Where to Spend
The Setup
Tokenmaxxing. This buzzword, which I first thought was a YouTuber’s handle name, has arrived unknowingly to my sphere of work.
It’s such an appropriate word that represents Goodhart’s Law.
An AI coding tool got switched off at work for me. Not a planned sunset. Just a “careful re-evaluation of all the AI tools we have.” I sensed the firm and teams were burning through its usage faster than anyone expected. Token budgets that were supposed to last a year were gone in four months. I was mid-use for an RFP. My first instinct was to make the (stern) case to get it back.
Then I evaluated what it had been doing with all that usage.
Every run, it was re-deriving the same things from scratch. Basic structural knowledge like how to parse a specific kind of internal report, what a recurring piece of shorthand meant represented by a spreadsheet, the stuff anyone on the team would learn in their first few days during onboarding.
None of that had ever been written down anywhere the tool could find it. So AI was solving the same problem over and over, at full token cost, because as far as it knew, nobody had ever solved it before.
My long-term business case was formed: Invest in codifying what we already know for those who can detect and form that knowledge for others, and give me the tool to do so.
My Approach + AI Role
I’d already been toggling my own AI usage by task type for months before this Tokenmaxxing thing became a buzzword. The most capable tools I keep for massively vague work, the kind where I don’t yet know what I’m looking at and need something that can hold a lot of context while I figure it out. Once that foundation is built, lighter and cheaper tools can handle the work on top of that. Drafting inside a template, running a checklist, producing something I can evaluate quickly against a clear standard… etc.
I hadn’t put language to why that split made sense until I started mapping it against a framework I’d been developing for how knowledge workers process information:
- orienting to a situation
- retrieving what’s known
- making sense of what’s ambiguous
- deciding
- producing
- aligning with others
- verifying the result holds up
Most of those moments have clear definitions and boundaries. But there’s one that doesn’t: sensemaking. That’s where you’re still making meaning out of what the problem even is with all the clues you’ve identified. Sensemaking is the moment that needs time and real AI capability, not the cheapest tool that will technically respond.
Not time for the AI to “tell you the answers” (yes, it will always confidently do that… you should always pressure test it). The AI tool helped explore patterns and ways of looking at a problem, it reminded me of ideas I’d forgotten, it helped me with more ways to make meaning.
Then I’d sleep on it and see what stuck out to me the next day.
What Actually Happened
Earlier this year I inherited an onboarding problem for a major account. What existed was a collection of old decks, fragmented strategy documents, and tribal knowledge that lived in the heads of people who’d been on the account for years. A new person joining the team had to piece it all together themselves, or find the right person to ask, or simply sit and wait as teammates struggle with out of date materials and the overwhelming feeling of “how do I explain this”.
I spent weeks reading the old strategy decks on and in-between meetings. Talking to team members. Mapping how the account actually worked versus how the decks said it worked. Figuring out which operating principles were real and which ones had been true last year. That whole process was super ambiguous… I didn’t know the shape of the answer going in. I was deciding what mattered, what to leave out, and how to connect pieces that had never been connected before.
What came out of it was a knowledge layer for the account. An interactive orientation walkthrough. Operating protocols. Principle-based learnings so team members could remember previous judgments instead of re-discovering them. Audience & Messaging files so the AI knew what mattered not just to our team, but to the people we were trying to reach.
That work was expensive (in tokens and in time). And it should have been. Because once it existed, the character of the work changed for everyone else on the account. The next person wasn’t re-figuring out what mattered. They were following what had already been figured out. Retrieving the right reference, producing inside a known template, checking against a written standard. That work didn’t need the same tool I’d used to build the foundation.
I’ve nailed down a formula… I thought.
Then I was building a different orientation piece a few months later, the artifact kept failing. I kept thinking I didn’t understand the material well enough. More research, more synthesis, more tweaking. Was it the prompt? Was it the logic? Then I realized what didn’t work weren’t things I’d misunderstood. They were decisions that hadn’t been made yet. I was trying to orient people to conclusions that didn’t exist. The artifact’s failure was telling me which moment I was actually in… still sensemaking, not ready to produce.
The AI tool that got shut down on me was the same problem at the org level. Nobody had done the sensemaking and architecture work to determine which and what type of work gets what resources. So every AI interaction ran at full cost, not because the questions were hard, but because the system had no memory.
The Real Insight
I assumed the usage budget guidance was the problem. It wasn’t. The missing groundwork for shared knowledge was.
When something a team already knows only lives in people’s heads, or gets re-figured-out by a tool every time someone asks, every interaction runs at the highest possible cost. Not because the underlying question is hard. Because nothing was ever captured. The fix isn’t spending less. It’s writing down what’s already known, once, and sharing it widely, so the expensive thinking only gets spent on things that are actually still open.
On the flip side, some constraint is required. I hit my own weekly usage limit while running a complex audit… triangulating contract documents, burn reporting, and interview transcripts to find a pattern I couldn’t see yet. When I saw that limit message, I wondered if the effort was worth it. (It was. It generated targeted knowledge files that sharpened pricing estimation and surfaced deliverable approach reminders that had been learned on the account before but only held by few experienced team members.)
A hard AI budget ceiling forces questions that unlimited access won’t reveal. And we need to get more imaginative with our answers to break our old habits.
Try This If…
If you or your team just got handed a usage limit on an AI tool, or you’re bracing for one:
- Ask where the actual re-work is happening. Where is a tool solving the same problem from scratch every time across hundreds of employees, because nobody ever wrote the answer down?
- Separate what’s truly still open (you don’t know the shape of the answer yet) from what already has a known shape (you’re just executing against something established). Spend your best capability on the first kind.
- If you can’t tell which is which for your own work, that’s the actual finding. It means the knowledge layer groundwork hasn’t been done yet, and that’s the first investment to focus on. Not more AI tools.
Systems Lens
The teams that come out ahead under tighter AI budgets won’t be the ones that simply use less. They’ll be the ones who can say, specifically, what’s worth using it on and what isn’t.
I’ve thought for a while that the old habit of tracking hours worked was going to fade out as AI took on more of the actual doing. What I’m watching now is that the habit isn’t disappearing. It’s just measuring something else. Instead of hours, it’s usage. Instead of a person’s time, it’s a machine’s output, tied back to a person’s judgment about how to spend it. The underlying question hasn’t moved at all. A limit by itself, with zero idea what it’s for, doesn’t create use discipline. Just a lot more confusion and mixed signals received by those in the trenches told to use AI for efficiency while headcount cuts are happening left and right.
What I’m still trying to figure out: If the result of all this efficiency always shows up as less spent rather than as more capacity handed back to people to do deeper work, then what was the actual point of all the change management around this? A token budget makes that question impossible to ignore. Either the savings get reinvested in the harder, better version of the work, or they just get pocketed and the vicious cycle of efficiency targets and reduced headcounts starts all over again.