Let Them Spend

The most tempting decision a leader can make about AI tools right now is also one of the most expensive ones. Give everyone the same modest budget. Cap it. Call it fair and responsible.

It is neither. And most organizations are about to make some version of this mistake.

Skill with these tools comes from one place only: using them, badly, until you are no longer bad. There is no course that replaces it. The developer who burns through a pile of tokens this month fumbling with an agent, getting garbage, adjusting, getting slightly less garbage, is the person who returns tenfold value next quarter.

On the balance sheet, this month looks like waste. It is tuition.

So when you cap everyone at a level designed to prevent waste, you are also capping the learning. The people who are already skilled stay skilled. The people who never got room to fumble stay exactly where they are. You have made the skill gap permanent and called it cost control.

The opposite decision is just as bad, and I have seen the enthusiasm version of it too. No caps, full access for everyone, trust the team. It sounds generous. Then someone leaves an agent looping over a weekend, the monthly bill arrives looking like a phone number, finance escalates, and leadership does the only thing leadership can do at that point: pulls access for everyone. Irresponsible freedom is not generous. It is how the whole organization loses the tool.

So the real question is not how much people should be allowed to spend. It is how you build so that spending freely is safe. My answer is three guardrails, and they do three different jobs. Most of the damage I see comes from confusing them.

The first guardrail is about competence. Not everyone gets to run every machine on an industrial floor, and the reason is not status. It is consequence. A breakdown is expensive, so authorization follows the ability to handle the machine. The same logic applies here, with one important detail people get wrong. A license is a threshold, not a ration. Your driving license lets you drive as much as you want. It does not grant you kilometers in proportion to your test score. So the model is not "the skilled get more tokens". The model is: everyone can spend freely, on one condition. That you know what you are doing.

And in a field moving this fast, knowing what you are doing has a short shelf life. What you learned in spring is partly wrong by autumn. So this cannot be a course you pass once and frame on the wall. It is a mandate you maintain. Stay current, or the threshold is back. That sounds demanding, and it is, but notice what it is demanding about. Not that you are the best. That you have not stopped learning.

So what do you actually measure? It is easy to build something here that backfires. Measure performance, tokens per delivered feature, value per krona, and you have built an anxiety machine. Nobody experiments in a system that grades their experiments. People will only use the tool for what they already know works, which is precisely the opposite of learning. Measure participation instead. Are you keeping up, are you trying things, are you sharing what failed. That is soft, a cynic will say, and unmeasurable. It is softer, yes. It is also the only version where the freedom produces the skill you are paying for.

The second guardrail has nothing to do with competence, and this is the one that must never move. There is an absolute ceiling, and it applies to everyone. The most skilled operator on the floor still cannot run the machine until it breaks, and still cannot burn the month's material budget in one shift. Not because anyone doubts their skill. Because no single person, however good, and no single agent, however useful, gets to consume what has to last for everyone. This ceiling does not rise as people get better. It is not about them. It protects the whole, and the moment you let excellence negotiate with it, it protects nothing.

The third guardrail is the one most discussions miss, because it does not restrict the human at all. It sits in the architecture. Not every question needs the most capable model. Most questions, in fact, need a fraction of it, and the price difference between the frontier models and the small ones is not a discount. It is orders of magnitude. So you build routing: the simple question goes to the cheap model, the hard one to the expensive one, decided automatically, invisible to the person asking. Done well, the bulk of your consumption quietly flows through models that cost almost nothing, and the expensive capacity is reserved for the work that actually needs it.

Notice what this does to the other two guardrails. It lowers the cost of every action, which means people can do far more before the ceiling is ever in sight. You did not raise the cap. You moved the day anyone hits it.

The first guardrail frees people through trust. The third frees them through technology. Only the second holds back, and it does that precisely so the other two can afford to be generous.

There are two good objections.

A CFO will say: an open sandbox times a few hundred consultants is an unpredictable bill, and unpredictable bills are what I am paid to prevent. Fair. But look at what the construction actually does. The freedom is open. The cost is not. The ceiling bounds the worst case in absolute terms, and the routing pushes the expected case down to a fraction of the naive one. That is not an appeal for trust. It is an architecture with a worst case you can read off in advance, which is more than most line items can say.

An architect will say: routing is not free. The classifier costs something, it is a new failure point, and sometimes it will send a question to a model that is too weak, and the answer gets worse without anyone noticing. This objection is better than it sounds, and the last part is the serious one. Silent degradation. The cheap model that was good enough in March and quietly is not in June. Routing without continuous checking of whether good enough still holds is just optimizing cost blindly. So the third guardrail is not a set-and-forget component. It is a commitment to keep measuring. If that discipline is not something your organization can sustain, then the honest advice is to run one good model, pay more, and skip the complexity. The economic guardrail is earned, like the others.

The position sounds almost contradictory. I want people to spend without fear, and I want the bill under control, and I refuse to choose between those. Cultivating experimentation and containing cost are not opposites you trade against each other. They are three guardrails doing three different jobs. One moves with the person. One never moves at all. One works so quietly that most people will never know it is there.

The art is in never confusing them. Loosen the ceiling because someone is brilliant, and you have gambled the whole. Keep the competence threshold high out of habit, and you have strangled the learning you claimed to want. Skip the routing, and you are paying frontier prices for questions a small model answers fine.

Get all three right, and something changes in how the organization behaves. People stop treating the tools like a scarce resource to be hoarded and start treating them like infrastructure. That shift, more than any individual productivity number, is what you are actually buying.