Read time: 4 minutes

Leader, welcome back.

The more instructions you give AI, the better the answer should get. Right?

Not always.

One of the easiest ways to limit a good AI model is to give it a detailed plan for how to solve your problem.

There is a better place to be specific.

Most people hear “be specific with AI” and instinctively start listing steps:

It feels rigorous.

It is also often the wrong kind of specificity.

Because once you prescribe the route, you have quietly constrained the model to your way of solving the problem.

You may get a cleaner version of your thinking.
You are less likely to get something you had not considered.

Instead, try being extremely specific about the outcome.

Tell the AI:

Then ask it to work out the possible routes.

That is not giving AI free rein. You still make the decision.

The model does not understand your organisation, risk appetite or the politics in the room the way you do.

But you are using it for something more valuable than following instructions. You are using it to widen the option set before you exercise judgement.

Instead of:

Try:

Same problem.

Very different job for the AI.

Or:

Instead of:

Try:

That second prompt makes the AI think much closer to the way a good analyst would.

That last sentence matters.

You are explicitly giving the model permission to disagree with your first idea.

And this works far better now than it did a couple of years ago. Modern models are increasingly capable of taking an end state, reasoning through different routes and working towards a stopping condition rather than needing every intermediate step prescribed.

The limiting factor is increasingly not the model.

It is the context we give it.

This also pairs well with the RACEF prompting framework I teach in my course.

RACEF helps structure the instruction.

Outcome-first prompting helps decide what job you should actually be giving AI in the first place.

Different ideas. Much stronger together.

See the full AI course here.

One of the biggest AI deals of the year is not about building a better model.

It is about deciding which model gets used.

Stripe has agreed to buy OpenRouter in a deal reported at around $8bn.

OpenRouter gives developers one interface through which they can access more than 400 AI models, switch between them and manage usage. It now handles more than 10 trillion tokens a day.

That makes the acquisition interesting for a very simple reason.

Stripe does not need to predict which AI lab wins.

It is betting that companies will increasingly use multiple models and care about what sits between those models and the customer:

routing, billing, usage, reliability and cost.

And there is evidence for that behaviour already.

On OpenRouter specifically, Chinese models overtook US models in token share earlier this year. DeepSeek went from around 9% of tokens in January to roughly 18% by June, while several other Chinese model providers also gained share.

That does not mean Chinese models now dominate AI globally.

OpenRouter is one platform, and token volume is not the same thing as enterprise market share.

But it does show something important.

When developers are able to switch models relatively easily, price and performance can move demand very quickly.

There is another signal.

Chinese lab Z.ai released GLM-5.3 last week. On one vulnerability-discovery benchmark it scored 84.5%, ahead of several leading US models, although it remained significantly behind the strongest models on harder exploitation tests.

Z.ai was sufficiently concerned by the model's cyber capability that it delayed releasing the downloadable weights for two weeks while it carried out additional safety work. These are primarily the lab's own results, so they should be treated accordingly.

The takeaway is bigger than any individual model.

For years, the normal enterprise question was:

“Which platform are we buying?”

AI increasingly introduces another question:

“Which model should perform this task, at what level of quality, and at what cost?”

A £0.02 task does not need the same model as a £20,000 decision.

A simple extraction task does not need the same level of reasoning as a board paper, contract review or forecast challenge.

So I think teams will increasingly need a metric that barely existed a few years ago:

Cost per completed AI task.

Not tokens.

Not seats.

Not which logo appears in the corner of the screen.

What did it cost us to produce a reliable outcome?

That is a much more useful number for a CFO.

SIGNAL: AI appears to be creating roles as well as removing work

A Lloyds survey of 1,200 UK businesses found 54% said AI had led to job creation, while one in five said they were creating AI-specific roles.

The caveat matters. Larger businesses are better positioned to capture the upside because they already have more technology, skills and investment capacity.

The leadership question is becoming less “Will AI replace people?” and more:

“Which skills become more valuable when AI becomes normal?”

SIGNAL: finance AI is starting to move closer to the P&L

SharkNinja is building an AI tool intended to monitor profitability far closer to real time.

Its current forecasting process can take weeks. The company piloted the new approach alongside its July North American forecast and expects broader implementation by Q4.

Its CFO described AI as a capacity multiplier, rather than simply a cost-cutting exercise.

That is the use case I would watch.

Not another chatbot.

Shortening the distance between something changing in the business and finance being able to see its financial impact.

AI at SharkNinja. Cutting the fat from the forecast without deep-frying the budget!

NOISE (IGNORE): “Model X just destroyed Model Y”

One model beats another by a point on a benchmark.

LinkedIn declares a new king.

Two weeks later, another model launches.

For business buyers, the leaderboard matters far less than:

Can it do your task reliably?

How much human checking does it need?

How fast is it?

What does a completed task cost?

That is the benchmark that eventually ends up in the business case.

Take one task you would normally give AI with detailed instructions.

Delete the instructions.

Write down instead:

The outcome.
The constraints.
What success looks like.
What would make the answer unusable.

Then add:

See what changes.