Decision latency in the AI age: More alternatives, slower choices

How cheap alternatives make choices expensive

In Decision latency: When choices that aren’t made slow the team down I described the pattern: questions that get identified, discussed – and never closed. In the follow-up (in Norwegian) I looked at what happens when the slowness comes from above.

There is a third variant. The newest. And perhaps the most seductive.

Because it doesn’t come from caution or a missing mandate. It comes from abundance.

“We could also try…”

The team needs to choose an approach for an integration. A discussion like this used to start from two or three alternatives someone already knew.

Now it takes five minutes to produce eight.

And in the middle of the discussion someone says: “Wait, I just asked about one more variant – listen to this.”

The meeting doesn’t end with a choice. It ends with more alternatives than it started with.

No one was being difficult. No one was stalling. Everyone was curious and thorough.

But the question is still open. And next week there are two more variants.

Postponement in new clothes

Classic decision latency dressed itself in caution: “we need more information”, “let’s hear all the voices”.

The new variant dresses itself in productivity: “let’s explore one more”.

That’s what makes it dangerous. Pulling in yet another AI proposal feels like progress. It produces text, diagrams, comparisons. The meeting notes look impressive.

But progress isn’t measured in alternatives generated. It’s measured in questions closed.

And the arithmetic has fundamentally changed: when alternatives were expensive to work up, the cost forced a conclusion. When they’re free, there is no natural stopping point.

The stopping point now has to be a human.

The good-enough point must be set explicitly

Teams used to trust that the exploration would stop by itself. Someone had to do the work, after all. Now the stopping point has to be decided.

Bezos’s rule of thumb, from Amazon’s 2016 letter to shareholders, is a usable default: most decisions should probably be made with around 70 percent of the information you wish you had, and if you wait for 90 you are probably being slow.

Which means the tech lead has to ask a new question before the exploration starts:

“What do we need to know to choose – and when do we know enough?”

Concretely:

• define the criteria first. What weighs most: operations, complexity, competence?
• set a frame: “we evaluate at most three alternatives against the criteria”
• set a deadline: “we decide Thursday, with what we know then”

It sounds rigid. In practice the frame is what makes the exploration safe: the team can go deep on three alternatives instead of skimming eight.

Reversibility has gotten cheaper. Use it.

AI hasn’t just made alternatives cheap. It has made rebuilding cheaper – the first version of it, at least. What generated code costs to live with is 10 September’s subject. A choice that turns out wrong still costs less to reverse than before.

That changes the decision calculus: more choices than before are effectively reversible. And reversible choices deserve fast decisions.

That line was drawn long before any of this. In Amazon’s 2015 letter to shareholders Bezos called most decisions two-way doors: “if you’ve made a suboptimal Type 2 decision, you don’t have to live with the consequences for that long”, and such decisions “can and should be made quickly by high judgment individuals or small groups”. What has changed is how many of a team’s choices now stand on that side of the line.

The tool is the same as in Handling disagreement in technology choices (in Norwegian): the revision point.

“We choose this now. We evaluate after the first release.”

Then the decision becomes learning-based. And learning-based decisions are easier to make: the choice can still be the wrong one, but then you find out at the revision point, while reversing it is still cheap.

The three sentences, updated

In the first latency article I wrote about the sentences that close questions: “We know enough now.” “We can adjust this later.” “Let’s pick a direction.”

They still apply. But the AI age needs a fourth:

“We don’t need more alternatives.”

Say it out loud. In the meeting. Kindly, but clearly.

Because someone has to say it, or no one will. The curiosity is infinite. The calendar is not.

The expensive part is choosing

AI removed the bottleneck in producing alternatives. It created a new one in choosing between them.

Teams that don’t see it drown in their own possibilities – with the best intentions, at high speed, with no direction.

Closing questions was always part of the quiet responsibility.

Now that opening them is free, closing them has become its very core.