I recently caught up with an old colleague who works at a bank.  He is not in an engineering role, but very technical and curious.  We were discussing the roll out of end user AI at the bank (365 Copilot).  He told me that he had already built 20 custom agents in the copilot environment.  In fact, he had brought his own notes taking agent to our meeting and it sent me a synopsis afterward - very nice.

This is the mythical 'AI Power User’ made real.  Someone who is curious and self-motivated to use  AI to solve all sorts of problems that others maybe don’t even know they have.  Previously, these power users were making Excel Macros, wiring up iPaaS automations, and building no-code apps.  Now, GenAI has unlocked their creativity by an order of magnitude - and there’s a lot we can learn from them.

How Wide is the Gap?

According to Microsoft, an AI Power user is someone who uses AI to save 30+ minutes a day at work. So, the productivity gain is backed into the definition, but the floor is set at a relatively modest 6% based on a 40 hour work week.  Mileage varies based on task  - 55% for coding, 37% for writing tasks, as much as 34% for customer support.   Stats like these are a constantly moving target as models and agent technologies improve, and they aren’t even taking into account the ultimate full automation of many tasks. What might be a 30% productivity gain for an AI assisted power user today could be a fully automated process tomorrow.

The limiting factor of these productivity gains right now though, is the distribution of power users, which is an unknown fraction of enterprise users.  Given that McKinsey only qualifies 6% of enterprises as AI ‘high performers’, it seems reasonable to assume that power users are a small fraction and most of those are distributed towards the bottom end of the productivity scale (6% gain vs 50%).  Otherwise, we’d see a much smaller gap between the innovation promise of AI and concrete gains for organizations.

Facing the Empty Text Box

What differentiates the Power User from everyone else? Key factors are:

​​Frequency & Depth of Use

Use AI daily across multiple workflows and start workflow with AI as the default.

Task Substitution vs. Task Enhancement

Leverage AI by delegating distinct work steps i.e. reimagining workflows vs. just speeding up tasks.

Prompting Behavior: Iterative, Multistep, Specific

Use iterative prompting with multi-turn refinement, providing extensive context, and using different personas.

Breadth Across Tools & Modalities

Use AI for multiple tasks such as coding, research, meeting notes, and image generation.

Literacy: They Know What AI Is Good At

Understand the strengths and pitfalls of AI and can structure their usage accordingly.

Mindset: Experimentation + Continuous Learning

Try new features and new tools as they roll out, learn from communities, templates, and examples, and capture and share best practices.

The most critical points here are the qualitative and behavioral distinctions vs the quantitative ones.  While you can’t be a power user without a certain level and breadth of use, it is how you approach it that makes the difference.  A normal user will likely be deterred by an empty text box, a power user understands how to break down a complex goal into a series of task-based prompts.

Power to the People

In enterprise, a big component of getting to full AI ROI is going to involve distributing the productivity of the Power User to everyone else.  In other words, a 30%+ productivity boost distributed to 80%+ vs 5% of an organization is the difference between game changing vs marginal gains.

Given that so much of what makes a Power User effective is behavioral, how do we distribute those productivity gains?  Training is slow, expensive, and not super efficient, especially given the rapid state of evolution within the technology itself.  

What if we could distill the patterns from the Power Users and bring them to everyone else as defaults and automations?  

We’ve actually done this already. Specifically, in coding workflows where the millions of coding examples are distilled into patterns that an AI can match on and offer the user intelligent defaults (autocomplete) and automations .  Coding has been the natural place to start and it’s no surprise this is where we’ve seen the greatest AI productivity gains so far.  What if we could apply the same methods to other domains?

Power users give us a preview of the near future: workflows where AI doesn’t just enhance tasks but orchestrates them end-to-end. The challenge—and opportunity—for enterprises is to make those capabilities accessible to everyone, not just the curious few. By distilling the behaviors, prompting patterns, and workflow designs of power users into embedded defaults and guided automations, organizations can unlock far more than incremental gains. Coding proved the model. Now it’s time to apply it across the rest of the enterprise.

‍