Insight / signal
AI token spend is the new payroll. Manage it like it matters.
If AI is replacing chunks of human labour, AI spend is not just software spend any more. It is payroll by another name.
We keep talking about AI replacing jobs.
Fine. Then we should follow that thought through.
If AI is replacing chunks of human labour, AI spend is not just software spend any more. It is payroll by another name. And most companies are nowhere near ready to manage it that way.
That is the useful AI story this week.
Not another model demo. Not another leaderboard. Not another confident LinkedIn post from someone who discovered agents last Tuesday.
The signal is in the money.
Marketing School has been talking about the OpenClaw founder reportedly spending $1.3m on AI tokens. The interesting bit is not the big number — big numbers are easy content bait. The interesting bit is what the number represents: tokens being used as operational capacity instead of headcount.
At the same time, OpenAI is publishing case studies about Codex turning customer requests into working code in minutes and helping one company compress requirements analysis from weeks to hours. TechCrunch is reporting that some coders now refuse to work without AI, while also pointing to evidence that AI-generated code creates maintenance debt downstream. The Decoder carried a reported story about an unnamed company allegedly spending $500m on Claude in a single month because nobody set usage limits.
You do not need to take every number as gospel. Some of this is vendor PR. Some is second-hand reporting. Some is probably polished within an inch of its life.
But the pattern is hard to miss.
AI has moved from “nice productivity tool” to “line item that does work.”
That changes the management problem.
A few years ago, if a team bought another SaaS subscription, the damage was usually contained. Annoying, yes. Wasteful, often. But the tool mostly sat there waiting for humans.
Agentic AI is different. It can chew through tokens while it reasons, searches, drafts, codes, reviews, re-runs, retries, summarises, stores context, calls tools, and works across a messy chain of steps. It can look productive while hiding cost, rework, and risk inside the process.
That is why “AI-first” can become stupid very quickly.
Not because AI is useless. It clearly is not. I use it every day. I would not want to go back.
The problem is pretending AI capacity is free, infinite, or self-managing.
If you hired a person, you would want to know what job they were doing. Who manages them. What they can approve. What systems they can access. What good work looks like. How much they cost. Whether they are helping the business or just staying busy.
AI workflows need the same discipline. Maybe more — because agents do not get embarrassed when they waste money. They do not stop and say, “This is probably a daft use of a frontier model.” They will happily use the expensive option to do a cheap job if nobody tells them otherwise.
That is where a lot of companies are going to get hurt.
They will give everyone access to powerful tools, celebrate the usage graph, and call it adoption. Then they will find out the work is inconsistent, the outputs need heavy review, the same context keeps being reprocessed, expensive models are doing low-value chores, and nobody can say which workflows actually saved time or made money.
That is not transformation. That is a bonfire with an API key.
The better way is less glamorous.
Give AI a job description. Not “help the marketing team” — that is mush.
Try: “Turn this sales call transcript into a CRM-ready summary, three follow-up angles, and a draft email. Use the account notes. Do not send anything. Log the source transcript, assumptions, and open questions. Escalate if the prospect mentions budget, timing, procurement, legal, or a competitor.”
That is a job.
Then give it a budget. How many runs per week? Which model for which step? What is allowed to use the expensive model, and when should it fall back to a cheaper one? How much context is actually needed? Can memory be compiled once instead of retrieved again every run? At what cost does the workflow pause and ask a human?
This is where the memory question matters too. Recent agent-memory research points to the same issue from a technical angle: retrieval-at-runtime is expensive when agents keep rediscovering the same facts. Compiled knowledge, task-specific context, and proper memory hygiene can cut waste significantly. Stop making the agent re-learn the business every time it does a job.
Then put a manager on it.
That does not mean a full-time AI whisperer sitting there admiring prompts. It means a named owner for each workflow. Someone who knows what the system is meant to do, reviews the logs, checks quality, watches spend, and kills or changes the workflow if it is not worth the cost.
This sounds obvious. It is not how most AI adoption happens.
Most adoption starts with enthusiasm. People try tools. They share tricks. Someone builds a workflow. Someone else adds a plug-in. A team starts using a coding agent. A marketer wires a content process together. A founder gets excited about an AI employee. Costs creep. Edge cases pile up. Nobody owns the whole thing.
Then leadership asks a horrible question: “Is this actually helping?”
And everyone looks at the usage dashboard as if usage is the answer.
Usage is not value.
Tokens spent are not work completed. Work completed is not work that mattered. Faster output is not better output. A busy agent is not a profitable process.
That is the bit I think business owners need to hear now, before the invoices get silly.
If AI is doing work, manage the work.
For a marketing team, that might mean one supervised workflow for research-to-draft content, with source links, approvals, and a hard line between internal drafts and public posts.
For sales, it might mean call summaries, CRM updates, lead scoring, and follow-up drafts — with human approval before anything leaves the building.
For ops, it might mean daily exception reporting: missed tasks, failed automations, customer messages that need attention, systems that look odd.
For software, it might mean coding agents on narrow tickets, with tests, review, scope limits, and proper tracking of any bug-fix work caused by AI output.
None of that is very sexy. Good. The sexy version is usually where the mess starts.
This is also where agencies and AI consultants need to grow up a bit. “We can save you headcount with AI agents” is an easy line to sell. It is also a dangerous line if you cannot manage the operating layer underneath it.
The better offer is more specific:
We will take one repeated commercial process, turn it into a supervised AI workflow, measure the time saved and output quality, cap the spend, log the decisions, and keep a human accountable.
That is less dramatic than promising an autonomous AI workforce. It is also much more useful.
Because the real shift is not that AI can write more stuff.
The shift is that AI is becoming part of the company’s capacity model. Part labour, part software, part infrastructure, part risk.
That means the post-agency opportunity is not “more content with fewer people.”
It is building the operating layer: workflow design, memory, permissions, approval gates, evaluation, model routing, cost controls, and evidence that the system is actually improving the commercial machine.
Not in generic AI hype. Not in prompt packs. Not in pretending every business needs a swarm of digital interns running around unsupervised.
Just practical, governed AI systems that make the business faster without turning it into a slot machine.
AI does not remove management.
It moves management into the operating system.
And if companies are going to spend like AI is payroll, they had better start managing it like payroll.