The DIY Tax on AI

Uber burned its entire 2026 AI budget by April. Not half. Every dollar, gone in four months. Priceline's routine AI contract renewal came back 4-5x more expensive than the prior term. One enterprise got a $500 million Claude bill after forgetting to set usage limits. By April, the FinOps Foundation was fielding calls from companies reporting they were 3x over their annual token budget. In Q1.

These aren't cautionary tales about reckless adoption. They're the predictable outcome of a structural mistake in how organizations are making the build-versus-buy decision on AI. They confused access to a model with having a product.

Token prices have fallen 280x over two years. Per-token costs are, by any historical measure, cheap. Yet enterprise AI spend has risen 320% in the same period. Total spend isn't a function of unit cost. It's unit cost multiplied by volume, and volume in agentic workflows doesn't behave the way anyone modeled it.

A sales rep researching a buying committee doesn't make one query. They make eight. They ask an incomplete question, get a partial answer, forget context, re-query when the output misses, start over with a better prompt. EY and Gartner document it clearly: agentic workflows consume 5 to 30 times more tokens per task than a standard chatbot. Teams that piloted AI on simple interfaces and then deployed multi-step agentic workflows at scale found their cost assumptions were off by an order of magnitude. Not after the pilot. After the production invoice.

Rep skill variance compounds the problem. Take a 50-person sales team, prompt engineering competence isn't uniform. Your top 20% have developed intuition for context framing. They get usable outputs in one or two attempts. The bottom 30% re-query 4-8 times per insight, building bloated context windows, abandoning threads, starting over. Each attempt is a billable token event. The gap between your best prompter and your worst isn't incremental. It's structural.

Then model drops hit. OpenAI, Anthropic, and Google ship new versions constantly. Each release breaks existing prompts in ways that are unpredictable and hit differently across skill levels. Top performers adapt in days. The bottom 30% spend 4-6 weeks getting back to baseline, burning tokens the whole way. Every model drop is an unbudgeted cost event. Two a year (very conservative) produce a cumulative disruption that CFOs are starting to recognize as a recurring liability, not a one-time adjustment.

That's the visible tax. It's real and it compounds, It also gets worse at scale and it's the wrong problem to fix first.

When the invoice lands, the instinct is immediate and predictable: cut usage, add guardrails, mandate prompt training, build internal prompt libraries, roll out monitoring dashboards. These are reasonable responses to an unreasonable bill. They're also treating the symptom and not the problem.

The root cause was an architectural decision made months earlier, before a single token was billed. Treating a large language model like a product rather than what it is, infrastructure. Infrastructure that commits the organization to a structural tax that governance can't eliminate. You can reduce it but you can't remove it.

Prompt libraries encode how to ask questions. They don't encode what your team knows about the accounts, the buyers, the market. Companies responding to token overruns with governance programs are making a second bad decision in attempting to fix the first one. They're doubling down on architecture that was misaligned from the start.

Gartner forecasts 40% of AI agent projects cancelled by 2027. Not technical failure. Not market fit. Economics. The builds that survive will be the ones where the architecture justifies the cost structure. In sales intelligence, almost none of them do.

The 80/20 holds here as well: roughly 80% of deal outcomes are shaped by 20% of the intelligence quality that drives them. The stakeholder map that finds the real blocker. The budget trigger that surfaces the actual funding window. The competitive insight that reframes the evaluation. These are the moments where intelligence changes outcomes. In a DIY agentic environment, they're randomly distributed. A roll of the dice.

When there's no persistent memory, the 20% of intelligence that actually moves deals happens by accident. It occurs when the right rep frames the right query with the right context on the right day. Not by design. Your top performer gets it more than the median rep. Average intelligence quality is set by variance and high variance doesn't just produce poor results. It produces unpredictable ones. And remember the original goal was to create less variance and higher predictability.

Three things follow from that:

The 80/20 split in sales intelligence isn't a law of nature. It's an architectural outcome.

Build on contextless, per-session queries with no persistent memory and no verification layer, and the 20% stays accidental. You can't train to consistency. You can't govern to a quality floor. You raise the average. The variance stays. And that's where the invisible tax lives.

A tool gives reps a better way to query an LLM. An intelligence layer makes the conditions for bad intelligence structurally impossible. It turns the 20% into a floor instead of a ceiling.

Fixed predictable pricing follows directly from well-engineered architecture. When the system produces consistent, verified, deterministic intelligence, the cost is predictable. No re-query loops. No blank-slate overhead. No model drop disruption at the rep layer. The token cost is absorbed, version-controlled, and included. It behaves like a line item because it is one.

Companies managing token overruns with governance programs have made two decisions: build, then treat the cost problem as operational rather than architectural. The second decision extends the liability of the first.

The question that resolves it isn't "how do we reduce token consumption?" It's "what problem were we actually trying to solve, and does this architecture solve it?"

In sales intelligence, the problem isn't generating text. It's producing verified, consistent, actionable deal knowledge — at scale, for every rep, on every opportunity, regardless of their prompt skills or whatever model version dropped last Tuesday.

An LLM access layer doesn't solve that. It gives individuals the raw material to attempt it, with widely varying results, at compounding and unpredictable cost.

The token bill is the invoice for that attempt. The DIY tax is what it costs to keep trying.