Justin Smith

Everyone bought AI. Almost nobody connected it to the business.

I didn’t realize how unusual it was until I started talking to other companies.

At Adventure Idaho, I can be pumping gas, say one sentence into my phone, and the operation changes. A guide’s day off gets logged. The trips that guide was on get flagged. The schedule checks itself for collisions. The crew board updates. A person confirms anything that touches money or safety, and that’s it.

That’s a 44-person crew, up to 272 people on the water in a single day, and a business that grew from 1,279 guests a year to more than 4,700, run from an office of three. For two seasons I assumed that was just what AI does now. It isn’t. It’s what AI does when it can reach the whole business, and almost nobody’s can.

It isn’t the software. It’s the system.

People assume I built some software. The software is the small part. What I built is a working model of how the whole operation fits together: which guides, boats, vans, trips and permits exist, how they depend on each other, and what’s allowed to change them. It’s 44 connected tables and about 207,000 data points. The booking software kept doing what it’s good at, recording the sale. Everything else lives in one place we own, and the AI can read it and write to it inside rules a person set.

It isn’t one app, either. The model sits in the middle, and dozens of automations and connectors, Zapier zaps, scheduled jobs and webhooks, tie it to the platforms where the work actually happens: the booking system, Slack, the phone system, email and forms. Real work still happens in real platforms. Guides still text, the office still answers the phone, and bookings still land where they always have. The system’s job is to connect them, and deciding what to automate and what to leave to a person was most of the skill.

There were no IT people or vendors between me and the work, so I could design it end to end.

Once the model existed, the software came fast. The crew got web apps, phone apps and dashboards built on it. The office whiteboard went digital. Staffing, fleet, gear and food all got tracked and tuned in one place. No more dozen notebooks, no more spreadsheets, and no more depending on somebody’s memory.

What an office of three can run

In most businesses the day gets eaten by questions. Who worked which days? Who went on which trip? Who is qualified to go where? When is that van back? Every one is a person stopping to look something up, or to find the one person who knows. Once those facts and the rules connecting them lived in one place, the operation could speak English. Anyone can ask it a question. I can update it in a sentence.

It also catches a person or a van committed in two places at once, puts unpaid balances in front of someone every morning before the trip launches, and sizes the food for every trip. At a bigger company that’s a dispatcher, a scheduler, a fleet coordinator, collections and an analyst. Here it’s one system, with a person checking the parts that matter.

The office went from two people to three while the business more than tripled. As the season got busier, those three spent less of their day on admin and more of it driving shuttles, which put more of our guides on the river.

Why almost nobody else can do this

It wasn’t the model. I use the same AI anyone can buy. It worked because the operation lived in one place, I controlled that place, and the AI was allowed to work there inside rules a person set. In a typical company the schedule is in one app, the customers are in another, the invoices are in a third, and the rules are in someone’s head.

Over the last few months, a few friends and I have been talking with a lot of companies outside the river: private equity firms, holding companies, and the IT departments that would have to sign off on any of it. From every side we hear the same thing. People are hitting wall after wall trying to make AI useful.

IT says no, and people go around it. Menlo Security found 68% of employees using free personal AI accounts, and 57% of those putting in sensitive company data. It isn’t new: in 2024, Microsoft and LinkedIn found three in four knowledge workers already using AI at work, most of them with their own tools. Nobody’s doing it to cause trouble. They have a pile of work, and the AI on their phone gets through it faster than anything IT approved, so a customer list or a pricing sheet gets pasted into whatever is open in the next tab. Insurers have noticed, and carriers have started adding generative AI exclusions to general liability policies.

Companies bought the AI, not the connection. Buying ChatGPT or Claude for everyone gets you the model, not your business. MIT found about 95% of enterprise generative AI pilots showed no measurable impact on profit and loss, and traced it to how the tools were fitted into the business, not to the models. Gartner expects companies to abandon 60% of AI projects that aren’t backed by AI-ready data.

The data is getting harder to reach. Salesforce changed Slack’s API terms to stop outside AI tools from storing and indexing company messages, and SAP now bars outside AI agents from its data unless it approves them.

Put those together and you get the worst of both. The approved AI can’t reach the data. The unapproved AI is getting it anyway, one paste at a time.

IT isn’t wrong to be careful. An AI that can write into your systems is a new hire with no training and every password. The answer is to decide exactly what it can touch, have a person approve anything that matters, and keep a record. Ours has an audit trail on every change: what changed, who asked for it, and whether a person or the AI made it. Its access is scoped to the job, and the keys stay locked down. That work is needed and it has to be done right, but it isn’t the hard part. The hard part is the rebuild that comes after it.

The models are ready. The businesses aren’t.

Get far enough into these conversations and someone asks whether it’s all a bubble. There’s a real case. Sequoia’s David Cahn calculated that the AI buildout needs to earn about $600 billion a year to pay for itself, and some of it will probably turn out to be too much, the way fiber was in 2000.

But economists have a name for what’s happening inside companies: the productivity J-curve. With a technology this broad, businesses first spend years rebuilding their processes and data around it, and productivity looks flat or worse while they do. The economist Paul David showed that factories took decades to redesign themselves around electric motors before the gains showed up. The gap between what AI can do and what it’s doing inside most companies isn’t hype. It’s integration work that hasn’t been done yet. I’ve seen what’s on the other side of it.

What comes next

In May, OpenAI and Anthropic each launched ventures backed by private equity firms to put their own engineers inside companies and do the installation work. Two of the most advanced AI companies in the world reached the same conclusion: the bottleneck isn’t the AI. It’s getting it connected inside the business, and that takes people on site.

Those ventures are built for very large companies. The trades and service businesses inside most portfolios, the ones with thirty trucks and a founder who knows everything, won’t get that attention. For a business that has just been bought, connecting it is the first thing worth doing, because the operation usually still lives in the founder’s head. For a company that owns several, doing it the same way in each one is how one good install becomes an advantage across all of them.

The work takes judgment more than code. Someone has to decide what the AI should do on its own and what stays with a person, which parts of the business are worth modeling as data, and which are better handled by hiring someone good and trusting them in the field. Get those calls wrong and you build something nobody trusts, or spend months modeling things that never needed it. That judgment can’t be exercised from a slide deck. It takes someone deployed inside the business, building next to the people who do the work, which is how I work. If you’re running into this wall, tell me what’s falling through the cracks.

I know The Innovator’s Dilemma well, and this looks like the pattern Clayton Christensen described. Incumbents don’t get caught because they’re slow. They get caught because everything they’ve built is tuned to how things work today, and keeping up means tearing a lot of it up.

Small teams don’t have that problem. With a connected operation and no old way of doing things to protect, they can move faster than a bigger company can follow. That’s where I expect the real disruption.

Sources

← All writing