Getting back into the rhythm of a commute into central London got me thinking. For many first-time visitors, the Transport for London (TfL) map of the London Underground derived from Harry Beck’s original can be overwhelming.

So many transport lines in a massive urban web. But look at it a little more closely and that same map gives you exactly the information you need to get from point A to point B efficiently. It's clean, it's colour-coded, it's genuinely useful and it’s elegant. Fun fact, the map only uses 45-degree angles, creating a symmetry that's easy on the eye.

But anyone who lives in London knows the map's a lie – or at least, it's not the whole truth. It's a model of London, not an accurate map of it. The capital's physical geography is different. Stations aren't equidistant. Central London isn’t nearly that large compared to the outer zones. Not all trains call at all stations. Lines run on different frequencies. And nobody – nobody – takes the Tube between Covent Garden and Leicester Square. You'd be faster walking. You'd probably be faster crawling.

The thing is, the map works brilliantly when you understand its limitations.

Sound familiar?

It should, because it's exactly what's happening with AI in healthcare right now.

AI tools are being rolled out across the NHS and the broader health sector at pace, because they bring a similar degree of surface-level simplicity and clarity. And they're genuinely impressive. They can triage, predict, flag, summarise, and support clinical decisions in ways that would've seemed extraordinary five years ago. But here's the catch: just like the Tube map, they approximate the world. They don't accurately reflect it.

Every AI model involves a necessary simplification. There's an inherent loss of detail in how it processes information and spits out an answer. In most settings, that's a manageable trade-off and in fact trying to mirror training data too closely can bring its own problems of 'overfitting'. But in healthcare, where the stakes are clinical, regulatory, and deeply human, theses trade-offs can come back to bite you.

So what does 'reading the key' actually look like with AI?

TfL doesn't just publish the map on its own. They publish guidance on how to use it. Reading the key – all those little symbols most people ignore: the little cross which shows limited opening hours for stations, national rail connections, step-free access, riverboat services – is what separates someone who can navigate the network confidently from someone who tries to use the Waterloo and City Line at the weekend or take a Piccadilly Line train from Turnham Green before 11 o’clock at night! 

For healthcare organisations deploying AI, the equivalent is the information that suppliers make available (or should make available) with their products. For management, this allows a candid assessment of the trade-offs involved in using the AI model leading to defensible choices:

• What has the vendor actually told you about how this model works, and what haven't they told you?
• Where in your workflows is the AI making a prediction, and who owns the decision about what that means for a patient?
• Are there populations for whom the model will produce less accurate predictions?
• What's your audit trail if something goes wrong?
• Have you tested the edges of the tool, or just its best-case performance?
• How might performance change over time?

These aren't abstract legal questions. They're the difference between an AI deployment that builds trust and one that ends up in a newspaper for the wrong reasons.

Frontline staff also need to understand how the model works and its limitations to avoid automation bias. 

The stuff nobody's written down

Here's the other thing about the Tube. There's a whole layer of knowledge that experienced users carry around that nobody's formally captured anywhere. Don't take the Central Line in a heatwave. The Victoria Line is fast, with trains coming almost every minute. That interchange at Bank looks quick on the map, but it isn't. 

The same tacit knowledge exists around AI tools. People who work with them day in, day out develop an instinct for where a model tends to drift, where the outputs need a second look, where it quietly underperforms. The challenge for organisations, and their legal and compliance teams, is capturing that knowledge and sharing it with their teams before it becomes a gap in your governance through which a patient can fall ('mind the gap' sounds familiar). The margin of error in the healthcare sector is incredibly thin. 

The bottom line

The Tube map works because it's useful and because enough people understand what it isn't. AI in healthcare has exactly the same potential – and exactly the same caveat.

The organisations that'll get this right aren't the ones moving fastest. They're the ones who've taken the time to read the key and understand its limitations.

If you're thinking about how to deploy AI responsibly in a healthcare setting, whether you're a provider, a supplier, or somewhere in between, Browne Jacobson’s healthcare and life sciences team would love to have that conversation.