The Terrifying Math Lurking Behind Your New AI Workforce

Two nights ago, a friend of mine in marketing sent me a WeChat message with just one line:

"Bro, we've gone all in on AI. We're twice as productive — a month now does what three used to."

I replied: "And your cloud bill for the last couple of days must look nice too."

He went quiet for a while. What he finally sent back was just a string of ellipses.

That's not clairvoyance. I've just seen this kind of situation far too many times.

Why Can't Anyone Do the Math at a Glance?

Let's start with a phrase: total cost of ownership. The acronym TCO sounds scary, but it's a simple idea.

It means: a system, every cent you pay into it from the day you start using it to the day you stop — all added together.

In theory everyone accepts that. In practice people still size things up the old way, the way meant for conventional software.

How does conventional software bill you? A subscription. Per user, per year. The rate is fixed in the contract: so much per year, in writing.

AI is not in that family at all.

AI bills per usage.

Ask it to check one client: a prompt goes in, the model runs one loop, then out comes the answer. One trip equals one charge. The more trips you give it, the faster the counter spins, like the meter of a metered taxi.

And here's what really bites: the machine never sleeps.

At two in the morning, you need to rest. Your AI worker is still in the background, watching your clients for "purchase intent."

That's not dedication — that's the meter still running.

What Does an AI Agent's Total Cost Look Like? Four Ledgers

Ledger One: The Token Bill

What is a token? You don't need to memorize a definition. Just treat it as the smallest unit of "reading" an AI works with.

What it reads in and what it writes back are all billed in tokens.

Give your AI worker a job: send a personalized greeting to ten thousand customers. For every customer, it has to read their record, write fresh copy, then double-check once more. Each back-and-forth burns a few thousand tokens — and you multiply that by ten thousand.

But the most expensive part isn't any of it. It's the "background loop."

Some agents scan your database all day for signals like "this customer is about to buy." You never told them to. They do it anyway. And every hour they spend on it goes straight onto your bill.

The number that really makes your stomach drop isn't in the emails you send — it's in the jobs still running while you sleep.

Ledger Two: The Middleware and Engineering Bill

An AI worker doesn't get the job done alone. It has to be connected to your CRM, your website's APIs, your ad campaigns.

Every vendor repeats the same slogan: "native adapter — just plug it in and you're ready."

And "ready" usually means: the generic connectors can talk to each other. But your own business logic — which customers deserve a coupon, what even counts as a high-quality lead — that whole pile of private rules is something the native adapters know nothing about.

Putting those rules into the machine takes engineers writing code layer by layer, then debugging and auditing it.

So where does that bill really land? Most of it lands on the engineering team's books. You assume it sits in the "AI project" budget, but it's actually been living in another department's pocket. The approval document never mentions it — your year-end profit report will point it out for you.

Ledger Three: The Maintenance Bill

You might think that once the AI is developed and deployed, the spending is over.

Not at all.

When outside systems bump versions and swap interfaces, the AI gets confused. And the prompts and rule documents you wrote once, never maintained since, slowly make it talk less and less like you. Emails sent to the wrong recipient, a discount granted to the wrong customer, a tone that keeps hardening — all of them are on their way.

You'll need a person to stay on top of it: audit today's outputs, fix the dead interface tomorrow, revise the prompts the day after, tune the guardrails the day after that.

Don't treat it as a one-time cost. It is continuous — at the bare minimum, one full-time person, all year round.

The most expensive misconception about automation is thinking that once it's up, it will manage itself.

Ledger Four: The Vector Database and Storage Bill

To make the AI's emails truly feel like they "remember you", it has to store the customer history into a "vector database".

That kind of database costs more than ordinary storage, and its price climbs right along with your customer count.

A thousand customers, one price; when you reach a million, the data has to be stacked across many layers, and every layer needs compute power to run and retrieve.

The longer your client list, the higher this "storage" item runs. Same as electric bill: the more you use, the more you pay.

So What Do You Do, Then?

I've boiled the key points into four steps you can carry to your boss:

One — work out the true cost of a single call, then multiply it against real scale — don't take an average, use the value from your few worst-case days.

Two — all the hours engineering spent on debugging, on making things work, on audit: turn the work into money per person and write it all in a separate item, so it won't keep on living hidden within other budgets.

Three — maintenance shouldn't be seen as "temporary". Keep at least one person on it all year long; the money needs to have been in the budget since the first day that project was created.

Four — storage. Don't figure it with "today's customer count", but with "the number of names you'll have three years from now", and remember to write the growth part in ahead of time.

Back to My Friend

When he'd taken it all in, all he could do was sigh: after all that hard work, automation had ended up digging him a pit.

Which is where I said:

You think you saved on labor costs. In reality, that money just changed its label and moved into the technology bill.

"Still — when it's time to go, go boldly. Just get the numbers squared away before you start."

Later he told me: the accounts came out clean, the budget went up, and his boss signed off fast. Fast as it was, the approval carried an extra clause: from now on, every quarter, the running cost has to be reported in a separate line.

You see — that's arithmetic a person can understand.

Sooner or later, you're going to hire a few AI employees yourself. Don't be afraid of the work they take on; be afraid of the billing you never measured.

Set the books straight in advance, and every cent after that is one that can answer for itself.