Category: Rich Pictures

It’s time for prompt engineers to step aside. Context engineers are taking over.

Big consultancies have stopped talking about model options and now talk about the importance of judgment, defining workflows, and making activities predictable.

The technology at the centre will do what it is told.

But that’s not enough.

To do the right things it needs more information.

Is the solution just more contextual information – policies, guidance, guidelines and guardrails – or is there still a role for humans in an AI-driven workflow?

I think there is, because the real context is rarely explicitly written down

Instead it is constantly constructed, reconstructured and negotiated by the people who are involved in a situation.

If we don’t know what people want and need we won’t know what to ask the machines to make.

Humans are needed, first to understand the situation

Then to get machines to take the right actions.

It took me years to realize that what we see is not all there is.

As an engineer I was taught that all you needed was a goal or objective – you could then build a solution to meet that goal.

That may be the case if the goal is simple – like losing weight.

But many people know that the process of losing weight is complex and problematic – we experience a fog of opinion, misunderstanding and cycles of success and failure.

Success depends on how you walk into the fog and find your way out again.

Now, the method I reach for, when trying to understand a situation, is to draw it – using rich pictures from soft systems methodology and rich notes from my own research.

Listening, talking and drawing together helps us explore the situation, peer through the fog, and understand what types of actions are possible given the constraints and limitations around us.

Navigating complexity becomes much easier if you can make unsaid and invisible factors explicit and visible.

Explore first. Then act.

Making a thing is easy. Making it useful is hard.

The big AI vendors appear to have realised this too.

Microsoft has announced its own version of Palantir’s forward deployed engineers – the Microsoft Frontier Company – saying that it will have thousands of people go out and help companies deploy AI.

The difference? Rather than selling AI as a product, firms are trying to sell outcomes and drive adoption.

It’s not easy to make the switch.

Organisations in the real world are not machines – you can’t just go in and engineer new systems and make everything work.

They are actually an intertwined mess of complexity and confusion as seen through the eyes of the people involved.

We need to engage with that confusion and have tools to explore and understand what is going on before we decide what to do.

That takes more time in the short run.

But it leads to better solutions and outcomes in the long run.

Between the creation of a strategy and a pipeline of projects is a chasm where opportunities go to die.

When we talk to managers they tell us that they have a strategy – they know their direction of travel.

They also know what needs to be done operationally. There is a pipeline of projects sat in various stages of approval.

Projects get stuck because things that makes sense to engineering don’t make sense to finance. Approval processes that meet risk requirements make it impossible to select a fully compliant solution. Procurement looks for innovation, but legal focuses on the risks and liabilities.

So decision makers push back – they ask for more data.

But what they really need is information, not data.

Data is the raw material – it needs to be collected, cleaned, processed and readied.

But for it to make sense – for it to turn into information – managers need to have the time and headspace to understand the data and figure out what it means for the organisation.

A good information system does two things: it makes it easy and quick to collect quality data; and it makes it easy for you to put what you’ve collected on the table and make sense of it.

Your data is like a pile of bricks.

Your information system is the bridge you build to go from strategy to operational delivery.

Choosing an AI today is a political decision, not just a technical one.

I asked three open weights models – Meta’s Llama 3.2, DeepSeek’s R1 and Alibaba’s Qwen – to tell me about Taiwan.

Llama provided an overview of the general position from a Western perspective. The two others responded with a version of the official position of the Chinese government.

The politics of the situation are deeply baked into the weights of the models.

They do not represent the world’s information neutrally but instead construct a very specific world model that they want you to believe.

That shouldn’t surprise any of us.

With a technology this disruptive it’s not surprising that governments will try and control how it’s used.

That’s why sovreignty is an issue at a national level.

And at a company level, we must select models and how we use them with an eye on what happens if someone decides to turn off the AI tap.

Information is power.

Who will you trust with the switch?

Humans are not obsolete. But we do need to adapt – once again.

A hundred years ago, we sold our labour.

The last fifty saw a move to services.

We saw that shift in the economy, as strong machines were succeeded by smart machines.

And now we have intelligent machines – another technological revolution.

What do humans add in this environment?

I’d argue that there are two essential skills: design and value.

Design thinking helps us create “things” that other people want.

Value creation is doing “something” that helps other people.

What’s central to being successful is not the machine you use.

It’s putting the people you help – your clients – at the centre of your work.

Human advantage in the AI era comes from design and value anchored in human needs.

AI output becomes slop only when we use it sloppily.

Most of us would regard mass-produced AI generated content that’s sent to us as slop.

But the same content becomes research if you work with an AI, iteratively talking with it and working through a problem or understanding a topic.

And if I create an initial draft and then talk through it with you, reading together and improving the draft – the AI output becomes a working draft.

Both these are useful examples where using AI increases personal and group productivity.

AI becomes slop when you try and remove the human aspect, when the output is judged on quality and completeness on its own.

In many situations AI doesn’t remove the need for people.

It’s the active participation of people that makes it valuable or not.

There’s one issue with solving problems when they are small.

We’re unprepared for big problems.

Climate change could have been a small problem – if we had acted sooner.

But now we’re going past the point where we can prevent the worst effects.

Heat. Floods. Wind. Wildfires. Drought.

These are physical risks. And the risk to us depends on where our assets are located.

The problems are getting bigger. And more complex.

So far we’ve talked about emissions.

Now we need to talk about exposure and financial impact.

I think we need to treat AI like the bear.

You know the story. Two friends are walking in a wood when they come across a bear.

One immediately reaches into her bag and starts putting on running shoes.

Her friend asks, “What are you doing? You can’t outrun a bear.”

“I don’t need to outrun the bear”, comes the reply. “I only need to outrun you.”

Today I came to the conclusion that AI is better at Excel than I am. It’s also better at programming, at writing guidance documentation, and checking if data in a system is correct.

AI has already replaced the workers I was going to hire.

Now it’s replacing the work that I was going to do.

But we all have equal access to AI.

So my only shot is to be better at using it to create value for my clients than my competition.

If you can’t outrun a bear, you really only have two choices.

Move to a place where there are no bears.

Or get really good at not being dinner.

Is anyone else finding that they’re settling into patterns with their use of AI chat models?

I haven’t done a real “search” in a while. Not the Google kind anyway.

When I search, it’s in academic databases for specific articles and authors.

I use an LLM to critique my writing and find weaknesses, but not to rewrite or suggest language.

I want to find my own words.

At work, it seems increasingly natural to ask a model to do work.

Add some calculations to a spreadsheet. Pull out action points from my notes. Create a first draft.

I’ve read of others using models as a way to process thoughts – talking to them about how they’re feeling.

I’ve started an experiment doing that as a food log instead. I talk to an LLM about the plan for the day. How many calories in this meal? Can I have this bar of chocolate or not?

It seems like this technology is now settling into some of our daily systems and routines.

I’m optimistic about the path we’re taking.

This technology may help us make better choices over time.