AI isn't magic. It's metered, and the meter is the most honest thing about it.
Everything you send it gets billed, in small pieces of text the industry calls tokens. Everything it sends back gets billed too. Ask it to draft, check, rewrite and check again, and you haven't bought an answer. You've hired a committee that bills by the word.
The industry calls this "agentic." Your accountant will have a different word.
The meter never sleeps
I listened to Friday's This Week in Startups, the episode on OpenClaw, and you can hear the shift happening live. We're not talking about chatbots anymore. They're talking about agents: software that doesn't wait to be asked. It watches, decides and acts, around the clock.
Jason Calacanis made the point nobody puts in the demo. The token bill on agents can quietly outrun a developer's salary. He's right, and it isn't close.
Let an agent read every Slack message, summarize every email, watch every document edit and search the web every hour, and congratulations: you've hired a shadow staff that never clocks out and invoices monthly.
Publishers already know how to watch payroll, hosting, insurance and software subscriptions. Add one more line. Tokens are payroll.
What actually runs up the bill
Mostly three things:
- How much you send. Attach a long document to every request and you pay to have it read every time.
- How many times you send it. Retries, tool calls and multi-step chains turn one question into a dozen.
- How many people send it at once. A pilot with three users is a hobby. A Monday-morning rollout is a budget event.
Then there's the overhead every serious organization adds: logging, security, policy checks, audits. All necessary, none of it free.
An agent multiplies the first two, and one that runs around the clock adds to the third. A plain summary is one call. An agent that summarizes, fetches documents, drafts a response, checks policy, rewrites for tone and then verifies its citations is a pile of calls before anybody looks. Multiply that by a team and the meter starts sprinting.
How we run it at the Mercury
We're building an AI-first newsroom, which means we design the work around what the machine can carry before we design it around headcount. AI handles meeting transcription, draft structure, research, headline testing and the search-engine plumbing.
And it runs on a leash:
- No standing watch over Slack or email.
- Material comes in batches, on purpose.
- Research loops have a cap. Draft passes have a limit.
- If an agent can't earn its keep inside those limits, it doesn't get deployed.
An AI-first newsroom is not an AI-unlimited newsroom.
The cheapest model that does the job
The only strategy I've seen work is unglamorous: don't pay for the smartest model when the job doesn't need it.
Routine work, like summaries, extraction, formatting and first passes, runs fine on smaller, cheaper models, sometimes on your own machines when policy demands it. Hard work, like complex reasoning, sensitive material and decisions that matter, gets escalated to the expensive one. Then measure the right number: cost per outcome, not cost per prompt.
Do that and AI is leverage. Skip it and AI becomes a line item that gets euthanized at the next budget review.
The agent era is here. The winners won't be the ones running the most agents. They'll be the ones who know what each one costs and have the discipline to say no. Everyone else is building a very expensive science project.
I'm pro-AI and pro-efficiency. I'm also pro-not-setting-money-on-fire. Now if you'll excuse me, I'm going to spend $6 on a coffee and experience metered consumption in analog form.
Peter Cellino is the publisher of The Charlotte Mercury and the creator of Mercury Local, a platform rebuilding local news into something that serves residents and the local businesses who need to reach them, without the ad-tech middlemen.