Sooner or later, every conversation I have with a CIO about AI turns to cost. What will it cost to run agents across the business? What happens to the budget when there are hundreds of them, then thousands? It's a fair question, and I don't think anyone can answer it with confidence yet.
The price of intelligence has been falling fast. Some people believe tokens will be close to free within a few years, and they may be right. But anyone who quotes you the price curve with certainty is guessing, however confident the chart looks.
So I'd suggest a different question, one you can act on today: if intelligence does become cheap, what stays expensive? That's where the advantage will be. It's also where the investments you make now will still be paying off, whichever way the price curve bends.
Cheap Power Didn't Fix the Factory
When factories first electrified, many simply put a big electric motor where the steam engine used to be. The power was cheaper and cleaner, but it still ran through the same shafts and belts across the same floor, and productivity barely moved. The real gains came later, when factories were redesigned around the new power: a motor at each machine, and work laid out by how it flowed rather than by where the drive shaft ran.
Enterprise AI is in its shaft-and-belt phase. Many organizations are dropping a cheaper, smarter model in where the old process used to be and leaving everything around it unchanged: the overnight pipelines, the copies of copies, the dashboards someone has to remember to open. A cheaper model in that layout gives you cheaper answers. It doesn't give you better ones.
Falling token prices change the economics of the model. They don't change the physics of the business. However cheap intelligence becomes, four things stay scarce: time, correctness, consistency and trust.
Time: The Business Sets the Clock
Every decision in your business has a window, and the model doesn't get to set it. A card authorization has a fraction of a second. A customer on the phone gives you a few seconds before patience runs out, and a replenishment order has a cutoff. Free inference makes none of those windows longer.
What fills the window is mostly not thinking. It's travel: pulling a schema, querying one system, checking another to see whether the answer still holds, reconciling a customer ID that almost matches. Cheaper tokens make each of those steps cheaper, not shorter. And when agents hand work to each other, the waits stack up, so a delay that looks trivial in one step can consume the whole window across ten.
The time that matters isn't how fast a model responds. It's how long it takes from something happening in the business to an action that reflects it. That number lives in your data architecture, not your model contract.
Correctness: Cheap Answers Multiply
When something gets cheaper, people use more of it. If asking questions of your data costs almost nothing, you'll ask far more of them, run more agents and let more of those agents act. That's the point. It also means every source of error gets multiplied by volume.
A model reasoning over data that's four hours old, or over a definition of revenue your CFO wouldn't sign, isn't fixed by being free. It's wrong more cheaply, and because you're using it more, it's wrong more often. A smarter model can make a stale answer sound more convincing. It can't make it current.
Correctness comes from context: the right definitions, the right relationships and the current state of the business. Those are properties of your data, and no price cut on the model will supply them.
Consistency: Many Agents, One Version of the Business
Cheap intelligence will give you hundreds of agents, working alongside your people on the same customers, orders and accounts. One agent reserves inventory while another prices the order. A service rep is talking to a customer whose account an agent updated a moment ago.
That only works if everyone sees the same version of the business at the same moment. When an agent can't see what another agent just wrote, or a person sees one number while an agent acts on another, you don't get a slightly stale answer. You get two actions that contradict each other, both taken with complete confidence.
Trust: Attention Is the Real Budget
As actions get cheap, the scarcest resource in the building becomes human attention. Nobody can review every decision a thousand agents make. So trust has to be designed in, not inspected after the fact.
Trust comes down to a few plain questions. Was this agent allowed to see what it saw and do what it did? What data was the decision based on? Who, or what, changed this record, and can we undo it?
The organizations that scale agents will be the ones that enforce permissions and business definitions on the same data the agents act on. Then every action is accountable by construction, and people can spend their attention on the exceptions that deserve it.
Four Questions to Ask Now
None of these depend on what tokens cost three years from now.
- Which decisions have a clock, and how old is the data feeding them? List the decisions that matter, the window each one has, and how old the data is at the moment it's made. The gap between those two numbers is your exposure.
- What breaks first if questions to your data increase a hundredfold? If the answer is the model bill, that may solve itself. If it's your databases, your pipelines or the people reconciling them, it won't.
- Do your people and your agents see the same row at the same moment? If they read from different copies, they'll eventually act on different versions of the truth.
- Can you trace any agent's action to the data it saw and the permission that allowed it? If not, you'll be limited to the agents you can afford to supervise, and that won't be very many.
The Bet That Doesn't Need a Forecast
Here's what I find reassuring. If tokens stay expensive, you'll want fewer hops between systems, less redundant context and data close to where decisions are made, because that's how you control cost. If tokens become nearly free, you'll want exactly the same things, because that's what's left to compete on. You don't need to know which future arrives to know where to invest.
At SingleStore, we didn't design our database for agents. We designed it for businesses that couldn't afford to be late: transactions, analytics, search and vectors running together on live data, with recent and historical data both within reach. It turns out that what those businesses needed and what AI needs are the same thing.
Intelligence will keep getting cheaper. Time, correctness, consistency and trust won't. Plan for what stays scarce.
