A tool is not a strategy. It’s a part. Picking a part before you know the system you’re building is how smart people set money on fire.

The most common question I get from founders is some version of “which AI tool should I use.” It’s the wrong first question.

I know, because I asked it myself and it cost me. Early this year I was running six AI tools at once. Claude, ChatGPT, Gemini, TypingMind, Genspark, Manus. I was paying for all of them, and every one held a different, half-outdated version of my own business. I started every session re-explaining who I was and what I did, in every tool, before I could get any real work out of it. It felt like the tools had a memory problem. They didn’t. I had a system problem.

That was my fire. The market is running a much bigger version of the same one. Let me show you that first, then the fix.

Why do companies burn their AI budget with nothing to show for it?

Because they bought tools and called it a strategy, so the spend has nothing to attach to. More budget on a missing plan just means the fire burns hotter. If the tool were the strategy, the biggest spenders would be winning right now. They aren’t.

Look at who’s spending. Uber blew through its entire 2026 AI budget in four months, and its own COO is now questioning the return. They’ve since capped it around $1,500 per employee per month, which tells you the reaction to “we’re not seeing it” was to throttle the spend, not to fix the plan underneath it. That’s the tell. When money moves faster than results, the problem was never the amount of money.

It’s not just Uber. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, and the cause they name isn’t the technology falling short. It’s unclear value and weak governance. The projects get killed because nobody could say what they were for.

And the sharpest number of all: MIT NANDA found that only 5% of enterprise generative-AI pilots drive measurable revenue. The other 95% show no measurable impact on the P&L. Ninety-five percent. You cannot look at that and conclude the gap is budget. The gap is strategy. Everyone bought the same tools and almost nobody built a system for them to plug into.

Three different sources, same gap: the money moved and the strategy didn’t.

Is choosing an AI tool a strategy?

No. Choosing a tool is a purchase. A strategy is the system the tool plugs into, and the system is the part almost nobody builds. The tool is a component. The strategy is the decision about which job it does, where it fits, and where a person stays in the loop.

Go back to my six tools. The context lived inside each app, so it was only ever as current as the last time I happened to feed that specific app. There was no single place that was just true. Adding a seventh would not have fixed that. It would have given me one more copy of the same confusion. And the tools were half-built underneath all of it, hallucinating, losing the thread of a long document, contradicting themselves, which is exactly why you never build the company on any one of them.

Sidebar: What’s the difference between an AI tool and an AI system? A tool is a single thing you use: a model, an app, a subscription. It’s rented, it’s replaceable, and it’s someone else’s product. A system is the thing you own that the tools plug into: your knowledge in files you control, the rule for who checks the work, the decision about where a human stays. Tools change every few months. A good system barely moves.

What are the most common ways companies misuse AI?

Three broken relationships cover almost all of it: they treat the tool like a genie, like a hammer, or like a line item in a big-company spreadsheet. All three come from the same mistake: believing the tool is the strategy. If the tool is never the strategy, none of these can work.

The genie. You make a wish and walk away. A founder tells the tool to “handle our support inbox,” expects it back perfect, and gets a pile of confident, wrong replies to customers. Genies in the old stories were a warning, not a gift. They give you something other than what you pictured, because you delegated an outcome and never stayed in the process.

The hammer. When the only tool you swing is a hammer, every problem looks like a nail. This is the founder who buys ChatGPT seats for the whole company and calls it an AI strategy. Same tool, applied the same rote way to every job, whether or not it fits. Nothing changes, because a seat is not a decision about which job needs which tool.

The big-company line item. Microsoft reportedly pulled Claude Code licenses from a major division and pushed teams onto internal tooling instead. Set aside whether that was right for them. The point is that a tool decision that size gets made on vendor alignment and cost, which are the wrong KPIs for “does my team ship better with this in their hands.” Giant orgs make that mistake because they’re too heavy to fit the tool to the job. You aren’t.

Here’s the part the tool vendors don’t want in the story: that half-built flakiness I hit with my six isn’t a quirk of one bad tool, it’s the current ceiling of the whole category. Every model in the field carries the same failure modes, and no amount of loyalty to one brand buys you out of them. That’s not a reason to sit out. It’s the reason you never bet the whole company on any single one of them.

Should I build my AI strategy around one tool or many?

Neither. Build it around the two things you actually own: your knowledge in plain files any model can read, and the human judgment loops around the work. The tools stay rented and swappable on top of that. That’s the whole move.

Once I saw the six-tool mess for what it was, I built the thing that was missing. It’s almost embarrassingly plain:

  • One private repo as the single source of truth.
  • Plain markdown, split by domain: services, marketing, SOPs, brand, me.
  • Any tool loads only the part it needs.

On top of that I route across 70-plus models from different providers, through OpenRouter, matching the model to the task instead of marrying any one company. I don’t collect models. I route to them.

Now, when a founder hears “own your system,” they picture something heavy and expensive. A platform. A gateway. Custom middleware to route seventy models. That’s the trap, and it’s the opposite of what I’m describing. Owning your system does not mean building an abstraction layer to insulate yourself from every vendor. It means two cheap, durable things: your data in a format any model can read, and your judgment kept in the loop at the points that matter. Files and a rule. That’s it. The expensive version is slow and it breaks. The cheap version is what actually survives.

The mechanics of how I run that (the portable memory, the multi-model review, the files instead of apps) are their own piece. If this article is the why, that one is the how: I Had a Frontier Model for a Week. I Used It to Audit My Own Systems. I’m not going to re-run it here.

What happens to my AI work if the tool shuts down?

If your work lives inside a tool, it leaves when the tool does. If it lives in files you own, with the discipline built around them instead of inside them, a shutdown costs you an afternoon instead of your business. I got to test this on my own stack last month.

On June 18, Google deprecated the command-line tool I ran every code review through. My system barely noticed.

Here’s why it didn’t. I run every serious build through a review from more than one model family, so no model grades its own homework, and for a long time the second, non-Claude reviewer ran through that Gemini tool. But the reviewer was never wired to that specific binary. It was wired to a thin adapter with a clean contract: give me an independent, non-Claude opinion on this. When the tool disappeared, moving to its successor, Antigravity, was re-pointing one small script. The review discipline never changed. My work was sitting in the same files it had always been in. What would have been a migration project for someone who’d built their process inside that tool was, for me, a slow afternoon. That’s the entire payoff of not marrying the tool.

I’ve watched a business owner lose an audience of half a million people overnight because it all lived on a platform he didn’t control, with no straight answer and no appeal, and he started over from zero somewhere else.

A deprecated dev tool and a deplatformed account are the same lesson at two different sizes. Rented ground can be pulled, and it’s never pulled at a convenient time.

Won’t I fall behind if I don’t jump on every new model?

No, not if you own the durable layer. Being early on a new model is a real edge, but it’s short-lived, because your competitor gets that same model the day it ships. What lets you win the race anyway is owning your files and your judgment, so you can plug in whatever model wins this month the day it lands and keep the gains after the edge is gone.

Be fair to the pull, because it’s real. When a sharper model ships, the people who adopt it first do get ahead. But that edge is short, because the same model is in your competitor’s hands the day it ships. So chasing tools is a treadmill. You get ahead for a moment, then everyone’s even, then the next model ships and you sprint again.

Owning your durable layer is what lets you win the treadmill anyway. Because your knowledge is in portable files and your review is a discipline, not a dependency, you can plug in whatever capability wins this month the day it lands, and drop it the day it stops being best. You are not refusing the cutting edge. You get it faster than the founder who has to rebuild around each new tool, and you keep the gains after the edge expires.

So here’s the gut check. If the tool you lean on hardest vanished tomorrow, what’s still standing? Whatever’s left is your actual business, and everything else was rented. If the honest answer is “nothing,” that’s not a failure, it’s the most useful thing you’ll learn all quarter, because now you know exactly what to build first.

Route, don’t marry. If your edge is just the tools, you don’t have one, because your competitor is buying the same ChatGPT and the same Claude you are. The edge is the system you put around them and the judgment about where your people stay. That’s the part nobody can buy off the shelf, because you built it.

If you’re staring at your own stack right now and you’re not sure what would be left standing, let’s start a conversation.