AI and Business Strategy
Adam Shostack, Shostack + Associates
How can we shape a business strategy in an ‘AI world?’
LLMs are an incredibly disruptive technology. Recently I talked to someone who used one to compress a project they estimate would have taken six months into six weeks. Every business we talk to is trying to figure out how that disruption impacts them. And while the conversations we have often center on software security and engineering, today I want to look at the situation more broadly.
Strategy, Richard Rumelt tells us in Good Strategy/Bad Strategy, is about understanding your strengths and weaknesses relative to competitors, and choosing actions to either magnify your strengths, or shore up your weaknesses. For example, Shostack + Associates is very focused on building security in via threat modeling, and that strength is built on decades of experience and is demonstrated by widely respected books and other resources that inform the whole field. I can talk about that openly because it's hard for anyone else to duplicate. Rumelt specifically talks about how bad strategy includes ideas like “work harder,” because anyone can try it, and no one has an advantage in doing so.
OpenAI and Anthropic are the only companies for whom “more tokens” approaches strategy.
”Spend more tokens” has almost the same problem. The “almost” is important here. It’s trivial for Microsoft to spend more tokens than Shostack + Associates. They have more revenue, they have more cash on hand, and they can borrow more than we can. Some companies are going to use LLMs to bury the competition that way, and maybe it will even work. But our strategy recognizes that cash on hand is not our business advantage. Unlike Microsoft’s services arm, we offer specialized advice based on focused experience, without having a product quota.
Economists have a term, “red queen race,” in which participants have to run faster and faster, just to stay in place. And if you’re selling sneakers, convincing people there’s a red queen race going on may help you sell a lot more sneakers. And if you're selling tokens, convincing people that spending tokens is the future will help you sell more tokens.
OpenAI and Anthropic are the only companies for whom “more tokens” approaches strategy. If they can convince their customers that they need to spend more tokens to stay relevant, they win. For everyone else, the same results with fewer tokens are better. And so maybe “getting better at spending tokens efficiently” is a strategy? That’s along the lines of “work harder.” Anyone can do it. It’s a fine tactic, it’s helpful in the short run, but the reason our AI-generated blog images usually include the prompt we use is that our strategy involves teaching, and sharing our prompts aligns with that. (NVidia might be in that category.)
If you’re going to pay someone to help you do something, you expect them to be accountable for the results.
The same applies to AI-driven engineering. Many people are sharing stories about how they’re getting exciting local acceleration (creating demos is now free, as are simple features). There are also stories about AI deleting prod, making up statistics, seeing painful drops in quality, re-hiring staff and more. We’re seeing stories about cognitive surrender, AI exhaustion from trying to keep up with AI. In the April roundup, I linked to Buying Back our Slack, and how AI is reducing our ability to reflect and make sense of the work we’re doing.
Which almost brings us back to strategy. But before we get there, let’s talk about what people pay for. People pay to have problems solved. No one wants a drill. No one even wants a hole in their wall. They need a drill to hang appropriate art. And today, if they’re going to pay someone to help, they want that person to solve a problem for them. And when someone get paid to help with a problem the buyer expects them to be accountable for the results. That’s why KPMG delivering a report with only 5 real citations out of 45 is so shocking. People pay KPMG for thoughtful advice; anyone can ask a chatbot for free, and they know what they’re getting.
When people pay to have their problem solved, they look for the lowest cost provider who can meet their needs, and that brings us back to spending lots of money, in two ways. First Anthropic, OpenAI, Google, and Facebook are all spending a lot of money to create “frontier” models that are not going to be very differentiated. This seems like the airline business. As Warren Buffet pointed out, competition is extremely expensive, but that doesn't make them good investments.
Sure, today Claude is better at coding, but Gemini is better at coding than the version of Claude that was available 6 months ago. And so we’re going to see price competition. In fact, we’ve already started seeing competition on price, and that means that whatever your strategy is, it probably doesn’t require most of your spending to be on the very latest models. You can probably get most of what you need at a huge discount, and learning to do that keeps costs low. (In fact, if you look at a leaderboard like Arena.ai's webdev, you'll see Claude Fable tokens costing $50 per output million to get a score of 1630, while Kimi-k3 costs 1/3 of that ($15) for a score of 1682, and GLM-5.2, with a score of 1588, costing $4.40 for the same million tokens. Are you confident that you understand those scores, and what you get for your extra 11x dollars?) Maybe we’ll get Frequent Prompter memberships that let us use LLMs for personal use...
Which, finally, brings us back to strategy: What are your business strengths, and how do you use LLMs to accelerate delivery, improve quality, or lower costs while executing on that strategy? What are your weaknesses, and can LLMs help you address them?
Image by midjourney: "A warm, slightly cluttered office space where two robots sit across from each other at a wooden desk covered in papers, books, and coffee cups, deep in a collegial conversation. One robot gestures thoughtfully at a whiteboard covered in strategy diagrams and token cost calculations. The other leans forward, engaged. Bookshelves visible in the background. Watercolor style, warm amber tones, the mood of two colleagues working through a hard problem together. Watercolor style impressionist colorism, clean lines."