Selling through machines to people
For years, most of the advice I heard about growing a business was about marketing. Agencies, ads, funnels, SEO, and lately a new AI trick every month. Some of it worked. Very little of it lasted.
What produced the most durable growth I’ve seen was more basic, and a lot less exciting. We made the operation better. Then we made the truth about the operation easy to see.
At Adventure Idaho we trained guides to a standard and kept training them. We kept raising the safety bar. We rewarded planning over heroics. We learned what the operation could really carry on a given day, and sold to that. We listened to what customers actually cared about, and fixed the parts of the experience that fell short first.
None of that is marketing. All of it ended up doing the work marketing is supposed to do.
I’ve been trying to work out why. This is where I’ve landed so far.
Selling through machines to people
More and more, the first thing that meets a customer isn’t the business. It’s a system that has already formed a view of the business.
Someone asks a search engine, an assistant or an app for a good option. A machine reads what it can find, decides what it believes, and hands back a short list. Most businesses never learn they weren’t on it.
You’re increasingly selling through machines to people.
The machine decides whether you get into the consideration set. It doesn’t decide the rest. A person still decides whether to trust you, whether to buy, whether it was as good as it looked, whether to recommend you, and whether to come back.
That second half is where the operation lives, and it feeds the first half. A good operation leaves evidence behind: specific reviews, clear answers, people describing the same experience in their own words. Machines can read that. Customers find you through it. They have a good experience, because the operation is good, and they leave more evidence behind.
Good operation → evidence of it → machines understand it → customers find it → a good experience → more evidence.
It runs in reverse too. A mediocre operation leaves mediocre evidence, and clever positioning has a hard time holding up once machines are reading everything customers say. I wrote about one version of this earlier: the lever wasn’t more content. It was an operation worth describing, described plainly. The specifics of how we did that are ours, and honestly they’re the least durable part.
Generic knowledge got cheap
There’s a second reason this matters more now.
Generic knowledge is close to free. Anyone can produce another article called “Five tips for running a better business” in seconds, and the internet is filling up with them. So are the systems reading the internet.
What can’t be generated is knowledge that came from doing the work. The kind you only get by running something:
- why the capacity you appear to have isn’t the capacity you actually have
- where a process fails, predictably, every season
- which training changed outcomes and which only felt productive
- which exceptions need a person, and which only look like they do
- which constraints only matter when they show up together
- what customers actually complain about, as opposed to what you expected
- which parts of the operation are fragile
If AI makes generic information abundant, I think it makes that kind of knowledge worth more, not less. It’s hard to fake, and it’s most of the difference between a business worth recommending and one that only sounds like it.
It’s also the hardest thing to get out of anyone’s head, including mine. That’s most of what instrumenting a business turned out to be about.
From readable marketing to readable operations
Everything so far is about what a business says about itself, and whether machines believe it. That stage is already here.
The next one is different, and I don’t think most businesses are looking at it yet.
There’s a large gap between a website saying what a business offers and software that holds the operating truth of the business. A website can say you run half-day trips on Saturdays. It can’t say whether this Saturday has a guide qualified for that stretch of river, a vehicle to get the group there, gear in their sizes, and room left once everything already sold is counted.
For a person, that gap is fine. They call, and someone checks.
An agent acting for a customer won’t call. It will need to know, directly:
- What does this company actually sell?
- What does it cost, for this group, on this date?
- Is it available?
- Does the company actually have the capacity to deliver it, or just an open slot?
- Is this customer eligible?
- What constraints apply?
- Can I reserve it?
- Can I change or cancel it later?
- Can I do all of that without inventing data or creating a duplicate someone has to clean up?
Most of those answers don’t live on a website. They live in a booking system, a schedule, a spreadsheet, a group text and somebody’s memory, often all at once and slightly out of step with each other.
Capacity is where it gets hard
The question I keep coming back to is capacity, because that’s where the gap is widest.
Capacity is not a number. An open seat, an open appointment, an idle employee, a parked vehicle, an empty room: none of those means the business can take one more sale. Real capacity is the intersection of every constraint that has to line up at the same moment. For us that’s guides and what they’re qualified to run, boats, vehicles, drivers, permits, timing, and everything already sold into the same day.
A person doing the scheduling learns to see that intersection over a few seasons. An agent looking at an availability field sees a number, and believes it.
That’s harmless while the agent is only showing search results. It stops being harmless when the agent books on someone’s behalf. Then the gap between apparent capacity and real capacity becomes an overbooking, a cancellation, a refund, and a customer who trusted the machine and was let down by the business.
So the operating model has to be explicit, and it has to live in the software, not only in the head of whoever does the scheduling.
Readable isn’t the same as usable
An agent that can read your operation is useful. An agent that can act on it (reserve a spot, change it, cancel it) without breaking anything is a different kind of customer.
That takes more than a better website. It takes the website, the booking system, the internal database, the staff, the customers and the agents all working from the same canonical truth, instead of each keeping its own copy.
Most small-business software wasn’t built for that, and that isn’t a criticism. It was built to record transactions for people, and it does that well. The side where an outside system can safely act on the operation is mostly still missing. I went into where that wall sits in more detail.
And the actions have to be governed. I don’t think agents should be able to do whatever they like. The real design problem is deciding which actions are safe to hand over, which need a person to confirm, and which should never be automated. That’s an operating question before it’s a software one.
Operational infrastructure is becoming distribution infrastructure
Put that together and something shifts.
A booking system, a CRM, an inventory database, a scheduling tool: for most of their history these were back-office software. They mattered to the people inside. Customers never saw them.
As agents become buyers and go-betweens, that changes. The quality and openness of the operational layer starts to decide whether a business can take part in the new distribution layer at all. If an agent can’t get a trustworthy answer about what you can actually deliver, it will send the customer to someone it can get one from.
Operational infrastructure is becoming distribution infrastructure.
That doesn’t make marketing irrelevant. It moves the foundation underneath it. I don’t think the small businesses that do well here will be the ones that added AI to a mediocre operation. I think they’ll be the ones where:
- the underlying operation is good
- its operating knowledge has been made explicit
- its software holds an accurate model of how the business really works
- people and machines can both question that model
- agents can eventually take governed actions against it
The order matters. You can’t make explicit knowledge the operation doesn’t have, and you can’t safely let an agent act on a model that’s wrong.
Where I’ve landed, for now
I’m not claiming to have this figured out. Most of what I’ve described I’ve watched from inside one operation, and I’m still working out which parts travel.
But this is what running an operation has made me pay attention to, and it’s where the pieces are starting to connect for me.
The businesses I’m most interested in aren’t the ones trying to bolt AI onto everything. They’re the ones with a real operation underneath, a clear model of how that operation works, and software that lets people, and increasingly agents, work with that reality directly.