Today's AI prices are subsidised and the demos are flattering. How to scale your team in the AI era without building on economics that will not hold.
Here is an unfashionable opinion: AI will never be cheaper than it is today.
That sounds wrong. Headline model prices keep falling, and every few months a new model does more for less. But the price you pay today is not the full cost of what you are using. A lot of it is subsidised by investors, by platforms fighting for market share, and by generous plans designed to get you hooked. At some point the bill arrives, and the businesses that built their plans on today's prices will feel it.
Impressive is not the same as viable
There is a lot of AI out there that looks incredible in a demo and could never stand up in production. The demo has one user, one clean example and nobody watching the meter.
Run the numbers on real usage and the story changes. We recently looked at an AI add-on sold for a few hundred dollars a month. Each answer cost a few cents in model fees, and each customer transaction triggered eight or nine calls. For a busy customer, the very behaviour the product is meant to encourage, more usage, is the thing that makes it lose money.
The same pattern shows up when AI replaces people. Plenty of AI-powered service operations still keep humans behind the scenes, because the AI cannot yet be trusted on its own. The saving in the announcement and the cost on the books are often two different numbers.
This is not an argument against AI
Teams that are not using AI are already falling behind. New businesses are producing in a few months what established teams have not touched in years, largely because they automated the busywork first. The question is not whether to use it. It is where.
Automate the grunt work, keep the judgement
The best returns we see are in the unglamorous work:
- Research that used to take weeks, done in hours.
- Reporting that compiles itself and flags what changed.
- QA that runs automatically, rather than living in a pair of Word documents.
- Content drafts grounded in real data, then edited by a person.
What stays human is the part clients actually pay for: judgement, taste, knowledge of the business, and being accountable when something goes wrong. AI makes good people faster. It does not make the need for good people go away.
Build on economics that hold
- Price your AI usage at scale, not in the demo. Multiply cost per call by calls per customer by your busiest customer's volume. If that number frightens you, redesign before you launch.
- Assume prices can rise. Plan with a margin for model costs going up, not only down.
- Keep people where trust matters. Customer-facing, legal and financial decisions need someone who can be held to account.
- Measure outcomes, not output. More content, more features and more code are not the goal. Revenue, retention and customer satisfaction are.
- Keep your team flexible. In a market moving this fast, a team that can grow and shrink with the work beats a big permanent hire made on today's assumptions.
Summing up
Bubble or not, the businesses that do well will be the ones that used AI to remove friction, not the ones that bet everything on today's prices and tomorrow's promises. Use it hard on the busywork, keep people on the judgement, and do the maths before the bill does it for you.