Axios C-Suite: AI's messy middle
Add Axios as your preferred source to
see more of our stories on Google.
I've experienced it myself and heard it from dozens of you: We're stuck in AI's messiest middle.
Why it matters: Yes, AI is more powerful than anyone thought possible two years ago. Yes, it does SOME things, like coding, magically well. Yes, it shows PROMISE to soon train and improve the next generation of itself.
- But right now, it's more expensive, faster-changing, harder to connect and harder to justify than hoped. Put simply, it's messy and costly.
The collective mood of dozens of CEOs I've talked to is hopeful but frustrated. So let me distill the frustrations — and the smartest ways through this weird period.
The five frustrations:
- It's too expensive. AI is great for repeatable, verifiable problems with rich data history. Most everything else costs too much for too little right now. Chamath Palihapitiya said his company's token costs are doubling every 45 days, while productivity is up just 5%.
- It's a competitive threat. Using AI means letting your vendor in on exactly how your business works. Microsoft CEO Satya Nadella admits buyers "essentially pay for intelligence twice" — once with money, then with "the proprietary knowledge you must reveal to make that intelligence useful."
- It's changing too fast. Anthropic, OpenAI, Meta and xAI constantly spit out new models and applications. Now, China's open-source models look promising, too. Every release comes with new capabilities, glitches and costs. This makes tradeoff decisions about integration temporary, complicated and fluid.
- It's creating new problems. Several CEOs told me Anthropic and OpenAI built and hyped new cyber capabilities, then forced them to spend unbudgeted money and time fighting those same threats.
- It's too clunky when connected to existing systems. Connecting AI to internal systems with reliable security and definite outcomes is still too hard.
I assume all of this gets solved once we sort through the mess, as it has before. In the meantime, five common-sense commandments to navigate this moment:
- Know thy models. We need smart people tracking model capabilities and costs, open and closed. We have no choice but to toggle between models. Nobody knows which will become the standard. And standards will likely get established by avid users.
- Know thy knowledge. Your alpha knowledge is the proprietary information, data and differentiators that make you you. Protect it and use AI to expand that moat and money-making muscle.
- Know thy priorities. Build a list of automation-ready tasks, ranked by value created. Run it against today's costs and capabilities, and assume both normalize. In past technological shifts, most lasting gains revealed themselves during the mess.
- Prepare thy team. Draft the organization plan for when AI can do what you need at a price you can defend. The companies that get hurt in every shift are the ones that treat the mess as a reason to wait.
- Protect thy experiments. Build a safe space outside the reach of frontier models to test and fine-tune internal models. Don't let costs, or fear, slow experimentation.
Your messy middle checklist:
- List your automatable tasks or products that would drive the most value (revenue or savings) over time.
- Get your CTO to figure out which can be done now at human-level or better efficacy using existing AI.
- Determine the rough cost of each one based on current pricing.
- Focus only on those things that can be done with reasonable ROI now. Put the others on a to-do list as capabilities rise and costs fall.
- Rinse and repeat regularly.
