The Feature Factory Has an AI Problem

By Joseph (@joem3.bsky.social)
Published:

The Feature Factory Has an AI Problem

What happens when we make it easier to build software without changing how we decide what to build?

Imagine a product organization where every team suddenly becomes twice as productive.

Features that took weeks now take days. Prototypes appear overnight. Backlogs shrink. Engineering capacity, once the constraint behind every difficult prioritization conversation, becomes a little less scarce.

What happens next?

If the organization operates as a feature factory, I suspect the answer is predictable: It produces more features. Roadmaps expand, expectations rise, and teams are celebrated for delivering more than before.

But are customers getting twice the value?

That's the contradiction I find interesting about AI-assisted development. We're becoming faster at producing code, without necessarily becoming better at deciding what software should do for people.

A faster factory is still a factory

In 2016, John Cutler published 12 Signs You're Working in a Feature Factory, describing organizations that celebrate shipping features without knowing much about their impact.

In his 2026 retrospective, he adds nuance to the critique. Sometimes a list of features is exactly what a team should deliver. If the customer problem and a workable solution are well understood, efficient execution makes sense. And an organization can perform all the rituals of outcome-driven development without making better products.

The difficulty is knowing when execution is enough—and when your assumptions need another look.

AI doesn't create that difficulty. It can make acting on those assumptions dramatically easier.

When teams ship their organizational structure

In June 2026, Eric Olive and I presented Tackling the Feature Factory: Scaling UX Research for Platform-Level Impact at UXDX, drawing on our work at DocuSign.

We described a familiar paradox: independently successful product teams can collectively produce a fragmented customer experience.

Each team has legitimate commitments and measures of success. Each competes for resources, roadmap space, and visibility in the interface. Over time, what gets built starts to follow the boundaries of team ownership. Nobody sets out to ship an org chart, but the product can end up looking like one.

Now of course, the customer's work doesn't respect those boundaries. In fact, it typically starts before they open our software and continues after they leave it.

This challenge is especially visible in platforms serving many industries. An agreement—to take the Docusign example—might be one step in a sales process, supplier onboarding, or hiring. The platform may offer common capabilities, but the surrounding work, people, and pressures are different.

Adding more isolated research studies doesn't necessarily solve the fragmentation. Teams need a way to connect what they know about the customer across product boundaries.

How to avoid shipping the org chart

In the Audience Research program I designed at Docusign, one organizing principle was to map the customer's end-to-end work at several connected levels.

Think of a procurement process. Its phases might include big chunky categories of things like identifying a need, selecting a supplier, negotiating terms, and managing renewal. These form a broad view of the process, whether the company is a small SMB or a large enterprise.

Within the supplier-selection phase, there are multiple core jobs: evaluate candidates, assess risk, review relevant terms, secure approval. Each describes meaningful work that has to get done, not a feature someone needs to click. Notice that they are more specific then the phases, but still work across industries and scale

Finally, each core job contains multiple job steps. Assessing risk might involve collecting supplier records, verifying requirements, and resolving an exception. Exactly how those steps unfold depends on the industry, organization, and use case. The same broad job can look quite different in a regulated enterprise and a small business. This is where the business strategy and customer targeting intersect the higher level, abstract understanding that has holds lasting value for the organization. 

The point isn't the taxonomy. It's the movement between levels. Phases orient multiple teams around a shared process; core jobs show where particular problems and capabilities belong across the customer process; job steps reveal what customers actually have to do, and where an intervention could help.

Job-performing archetypes provide another anchor: the work stays recognizable whether responsibilities are spread across five departments or handled by one owner-operator.

You can now ask a more useful question than Which team owns this feature? Where in the customer's work does this capability help—and what happens immediately before and after it? How does the feature need to be tweaked or oriented when selling to a new customer profile?

These are questions that go beyond product teams. Field teams including Solution Engineers, Customer Success, and product marketing teams are also enabled around this model. In fact, it helps to align the whole, cross-functional organization.

For UXR teams, the broader structure gives a relatively durable frame, while more specific observations can be updated as customer practices change. It isn't a perfect map of reality. It is a shared model that teams can challenge and improve.

One model, many product decisions

The bigger opportunity is to make that model useful throughout development, not just during an initial discovery study.

In strategic discovery, a research team can work across product areas to investigate end-to-end processes, identify underserved jobs, and help partners choose which opportunities warrant investment.

In iterative development, teams can use the same job context to test concepts or working prototypes. Does the intervention actually help with the difficult step? Does it create new friction elsewhere in the process? Evidence can justify a redesign—or stopping.

During launch and adoption, the customer's desired outcomes can inform candidate UX measures: early signals of progress in their work. Teams can then compare those signals with lagging business measures, such as adoption and retention, and refine both the product and what they measure.

Those aren't three disconnected research services. They're different ways of using and revising a shared understanding across the product lifecycle and across multiple teams.

AI makes that practice more attainable. I've returned to hands-on software development with AI coding tools, and the speed of moving from an idea to something I can interact with and prepare to test has changed what's practical for me.

Beyond prototypes, AI can help organize evidence and maintain links among jobs, pain points, hypotheses, and proposed capabilities. I'm actively exploring how to extend those links into the codebase, so an implementation remains connected to the customer problem it is meant to address. A structured job-process map offers a promising foundation for that work—not a substitute for customer evidence.

What are we making more productive?

Let's look again at the organization that can build twice as fast.

It might ship twice as many features. It might use the capacity to cut costs and expect fewer people to maintain the same output. Neither pressure is new.

Or it could use some of that capacity to understand customer work more deeply, test alternatives earlier, and connect decisions across teams.

Of course, more research isn't automatically better. As Cutler reminds us, outcome language and process rituals can become their own theater. A map matters only if it helps someone make a better decision—and revise that decision when the evidence changes.

In From Customer Jobs to Code, I explore what it might mean to make customer-job knowledge part of the context used to build software. Here the question is more immediate:

If your team could build twice as fast tomorrow, would you know which customer problems deserved that capacity—and what would make you change your mind?

AI can make the feature factory more productive. Making the product organization more effective is still up to us.