Less Is More

A few months ago, I was walking a customer through how UserApproved could send experiment ideas into their backlog.

We had built enough of the agent workflow to generate experiment ideas. The obvious next step seemed to be automating more of the development and experiment setup.

The customer pushed back and shared what had happened after they connected Claude to Jira.

“Our dev team is freaking out because they went from having like 50 tickets to now like 5,000 … this is getting really out of hand. I don’t want to replicate that here. It really is a waste of time and actually ends up getting more work than less.”

His request was simple. Do not send everything into the backlog. Build a connector to Monday and Claude so they can add only the ones that matter most.

That conversation changed how I looked at what we were building.

The valuable AI system is not the one that creates the most work. It is the one that helps a team confidently ignore most of it.

I thought more insights meant more value

When we started UserApproved, the problem seemed clear. The companies we worked with had plenty of analytics, customer feedback, session replays, and research, but they still struggled to turn all of it into action.

We built long-running agents to analyze, simulate, and debug opportunities across ecommerce operations and shopping journeys. The idea was to review more situations than a small team could cover manually.

At first, we optimized for coverage. We produced more insights, deeper cross-referenced analysis, and more checks. But the reports became deeper too, and the list of opportunities kept getting longer.

Then reviews started taking longer.

The feedback sounded different, but it kept landing in the same place:

“This is an amazing consultant-like report. Let me find an hour to read it next week.”

“This is a great list. Where should I look first?”

“It is spectacularly detailed, but overwhelming. Can your human consultant tell me the top three things to act on tomorrow? We already have a huge backlog. Can you review that too before calling anything top three?”

Customers did not want another list of ideas. They already had guesses and suggestions from marketing, product, executives, and external consultants. Their problem was still the same: what should I prioritize this week?

Even when most findings were useful, a few weaker ones could make customers stop reading and trust the whole list less. Every additional finding asked someone to understand it, compare it with everything else, and decide what to do.

Plausible work is almost free now

AI does not just make good ideas faster. It makes plausible ideas so cheap that people start to ignore them.

Most are not obviously bad. That is what makes the problem difficult.

Someone still had to decide whether the issue affected enough customers, whether the evidence was strong, whether the fix was realistic, and whether it mattered more than the other work already in progress.

The cost did not disappear. It moved. AI sped up one step and slowed down the whole system.

Across the growth team’s full workflow, our automation was starting to add overhead instead of removing it.

The product became the filter

This changed our product principle. We could not optimize one part of the customer journey and call the job done. We had to improve the whole decision workflow, with AI and people owning different parts.

The workflow I want looks like this:

Agent: Signals → Hypotheses → Evidence → Ranked experiments
Growth lead: Approve, reject, or ask for more
Team: Design → Build → Run
Agent + team: Monitor → Learn → Update context

Agent: Signals → Hypotheses → Evidence → Ranked experiments
Growth lead: Approve, reject, or ask for more
Team: Design → Build → Run
Agent + team: Monitor → Learn → Update context

Agent: Signals → Hypotheses → Evidence → Ranked experiments
Growth lead: Approve, reject, or ask for more
Team: Design → Build → Run
Agent + team: Monitor → Learn → Update context

The agent can handle data analysis, hypothesis generation, simulation checks, and technical debugging. But the user-facing result should not be everything the system found. It should be a grounded, ranked backlog that a growth lead can confidently act on.

The deeper context still matters. The system should preserve the analysis, rejected ideas, decisions, and results so future work does not restart from zero. Depth stays underneath. Attention goes to the few things that matter now.

The product is not the list. It is the filter and the learning loop that help the customer do less work.

When we design any product, we ask:

Did the team decide faster? Did the right work get an owner? Did someone act on it tomorrow? Did the result improve? Did we reduce the investigation people had to repeat?

If an agent gives your team 5,000 new things to do, it did not save time. It just moved the work.