Are We Teaching Workflows — or Teaching People to Think? -

The plan made sense.

Structured rollout. Clear workflows. Defined use cases. A way to make GenAI usable without asking everyone to figure it out from scratch.

Al Hounsell and Michelle Fernando at Gowling WLG outline exactly this kind of approach — disciplined, thoughtful, grounded in real practice.

They even point toward something beyond workflows: the idea that effective GenAI use is learned through practice, not assumed. It's worth reading. And it surfaces a question worth sitting with.

Because legal organizations are navigating two pressures at once right now. The first is adoption — GenAI is here, the investment has been made, and progress needs to be visible. The second is risk — incidents like the recent Sullivan & Cromwell filing are a reminder that in legal practice, output that leaves the building unchecked carries real professional consequences.

Structured workflows are the instinctive answer to both. They make adoption visible and risk feel managed. But they only fully solve one of those problems. Understanding why is what this piece is about.

What workflows can't anticipate

The strongest GenAI examples emerging inside law firms share something in common.

They didn't start with a predefined path.
They came from attorneys who understood the tool well enough to reach for it in an unscripted moment. Who changed what they did first, skipped steps that no longer made sense, or created something reusable on the spot.
Who used GenAI not to complete a process — but to navigate a situation.

That kind of use doesn't come from following a workflow more closely.
It comes from developing a different kind of fluency.

When workflows create a false floor

The recent Sullivan & Cromwell filing — and the apology that followed — has been circulating inside legal circles for a reason. Not just because of the hallucinations themselves, but because of what they quietly surface.

The instinct is to ask: how did this get through? Was it the tool? The process? The review?

Those are reasonable questions. But there's one that tends to get asked less often: how were the people using the tool taught to work with it in the first place?

Structured workflows do something important — and something unintended. They make GenAI usable. They also make it feel supervised. When a process is in place, there's a natural assumption that the process is doing the protective work. That assumption is usually fine. Until the output looks right but isn't.

Hallucinations rarely announce themselves. They appear inside otherwise plausible work — often in exactly the kind of document where a reader's guard is down because everything around the error looks correct. Catching them requires something a workflow can't fully supply: a practitioner who understands how the tool reasons, where it tends to fail, and why confident-looking output sometimes isn't.

That's not a process gap. It's a fluency gap. And it's worth naming — not to slow GenAI adoption down, but to be clear about what structured rollouts do and don't build.

A different kind of value

It helps to see this distinction mapped out. When we work with legal teams on GenAI development, a pattern emerges consistently enough that we tried to capture it visually.

Not as a model. Just as a way to capture what we were seeing.

Most structured approaches tend to stay where the work is predictable. Where the path is known. Where consistency matters most. That's exactly why they work. They make GenAI usable. They reduce friction. They help more people get started.

But the examples that begin to change how work is approached tend to show up elsewhere. Where the path isn't fully defined. Where judgment plays a larger role. Where the focus moves from producing outputs to shaping outcomes.
That's where legal AI fluency starts to look different from AI adoption.

The gap between the two isn't about effort or intent. It's about what the firm has deliberately built capacity for — and what it's assumed would develop on its own.

The risk isn't failure. It's settling too early.

Settling too early is where the dynamic becomes self-reinforcing — and worth watching carefully. Structured workflows do what they're supposed to do.
They scale. They standardize. They make adoption visible.

But they also create a kind of gravity.

Once a set of workflows works well, it's natural to build around them. Expand them. Optimize them. Measure them. And over time, they can begin to define how the tool is used — not because they're limiting by design, but because they're successful.

So the risk isn't that firms get GenAI wrong.
It's that they get it right — within a narrower frame than what's possible.

Holding both

None of this is an argument against structured approaches.
Without them, most firms never get past early experimentation. The Gowling example shows what that discipline looks like in practice, and it matters.

But structure alone doesn't determine how far the firm evolves.

The firms that seem to move further tend to hold both: enough structure to make GenAI usable across the firm, and enough space for attorneys to develop their own fluency.

Workflows help people get started.

Fluency changes how the work itself is done. It's also what determines whether the next Sullivan & Cromwell moment happens on your watch — or gets caught before it leaves the room.

A practical signal to watch

When a GenAI example gets shared internally, the question is often: Which workflow was used?

It's a sensible question. It helps others repeat the result.

But there's another question, asked less often: What changed in how the work was approached?

That question is harder to answer. It doesn't scale as neatly. It doesn't fit into a predefined path. But it tends to point toward where the more interesting shifts are happening — and where the more important protections are developing.

Firms that only ask the first question learn to repeat what works. Firms that ask both learn to understand why it works — and where it might not.

Over time, that distinction shapes more than what the firm learns how to do. It shapes what the firm is actually protected from.

Adoption without fluency scales the work. Fluency is what makes that scale safe — and what determines whether the next hallucination gets caught before it leaves the room, or after.

That's not a workflow question. It's a strategy one.

And, it's the question the Edge program at O4I was built around.