AI Spend Hit 2.5x Budget. How Much Went to Agents Guessing at Your Data?
Agents now drive most AI usage, and unclear data multiplies the bill
Most organizations attribute runaway AI costs to model pricing. Pick a cheaper model, negotiate a better rate, cap the seats, and the bill comes back under control.
The larger problem is who is doing the spending. Menlo Ventures' 2026 State of Enterprise AI report, released November 3, found enterprise AI spending rose from $33 billion in 2025 to $176.8 billion by September 2026. On OpenRouter, agents now generate 80% of all tokens, roughly seven times human usage. And companies that budgeted nearly twice their 2025 spend "ended up spending 2.5 times what they'd budgeted."
Much of that agent work is reading your data: looking up a customer, reconciling a figure, finding the right table to query. When the answer is ambiguous, an agent does not stop and ask a colleague. It searches again, tries another table and checks its own work, and every step is billed.
That gets worse as you scale. A human analyst who learns which "revenue" table is the right one remembers it. A thousand agents running a thousand times a day relearn it on every run, unless the context is written down somewhere they can read.
As agents overtake people as the main users of enterprise AI, every agent needs to find the right data, and the right other agents, without guessing. That is what we are building at OpenMap: The Agent Network, the connective tissue for the world's AI agents, starting with every person and every local business. We are raising a £5M seed. Meet us at openmap.ai/investors.
Why Does an Agent Spend More When Your Data Is Unclear?
Picture a finance agent asked for last quarter's revenue by region. It finds four tables with "revenue" in the name. Two are deprecated, one excludes refunds, and the owner of the fourth left last year.
A human analyst would message someone. The agent queries all four, compares the results, notices they disagree and tries to reconcile them.
The answer arrives.
It cost several times what it should have, and it may still be wrong.
This is not a model pricing problem. It is a metadata problem.
The Cost Isn't More Users. It's More Guesses.
Menlo's Derek Xiao put the shift plainly: enterprise spending "is no longer constrained by the number of employees using it, but by how much work it can get done."
That cuts both ways. When agents do the work, every gap in how your data is documented turns into repeated, billable effort.
Before an agent can answer a routine question, it often has to:
Work out which of several similar tables is current
Guess what a column like "status" or "amount" actually means
Decide whether a dataset is complete or still loading
Check whether it is allowed to use the data at all
Reconcile two sources that disagree, without knowing which one wins
None of these tasks are particularly difficult. A person with context resolves them once and moves on. An agent without context resolves them on every run, and you pay for it every time.
How to Cut Agent Data Waste in 30 Days?
You do not need to redesign your data platform, or switch models, to bring agent costs back in line.
Within the first month, focus on the handful of datasets your agents query most. Your query history and token logs will show you which ones they are.
Start by certifying one source for each core concept, such as revenue, customer or active account, and deprecating the look-alikes. Give each one a named owner.
Next, expose this context through the same catalog or retrieval layer your agents already use, so they find the definition before they start searching.
DataManagement.AI accelerates this process by automatically discovering metadata, mapping lineage across the modern data stack, identifying downstream dependencies, and creating a centralized knowledge layer that stays continuously updated as your environment evolves. It resolves duplicate and conflicting records, and ownership and policy travel with every entity, so an agent knows which table to trust before it runs a single query.
Instead of spending months documenting assets manually, your teams get a context layer agents can read from day one. Every query that lands on the right table the first time is a query you only pay for once.
AI Cost Overruns Are Actually a Data Context Problem
Model prices will keep moving, and switching providers will always save something at the margin.
But when agents do most of the work, the biggest variable in your AI bill is how often they have to guess.
Organizations that scale agents successfully do not just buy cheaper tokens. They give their agents data that never needs to be guessed at.
Warm regards,
Shen and Team