AI Is Fast. Why Are Decisions Still Slow?

When I read Boris Cherny's Steps of AI Adoption, I immediately thought of an internal memo I had shared with our team, advisors, and investors.

Boris keeps hearing the same story from engineering teams: one person has dramatically increased output with AI, while the rest of the organization has not caught up.

We were seeing the same problem from another angle. AI could make every function faster while the company still took weeks to decide what to do.

His article motivated me to share our reflection publicly.

We built UserApproved on an early assumption that coding would stop being the main bottleneck in building software. As implementation gets faster, the harder questions become: should we build this, do users care, and what changed after we shipped?

We focused on helping teams answer those questions from customer and product evidence. Along the way, we learned that finding an answer and getting a company to act on it are two different problems.

The company had data but no shared decision

At one company we worked with, leadership wanted to expand our work after we found useful opportunities in one part of the customer journey.

As more functions joined, communication channels multiplied. Each team brought its own data, goals, language, and constraints. The evidence for one decision was split across teams, tools, and decision rights. No one began with the complete picture.

I have seen the same pattern in many leadership standups. One person brings an analytics chart. Someone else brings customer feedback. Another person explains a technical or operational constraint. During the meeting, the group discovers that an important question is still unanswered.

The meeting ends with a request for more information. The next meeting begins by rebuilding the context.

A decision that looks small from the outside can take weeks or months.

Giving every team an AI copilot can make this worse. Each function can produce its own analysis, prototype, and recommendation faster. Leadership then has more material to reconcile, with the same missing context and unclear ownership.

The company has more output. The decision takes just as long.

Time to decision is the missing productivity metric

A six-month randomized field experiment from Microsoft illustrates the divide. Half of 6,000 knowledge workers received access to a generative AI tool inside the applications they already used. Active users spent about three fewer hours each week on email and appeared to complete documents moderately faster. Time in meetings did not significantly change. Shifting Work Patterns with Generative AI

The result matches what I have seen. Work that one person controls moves first. Meetings move more slowly because they carry dependencies, authority, negotiation, and shared risk.

This is the distinction I now use:

Individual work:  Prompt → Output

Company work:     Signal → Shared evidence → Decision → Action → Learning

Individual work:  Prompt → Output

Company work:     Signal → Shared evidence → Decision → Action → Learning

Individual work:  Prompt → Output

Company work:     Signal → Shared evidence → Decision → Action → Learning

For one person, time to output is useful. For a company, I care much more about time to decision, then time to a verified result.

Imagine Monday working differently

Imagine that before a Monday leadership standup, everyone spends fifteen minutes reviewing the same backlog of decisions.

Each item has already been investigated enough for the group to see the evidence, options, and remaining uncertainty. The meeting does not start by reconstructing what happened last week.

The meeting begins with three possible responses:

Yes. No. More information.

Yes assigns an owner. No records the reason. More information becomes a specific question instead of restarting the whole discussion.

A thirty-minute meeting could resolve several decisions that might otherwise move through weeks of documents, messages, one-on-ones, and repeated reviews.

This is one direction we are building toward at UserApproved. A team can also begin without replacing every system or centralizing every piece of data.

Start with one slow decision

Choose one recurring decision that takes too long. Name the owner and the minimum evidence required. Make the source and uncertainty visible. Agree on what the team needs before it can say yes, no, or ask one bounded question.

Then measure how long it takes to move from signal to decision, and from decision to a verified result.

That will tell you more about company productivity than the number of AI seats, prompts, or generated artifacts.

I expect the next phase of workplace AI to focus less on helping one person create another artifact and more on helping the right evidence reach the people who can make a decision. Managers will spend less time routing information. Meetings will spend less time recovering context and more time making choices.

If Monday still starts with everyone explaining what happened last week, AI has made the people faster, not the company.

Personal AI shortens time to output. Organizational AI has to shorten time to decision.