The premise
The pressure is real, the groundwork usually isn't
There is enormous pressure right now to show that your company is using AI. It comes up on earnings calls, in board meetings, and from the owners and shareholders who want to know how you are leveraging it to improve operations and products. "Not yet" is a hard answer to give, and that pressure is real.
So the natural move is to act fast: buy the tool, hand it to your people, and wait for the numbers to move. Often they don't move much, and it is tempting to conclude the team just didn't adopt hard enough. More often, the groundwork for AI to succeed simply wasn't in place yet.
That gap is not a failure of will. What gets rolled out is a chatbot on every desk, an exosuit for the individual worker, the kind Ripley straps into to fight the alien at the end of Aliens. It makes one person much stronger, and that is real, but it is still bolted to a single human doing the same job, and lean on it too hard and the underlying muscle starts to atrophy. A workforce in exosuits is a genuine boost. It isn't transformation. Real transformation is a fundamental shift in how decisions get made, and decisions are what businesses run on.
Decision economics
Every process is a stack of decisions
Strip any business process down and you find a sequence of decisions. Some are obvious. Most are buried in email threads, spreadsheets, work queues, dashboards, and the tribal knowledge of whoever has been there the longest.
Before AI, you had three ways to make them:
- Deterministic rules: if this, then that.
- Traditional analytics and machine learning: forecasting next month's demand.
- Human judgment: approving an unusual refund.
The real question
AI adds options between the rigid rules and the expensive humans. So the question stops being "where can we use AI?" and becomes "what is the cheapest, fastest, safest way to make this decision at the quality we need?" That is optimizing decision economics.
Routing
The frontier model isn't the answer to everything
A frontier model from Anthropic or OpenAI is good at almost everything, so the easy move is to route everything to the best one available. Most people do, and it is a reasonable instinct. They don't know exactly where the quality line falls between the top model and a cheaper one, and figuring that out takes time they don't have. Defaulting to the smartest model avoids the risk of rework, so that is what everyone reaches for.
The trouble is that this is really a question of token economics, and sending everything to a frontier model is how you run up an exorbitant bill. Match each decision to the cheapest tool that clears the quality bar:
- Known, stable patterns: deterministic logic. Cheap, fast, testable, auditable.
- Data and a narrow target: traditional machine learning is often enough.
- Broad reasoning, ambiguity, synthesis: where frontier inference earns its cost.
- Accountability, ethics, trust, strategy: a human stays on top.
The goal isn't maximum AI
Good AI design is the right decision structure for the work, not maximum AI usage. AI augments around its strengths; it doesn't replace everything in sight.
The long tail
The interesting work is the long tail
For years, engineering automated the high-volume happy path and routed everything strange to a team of humans. The long tail of exceptions was too expensive to analyze, code, test, and maintain by hand. Diminishing returns.
That math just changed. AI can read years of historical exceptions, surface the recurring patterns, draft candidate rules, write the tests, and document the logic, then push deterministic coverage deeper into the tail than was ever worth doing by hand.
The catch: someone still has to verify the rules the AI drafts before they go live. The work moved from writing logic to checking it. That's a better trade, but it's still a trade.
Hybrid orchestration (deterministic systems and AI working side by side) is how you optimize cost, speed, safety, and compliance at the same time, instead of trading one for another.
The part leaders skip
You're redesigning decisions, not just buying licenses
Redesigning decision architecture means redesigning work, and work is done by people with incentives.
The companies failing at AI usually aren't failing on the technology. They're failing because nobody designed the work, and the people asked to adopt the tools can see that "successful" adoption might cost them headcount or the craft they're proud of. Hand someone a tool that threatens their job, put them in charge of its rollout, and you shouldn't be surprised when it stalls.
A decision redesign that ignores the humans inside it is just a slower failure. Get the incentives honest first, or the architecture never ships.
The gating function
AI is only as good as it's let in
A new hire spends weeks being shown how the work really happens: the unwritten rules, the edge cases, who to ask when something looks wrong. AI gets none of that by default. It only knows what someone takes the time to write down. Onboard it well and it can perform like a strong hire. Leave it to guess and it stays a clever stranger.
That puts the people already in the role in a hard spot. They hold the operating knowledge the AI needs, and they're the ones being asked to hand it over. It's human nature to protect what you value, and plenty of people value the expertise that makes them hard to replace. When documenting how you work starts to feel like writing the manual for your own replacement, hesitation isn't sabotage. It's self-preservation.
Companies make this worse when they announce layoffs and pin them on AI. Once "we cut staff thanks to AI" is the public story, asking those same teams to document and train the tools is asking them to train their replacement. That's not paranoia. It's reading the room. If you want people to let AI in, you have to make being good with it safer than staying quiet about it.
Timing
Timing is part of the design
Here's the part nobody likes: you're redesigning around a moving target.
"This is the worst AI is ever going to get" sounds like strategy. It's really an assumption about the shape of the curve. Combustion engines improved for a century, but most of the last few decades were marginal. The next real jump came from pairing them with electric systems, not from the engine becoming a different machine.
Changing systems, processes, and people is expensive. That's the change tax. Pay it too early and the premise shifts before you capture the value. Pay it too late and a competitor banks the learning curve first.
Why your timing isn't their timing
Your moat sets your clock
The right timing isn't the same for every company. It depends on what you own.
A telecom with fiber in the ground can watch the curve longer. Nobody prompts their way into a national network, so the moat holds while the models improve. A company whose moat is mostly software workflow has less room. If AI keeps lowering the cost for customers to build the narrow internal tools they used to pay for, the rent gets harder to justify.
So the change-tax math is personal. Read your own moat honestly before you decide whether you're early or late.
The bottom line
It comes down to three questions
Pull it all together. The pressure to "do AI" is real, but handing everyone a chatbot is augmentation, not transformation. The real work is redesigning how your organization decides, then re-engineering the process around where humans add the most value. And the timing is yours to judge, set by your own moat, not by the hype cycle.
Strip away the noise and it comes down to three honest questions:
- What is the cheapest, fastest, safest way to make each decision well?
- Where do humans add the most value, and how do we re-engineer the work around that?
- Are we building around durable capabilities and our own moat, or around this quarter's model limits?
The takeaway
Answer those honestly and the tools mostly pick themselves. Skip them, and no number of seat licenses will save you.