Putting automation in the hands of the people who know the work

During the past months, I have been exploring a simpler way to automate manual work: give knowledge workers the tools and support to create and own their “zero-code” automations themselves.

Our team needed to find a way to reduce the hours knowledge workers spend on manual tasks across a company of around 1100 employees, with a team of just four people. Enabling teams to create and own suitable automations was our way to extend that small team’s reach.

With this challenge ahead, after some months of research, experimentation and plenty of evals, I was able to create what we call the Automation Framework. It helps people use agentic AI to turn their knowledge of a workflow into an self-improving automation they can test, iterate and reuse, without generating code or starting a dedicated engineering project for every suitable use case.

The idea is simple: The people closest to a process already understand its steps, exceptions and desired outcome. Instead of waiting for those requirements to be translated into a software project, they can take a more active role in improving the work themselves. The team remains involved in validating the result and refining it as the process changes.

Built on enablement, not just tools

Making automation accessible starts with preparing people to use AI responsibly. Our enablement framework provides a foundation in agentic AI, usage guidelines, compliant tools and AI governance. From there, teams can explore their own workflows with support from our task force team.

Through this enablement work, we have onboarded many knowledge workers to AI and helped them develop more mature ways of using it. Within just three months, their usage had surpassed engineering across adoption, credits consumed, plugins created and agents created.

This combination makes it easier to move from learning about AI to applying it to real work. It also gives teams a shared understanding of the boundaries: what data and tools they can use, when a result needs review, and when a connection or higher-risk use case needs further approval.

Working where people already are, without extra cost

While FDEs and AI engineering work delivers high value to the most critical projects, we understand that the major of automations within a company, can be done through enablement and a clear, safe and scalable automation framework that people can apply independently.

We built the approach using OpenAI solutions that our people already use. That means teams can develop their automations in a familiar environment, rather than having to learn and adopt a separate interface for each opportunity. We developed this approach without using forward-deployed engineers (FDEs).

It also gives us an alternative to relying on a third-party company’s own automation solution, which could mean spending tens of thousands of dollars more and becoming dependent on a specific UI tool. By building on the capabilities we already have, we can explore suitable workflows while keeping more ownership of how they are created and improved.

From workshop to working workflow

Our first execution workshop focused on a metadata workflow that was composed of many manual hours and a team composed of no previous engineering expertise. During the session, the team automated part of the manual process and learned a repeatable approach they can apply to other opportunities.

Our current estimate is that the automated workflow takes around 40 minutes to complete work that would otherwise require around 40 hours of manual effort.

The workflow can also run tasks in parallel, creating an opportunity to increase throughput as the team develops it further.

As a point of comparison, just a part of this automation was expected to take 4 months of different engineering teams to get it sorted. The full automation wasn’t even brought to the scope discussion because it would take too many engineering resources.

A different kind of ownership

My goal is to democratize AI automation, so everyone has the opportunity to discover new ways of getting their work done. With the right foundations, tools and support, people can create and own suitable automations, freeing up time to question how work is done and improve the processes around them. I want people to have room to experiment and rethink their work, rather than feeling that meeting every deadline depends on more manual effort.

There is still a place for engineering, especially for integrations, complex systems and work that requires additional controls. But many opportunities should no longer have to wait for a traditional development project before a team can explore them.

This is an early step. As we learn from the experimental teams, we will continue improving the approach and measuring the value it creates.

Next steps

We are expanding the experiment to a control group of six different teams and sharing the workflow with an internal network of around 50 AI Champions across different parts of the organization. This will help us learn how the approach works across a wider range of tasks and how teams can support one another as they adopt it.

We are also researching new capabilities that could soon allow knowledge workers to run these automations at greater scale through coordinated swarms of agents. One inspiration is OpenAI’s Navier–Stokes research, where groups of agents explored different approaches and Codex consolidated their intermediate insights. We are drawing on that coordination principle, rather than aiming to replicate the scale or capability of the research system. Our aim is to investigate how that idea could translate into everyday knowledge work, with clear responsibilities, shared goals and human oversight. This is a research direction, rather than a capability we have already rolled out.