AI's jagged frontier

Ask most people whether AI is good or bad at its job and you'll get one of two answers.

Either it's extraordinary, the dawn of a new era, doing things no human could manage at speed and scale. Or it's nonsense: it hallucinates facts, gets names wrong, and produces text that sounds vaguely right but is fundamentally hollow.

Both are right and that's the problem. It's not as smart as you think nor as dumb as you've heard.

AI doesn't improve uniformly across everything it does. Its capability looks less like a steadily rising tide and more like a jagged outline. Extraordinary peaks in some areas, embarrassing troughs in others. Understanding that shape is arguably the most practical thing you can do with AI right now.

Take writing. Ask AI to draft a clear, professional explanation of cremation for a family who have never arranged a funeral, and it will produce something well-structured, in an appropriate tone that provides a genuinely useful starting point. It's excellent at this. It works from patterns, language, tone and structure. Give it the right prompt and the right context, and it saves time.

Now ask it to count the number of letters in the word strawberry. Until recently it would get this wrong – it was not a quirk but a structural limitation of how these systems process information and, in the wrong setting, these kinds of errors could be catastrophic.

Perhaps this is why Microsoft recently updated its terms of service for the consumer version of Copilot to state that it is for entertainment purposes only.

The term jagged frontier was coined by researchers at Harvard who discovered that AI outperformed human consultants on some tasks by a significant margin and fell apart on others in ways that were hard to predict. Crucially, the tasks AI failed at weren't always the complex ones.

It's not as smart as you think nor as dumb as you've heard.

For funeral professionals, this means spending time finding what AI is good at and focussing on that.

Repetitive written communication is a good place to start. Acknowledgement letters, standard information packs, responses to frequently asked questions from families – tasks that draw on consistent language, a defined tone and a fairly predictable structure.

Once you've shaped the instructions carefully and checked the output, the same template can be handed to other members of your team and the time saving quickly mounts up.

Where the frontier dips sharply is where facts, accuracy and emotional nuance are required without a human in the loop. AI doesn't know your firm's specific practices, your pricing, your local authority's rules or the name of the family you're dealing with. And unless you tell it, it will fill in those gaps – with potentially embarrassing results.

Best to treat AI like a capable but overconfident colleague: brilliant at the first draft but unreliable without supervision.

Maybe this is why a recent Gartner report jokingly proclaimed that employees should be banned from using AI on a Friday afternoon – a time where perhaps they are more likely to be checked out and therefore less likely to check the outputs.

The firms getting the most from AI right now aren't the ones that deployed it everywhere at once. They are the ones that found two or three things it's genuinely good at, made those part of the everyday and checked the output every time.