You just hired a million bad employees
Companies handed everyone unlimited agents and token budgets. Bad processes now replicate by the second. The next lesson isn't buying more model — it's learning to manage this digital workforce.
- Hebbia founder George Sivulka argues that once companies hand every employee near-unlimited agent and token budgets, the bottleneck shifts from "is the model strong enough" to "can you manage this digital workforce."
- His core analogy: 19th-century railroads. The tracks (capability) arrived first; dispatching (management) came later. Agents amplify vague tasks and broken processes by the second.
- Seven parallels: burning tokens ↔ throwing bodies at a problem, loops ↔ meetings about meetings, wasted tokens ↔ headcount bloat, 100x context ↔ 10x engineers, hoarding context ↔ job security, evals ↔ OKRs, transformation firms ↔ the next big industry category.
- Scaling depends on evals — repeatable acceptance criteria. Coding works first because pass/fail can be judged automatically.
- His business conclusion bets on "AI transformation" firms that codify legacy companies' differentiated processes into agents; several charts carry extrapolation and "illustrative" labels — read them as a management framework, not an audited report.
Disclosure: the author is founder and CEO of Hebbia. His company embeds AI into real workflows for finance, legal, and other teams — a business shape close to what the piece calls "AI transformation." Later on he argues transformation firms will outgrow Neofirms by an order of magnitude, which lines up with his own commercial position. Read this as an opinionated practitioner's essay, not neutral market research.
The books say productivity. On the ground, it's a workforce that never stops
Companies think they're buying "productivity": signing model licenses, rolling out assistants to everyone, and plugging in agents that can click through tools on their own. What's actually happening is something else: they're handing every employee, including the ones least able to spell out a task clearly, near-unlimited digital headcount and budget. The moment a task is vague, an agent won't stop and ask — it just keeps working, keeps burning tokens. Bad processes used to spread slowly, through people. Now they replicate by the second.
In short: you've hired an employee who'll work unlimited overtime even when it's doing the wrong thing.
Illustrative example (not from the original piece — for comparing against your own workplace)
"Go organize this client folder and give me an update."
- No defined deliverable: minutes, a risk list, or an email ready to send the client?
- No defined scope: which three files count, which chat logs can be ignored?
- No defined stopping point: what counts as "good enough," and when should it pause and ask?
- No defined acceptance criteria: who judges "done," and by what standard?
The same vague instruction goes wrong differently depending on whether a human or an agent gets it. Click the two tabs below to compare (the heading stays visible either way):
People get stuck too, but slowly
A colleague getting this instruction would usually ask first: "Who's it for? Do you need numbers? Is this for tomorrow's meeting?" If it's still unclear, they'd call a short meeting — a day lost, at most. It burns calendar time and patience, but the bill doesn't snowball on its own.
An agent won't complain — it'll just retry forever
Feed the same instruction into Claude Code, Copilot, or any agent harness, and it starts drafting an outline, restructuring, self-critiquing, calling itself again to fix its own output. With no stopping condition, there's no "done." Even a failed attempt gets handed back in flawless formatting, delivered with total confidence. A token is the basic billing unit for how a model reads and writes — roughly the "word count" an AI burns through while working — and that count keeps climbing on the bill by itself.
Unpack the "agent" side above, and a bad loop on the ground often looks like this (one illustrative loop, not audited data):
Receives a vague task
"Organize the client folder" has no defined deliverable and no sense of what "done" looks like.
Produces something that looks complete
Outline, summary, tables — all there, nicely formatted, but nobody knows if it clears the bar.
Self-checks, fails, starts another round
The agent calls itself to fix itself: adding material, restructuring, re-running retrieval. With no human in the loop, the loop doesn't break.
The bill climbs, the business outcome barely moves
Tokens keep burning while the actually usable deliverable is still missing. That's "spending tokens to buy more token-spending."
Per the original piece: loops are a band-aid for the fact that almost nobody writes a clear task spec. Brute-force iteration becomes the system's only way forward.
The data side of the story matches this pattern. The two charts below come from the original piece: per-employee AI spend is climbing, and headcount at high-adoption companies hasn't shrunk either. Read this as "money and headcount are both rising" — not as "AI is reliably delivering results."