Stupid.ai - Quick! Do some AI before everyone else does.

SHHHH. Do you hear that? It’s every company you can think of racing to start doing … stuff … with AI. They stood up an AI innovation group or center of excellence or something. They hired a bunch of people (probably contractors or consultants) to help them move faster. The launched a bunch of pilot projects with small teams trying to move quickly focusing on the highest impact projects.

Now here’s my magic trick for today: I bet I can guess what your company’s main AI projects are. It’s probably one (or all) of the following:

• A bespoke coding agent that is “Copilot or Claude Code for company X”
• A chatbot that is “Company X - GPT” for internal use
• An agent that automates a boring business process (like creating Jira stories or summarizing things for a monthly business review)

Ta-da! Nailed it, right? Your company is doing one or all of those things and possibly has even cancelled one of them already. But according to an MIT study, although there has been $30–40 billion in enterprise investment into GenAI, 95% of organizations are getting zero return. Not just a low return. ZERO.

So let’s talk about what everyone is getting wrong here and what they should be doing differently.

Mistake 1 — A tool is not a strategy

As I’ve written about previously, most big companies don’t have a strategy. And that problem carries forward into the usage of LLMs. And the key word here is tool. An LLM is a tool. It’s a powerful tool, but it’s still just a tool. It’s not magic, and it won’t solve your problems by itself.

Here’s one way you can check if you’re making this mistake. Whenever someone talks about “AI” as a strategy, just replace AI with the name of a different tool. For example:

AI is a core part of our strategy going forward. We are investing heavily in AI and have a dedicated team working on it.

becomes

Excel is a core part of our strategy going forward. We are investing heavily in Excel and have a dedicated team working on it.

or

Laptops are a core part of our strategy going forward. We are investing heavily in laptops and have a dedicated team working on them.

Sounds stupid right? That’s because it is. No company has every had a successful strategy like that.

“Now, Matt” you might be saying, “what about smart phones or the internet? Those were tools too and lots of companies had successful web-first or mobile-first strategies.” That’s true, but there’s a key difference here (at least in terms of how companies are currently thinking about AI).

Web and mobile platforms fundamentally changed how customers interacted with businesses. Being a “web-first” or “mobile-first” company didn’t mean “we’re going to use the web/mobile inside the company to make things faster.” It meant “customers and the world at large are changing their behavior. We’re going to structure our whole business around that new behavior.” Building an internal coding agent changes nothing for your customers other than your bad ideas get in front of them faster.

Now I suppose that someone could actually have an agent-first strategy. This would assume that the primary means that a customer uses to interact with your product is through an LLM in some way. I’m a bit skeptical of this for a variety of reasons (will probably write a whole other post on this). I also haven’t seen any companies do this well yet and it’s certainly not the norm.

Mistake 2 — Trying to be someone you’re not

This mistake is again, not unique to AI, but it’s especially common here. Every company seems to be trying to build a coding agent or chatbot or an AI plugin for a tool they already use. And oftentimes, they may get some traction but one week after launch, Anthropic releases the same feature under your existing license but it’s 10 times better.

So here’s what I would ask these companies:

Are you Anthropic? Or Google? Or OpenAI? No? Then why in the f**k are you trying to compete with them?

An insurance company will never make a better coding agent than OpenAI. A bank will never make a better chatbot than Anthropic. A retailer will never make a better AI plugin for Shopify than Shopify. So why are you trying to do those things? You didn’t build your own Excel or Salesforce or Slack. There is plenty of stuff that you probably ARE good at that Anthropic is not. Focus on what’s unique to you, ignore the rest.

Mistake 3 — The pilot graveyard

This is the most common mistake of all. Every company has a pilot graveyard. They have hundreds of projects they started but never finished. They launched a chatbot for internal use but it’s easier just to ask ChatGPT. They built a custom agent to automate some process but it was easier just to do it manually. They built an AI plugin for their CRM but the off-the-shelf one is good enough. Fundamentally, this is a lack of true product management. They weren’t solving the right problem, or they weren’t solving it in the right way, or they have no way of measuring success.

Here’s the reality: for any new technology (and even old ones), most of the projects you start will fail. That’s just how the world works. We are usually wrong. We have bad ideas. We made a wrong assumption. As Rafiki would say, “You can either run from it, or learn from it.” The risk here is not that your pilots fail. It’s that your failures are expensive and slow. You need to assume that most projects will fail. So, you should do a LOT of them as quickly as possible and not bet the farm on any one of them.

One way I’ve seen this go wrong is when there is a lengthy gatekeeping process for a new idea. You have to submit an intake. That intake gets reviewed by 3 layers of management. Your project gets weighed against the 50 other submissions (only 10 of which will be approved). Then 6 months later, you finally get access to the model API’s you need. By that time, one of two things has happened: either you realize that your idea won’t work OR your idea is now obsolete (especially common now with the pace of LLM model development). In either case, you just wasted 6 months and probably millions of dollars when you account for all the people involved.

Instead of spending time on processes and governance, spend time on platforms and tooling. Make it easy for anyone to get access to the tools (in a controlled way) and test out ideas. Make it easy to measure the impact of those ideas. Make it easy to kill the bad ones and double down on the good ones. The faster you can do that, the more likely you are to find a few winners and the less money you will waste on losers.

So what do?

The common thread across all three of these mistakes is the same: companies are reacting instead of thinking. They see competitors launching AI initiatives, they read the press releases, they watch the demos and they panic.

The companies that will actually win with AI are not the ones that move first. They’re the ones that ask the right questions first. What do we know that nobody else does? What data do we have that nobody else has? What problems do our customers have that AI could genuinely help solve in a way that is unique to us?

That’s your strategy. The LLM is just the tool you use to execute it.