Who decides which model reviews your contract.
Choosing a destination
Most people do not want to choose the car. They want to get somewhere.
Yet that is increasingly what we ask people to do when they use AI tools.
First, choose a provider. OpenAI, Anthropic, Google, or someone else.
Then choose a model.
Then decide how hard it should think. Low, Medium, High, Extra.
If all you know is that you want the best possible result, you will probably pick the best one every time.
Why wouldn’t you?
That may make sense for the user. It makes less sense for the business paying the bill.
Choosing the car should not be my job
The user has little reason to optimise.
Imagine I am an employee using an AI tool through my company’s enterprise agreement.
I want the work done properly.
I am not thinking about token costs or whether a smaller model could have done the job. I just want the best answer.
So I choose the most capable option.
Of course I do.
The problem is that the person making that decision is often not the person paying for it.
Scale that across an enterprise and the cost adds up.
I don’t think the answer is to make every employee better at choosing models.
The system should make the choice.
Not every ride needs a Ferrari
A simple extraction does not need the same resources as a complex contract review.
The agent should decide what the task requires.
That might mean a different model, more reasoning, more context, additional tools or another pass to check the result.
This is where orchestration matters.
The goal is not to use the cheapest model.
It is not to use the most powerful one either.
It is to use what the task needs to deliver the outcome at the expected standard.
Sometimes that means a Ferrari. Sometimes it doesn’t.
Either way, I should not have to choose.
What does the job actually require?
Model choice is only part of the problem.
The harder question is whether the agent understands what it needs to do the job properly.
That matters in legal work because the document in front of you is not always the whole story.
Take DAZN Ltd v Coupang Corp [2025] EWCA Civ 1083. The Court of Appeal upheld the finding that a binding contract had been formed through the parties’ communications, including emails and WhatsApp messages, even though a formal agreement was expected to follow.
Now imagine asking an AI agent to review the draft agreement.
It could compare every clause against your playbook, flag deviations and suggest fallback language.
But what if something material had already been agreed by email?
An agent that only reviews the document has followed the instruction.
It may also have missed the point.
A better agent might recognise that the document is not the whole story. It might retrieve earlier drafts, look for relevant correspondence or tell you that important context is missing.
The agent should not just decide how to answer. It should work out what it needs to answer properly.
Knowing what it doesn’t know
There is a big difference between:
“I don’t know.”
and:
“I don’t have enough information to know.”
The second is much more useful.
We expect this from people we delegate work to.
A lawyer might spot that something is missing, find another document, ask a question or check their answer before sending it back.
The agent should do the same.
Sometimes it should answer.
Sometimes it should retrieve more context.
Sometimes it should use a different model, think for longer or check its work.
Sometimes it should come back to you.
The important part is that the agent makes that judgement.
Outcomes > Output
At Flank, our focus is on the outcome.
So if I ask it to review a contract, I care about the review.
I care that it found the important issues, had the context it needed and checked its work when necessary.
I do not particularly care which model got it there.
Those are decisions about how the work gets done.
They should not become my work.
That is the promise of agentic AI.
Not simply giving users more models and more settings.
But taking the complexity away.
Tell the agent what you need. Let it work out what it takes.
Sometimes it will need a Ferrari. Sometimes it won’t.
You choose the destination. Let the agent pick the car.
