Part 4: Helping People Build The Skills They Need To Leverage AI In Their Work

People can only use AI effectively if they can define their work, and that's a skill they build through practice. In this part of the series, I explain what this practical workshop needs to achieve, and the learning moments that make it work.

This is the fourth part of a series on building workflow capability for enterprise AI, the previous three parts being:

In Part 2, we saw that the way to build this capability is through an experience: an activity that lets people experience the skills gap first hand with a safe place to practise.

In Part 3, we looked at the unhelpful, subconscious assumptions that stop people building this skill. How do we design an experience that helps dismantle them?

In this article, we'll unpack what that experience needs to achieve and the learning moments that make it work.

What does the workshop need to achieve?

The workshop has two jobs:

1. Dismantle assumptions

Each of the assumptions from Part 3 stop people moving on to the next one - for example, "I don't need help defining my work". The workshop's job is to dismantle those assumptions one at a time, until people reach complete transformation - someone willing and motivated to master the skills needed to define their workflows, because they understand this is necessary to effectively leverage AI.

Not everyone will get there, and that's okay - some will require solutions to be built for them. But to build AI solutions, this is not negotiable (and one might suggest that to stay ahead of the curve when it comes to the tech, it’s probably a sensible career move to be the one building solutions).

2. Give people a safe place to practise

The two jobs feed each other. When people have a go at breaking down a workflow, they discover it's harder than they thought and that they need help. This realisation is what creates the mindset shift.

Then we give them frameworks and guidance, and let them practise on one of their own workflows.

What the workshop isn't doing is building solutions - that will comes later. Instead, people leave with a cleanly broken-down, best-practice workflow they can evaluate.

And once they've pinned down which parts are most difficult, the solution they need usually becomes clear. In fact, it’s at this point they might even decide they don't need AI at all, and that fixing the process is enough.

To what standard does a workflow need to be defined?

Before we start designing activities, it’s helpful to understand what we’re trying to achieve here. How immaculate do our workflows need to be before AI can help?

Well, if we want AI to help with our work, it needs to know how we do that work. That means getting it out of our heads and defining it clearly enough for AI to follow. So the obvious question is: to what standard does a workflow need to be defined? How do we know when the definition is good enough?

It's a difficult question, because we're generally not too good at marking our own homework. We already know the work, so what we've written always looks sufficient to us. And unless someone (or something) actually tries to follow it, we struggle to see what's missing.

So we need a lens - a perspective that tells us whether we're in the right ballpark.

The lens we'll use throughout this playbook is ‘Eager Eddie’.

Eager Eddie is capable and enthusiastic, but he's never done this work before. He has no experience, and no context about how things are done. Give him a definition and he'll follow it faithfully - but he can't fill in the gaps, because he doesn't know what's missing.

And that's exactly what AI is like - capable and enthusiastic, but has never done the task before, and has none of our context or tacit knowledge. It can only work with what we give it.

So, whenever we picture Eager Eddie following our workflow, we're really picturing AI following it.

A workflow is defined to the standard if Eager Eddie could follow it from start to finish without us standing next to him to explain. That means the steps are small enough, he has everything he needs to know, and he'd know what to do if something went wrong.

It isn't scientific. But it's a really helpful frame of reference, because it gives us someone concrete to picture instead of an abstract standard.

Activity 1: Define how to make a lasagne

The first activity is where people feel the gap for themselves, and it only works if they do it themselves. People who sit back and watch won't build this skill. You can't learn to drive by watching someone else drive - you have to sit in the driver's seat.

So we start with an everyday workflow that everyone has some familiarity with: making a lasagne.

Here's the scenario. Eager Eddie has never cooked lasagne before. He has a regular kitchen and he knows how to use basic tools. The goal is to define a workflow with enough guidance for Eager Eddie to make a lasagne without us there to help.

The one rule: no AI.

That might sound odd in world in which everyone has AI, but this gets to the heart of what we’re doing here. We’re trying to avoid outsourcing our thinking to AI and defining what we - the human - is doing when we execute this workflow. If we ask AI to tell us how we do our workflow, it fills in the gaps for us - and does a pretty average job of it, because it doesn't have our experience, our tacit knowledge or our judgement. And we'd never find out how much we couldn't define ourselves.

The skill we're building is getting that knowledge out of our own heads, so people have to do the thinking.

So, in the context of the activity, Eager Eddie stands in for AI, so everything people write is what AI would have to work with.

The learning moment

People tend to start this activity thinking it'll be easy. Most people could have a stab at making a lasagne, couldn’t they?

But as they try to define the steps, they start to discover things. They stay too high level. They merge steps. They make assumptions about what Eager Eddie already knows - which tools to use, how to measure, what "done" looks like. Even a recipe makes these assumptions, because it expects the person following it to understand a lot already.

That's the first mindset shift, and it's one we saw in Part 1: doing the work and defining the work are two different skills. The difference is that now people have felt it themselves, rather than being told.

Why does our first attempt fall short?

Once people have had a go, we look at what they've written through Eager Eddie's eyes. And almost every time, the first attempt falls short.

And there are usually three reasons for this, each one being a gap AI would fill with its best guess - thus creating the likelihood of a substandard outcome.

1. It's at the wrong altitude

Most people start at too high a level, like hosting a dinner party. But look closer and that's lots of smaller workflows rolled into one. Making the lasagne is one of them.

In fact, if you’re making a real lasagne, making the béchamel sauce from scratch might be the trickiest part - which would probably be a workflow in its own right.

So, someone might write ten steps for making a lasagne and feel they've covered it. But ask Eager Eddie to follow the "make the béchamel sauce" step and he's stuck. Each of those ten steps is a workflow in itself, with around fifteen steps of its own. That's roughly 150 steps in total!!

Now, without evaluating each workflow individually, there's no scientific way to know how far to break each workflow down. But we can use a quick rule of thumb: could one person do this workflow alone, within a few hours? If not, it's too broad. We're probably looking at several workflows, and we need to pick one and break it down further.

That also solves a problem. 150 steps is far too many to define in depth. So, we pick the part that matters most - the part that's most important, easiest to get wrong, or easiest for AI to help with - and get more granular there. For our lasagne, that might be the béchamel sauce.

2. The steps aren't discrete

The second reason people struggle is because they bundle several steps into one. "Make the béchamel and layer the lasagne" reads like a single step, but it's two different pieces of work. The person writing it knows where one ends and the other begins. But Eddie doesn't.

Each step should be one thing Eddie can do, and know he's done, before moving on to the next.

3. There isn't enough detail

Even when the steps are the right size and separate, they usually don't say enough. We can sort what's missing into four questions:

  • Instruction: Is the action clear and specific enough that someone would know exactly what to do next without you there to fill in the gaps? For the béchamel sauce, that might be: "Add the milk a little at a time, whisking constantly, and wait until it's smooth before adding more."

  • Context: Is there anything someone needs to know to understand, complete or judge this step correctly? For example: "The sauce is ready when it coats the back of a spoon and a line drawn through it holds."

  • Guardrails: Is there an exception or unusual situation that could happen during this step that would require them to do something differently? For example: "If lumps form, take the pan off the heat and whisk hard, or pass the sauce through a sieve."

  • Templates, diagrams and blueprints: Would an example, diagram or reference make it easier to understand what good looks like? Would a template help? For example: a photo of the right sauce thickness, or a diagram of the layers in order.

Context and guardrails are the two people most often leave out. They're so ingrained that we don't realise they're missing. And it's the same for AI - the more of this we give it, the better the result.

These four questions are the simplest version of a rubric we'll cover fully in Part 5: The Easy Way to Define a Workflow to Leverage AI.

Activity 2: Define one of your own workflows

The lasagne gives people a safe place to fail. But the real goal isn't to get better at explaining how to make a lasagne! It's to define their own work.

So, in the second activity, each person chooses a workflow they actually do and has a go at defining it, using what they've just learned: the right altitude, discrete steps and the four reflective questions.

Then comes the important part: a second person to interrogate it. Working in pairs, the role of the second person is to look at the workflow through Eddie's eyes - capable and enthusiastic, but new to the work, with no experience and no context - and challenge anything he couldn't follow.

This is where two more mindset shifts land.

The first is that defining work is a skill that needs practice. Knowing the gap exists isn't the same as being able to close it. Even with a better method, people still miss things, and it takes repetition before the thinking becomes natural.

The second is that AI can't fill the gaps for us. When someone interrogates our workflow, they ask what we meant, and we find things we'd never have thought to write down - things that only exist in our heads. But AI wouldn’t ask those things - it would make its best guess and carry on.

So, if we want AI to do the work well, those questions have to be answered in the workflow first.

What do people take away?

Once both activities are done, the group steps back and reflects. How did the first activity change the way they approached the second? Where did the person looking through Eddie's eyes challenge them, and what blind spots did that expose? And what became clearer about their own workflow by breaking it down?

From there, they can look at their workflow and consider where could AI help.

And often, seeing their clearly defined workflow steps is enough to start to figure out what the solutions could be.

Over the next few parts, I'll share what this looks like in practice:

  • how to define a workflow in enough detail for a person or AI to execute it,

  • how to build tools that provide personalised recommendations on how to leverage AI for their specific use case, and

  • how to help people turn a defined workflow into a working AI solution.

Next article in the series: Part 5: The Easy Way to Define a Workflow to Leverage AI.

And if you're an L&D professional who wants to master workflow improvement and become the person responsible for unlocking the value of AI in your organisation, join my email list today.

Previous
Previous

Part 3: The Mindset Shift Needed To Build AI Workflow Capability

Next
Next

Part 5: The Easy Way to Define a Workflow to Leverage AI