers knows things the team still waiting for clarity can’t: where people trust it and where they don’t, which data was never as clean as everyone assumed, what the regulators actually ask about. You can’t see the horizon from the sidelines.
The companies waiting for the future to become clear will be the last to see it.
The same question, inside the building
Everything above is about agents that face customers. The same order applies to the agents you point inward, at your own operations, finance, HR and engineering, and there the questions of control and trust are sharper still. A customer can ignore an agent’s suggestion. An internal agent with access to your systems can act on it. Commit first to what pays back, plan for platforms you won’t control, and get the data and the permissions right underneath both. But go slower on how much you let it do on its own.
Yes, and
So: yes, start making yourself easy for other people’s agents to talk to. And keep committing to what pays for itself, whether you build it or borrow it, while the rest of the world catches up to my friend’s trip planning.
For the record, when an agent proposes something to me, I still look at it myself before I say yes. Not because I’m slow to adopt. Because I know agents still make mistakes, and because, even when it’s acting on my behalf, I’d like some actual say in the choices it makes.