Graceful degradation is the only AI strategy that survives contact with regulators

When I joined BT, the order-management system was running on roughly two thousand hand-written rules. Every edge case in a decade of broadband, mobile and TV orders was a rule. Some of them contradicted each other. A few of them were written by people who had left the company years earlier.

The temptation was to replace the whole thing with a model - give it enough historical data, let it learn the patterns, and throw the rules away. We didn't do that. Or rather, we did do it, but only for the orders where the model was confident. Below a threshold, we fell back to the rules.

Confidence is a design primitive

In regulated consumer infrastructure, you don't get to be clever at the expense of predictability. The interesting question isn't whether your model is more accurate than the rules on average - it's whether your model, plus your fallback, plus your monitoring, is more accurate than the rules in the worst case.

So we wrote the AI layer to be honest about confidence. When it scored an order and came back with a low-confidence result, we didn't pretend. We routed to the deterministic path and logged it. Over time, the low-confidence pile became a backlog for the team: each entry was a signal that either the model needed retraining, or the rules needed rewriting, or there was something genuinely new in the input distribution.

Graceful degradation is not a safety feature. It is the product.

What we actually shipped

Order fallout dropped by about 18%. Manual interventions fell by around 30%. The cut-over happened without a customer-visible disruption, which in a live consumer estate with 20 million customers is the only metric that matters.

None of that is because the model was especially clever. It's because the system as a whole had a coherent theory of what "wrong" meant - and what to do about it.

The general lesson

Every team I've worked with that tried to launch an AI feature "and then figure out the edge cases later" has ended up either rolling it back, or quietly absorbing a lot of invisible risk. The teams that ship durable AI products design the fallback first. Then they train the model into the space the fallback leaves for it.

If you're a leader pushing your organisation toward AI-augmented systems, start with the question of graceful degradation. Everything downstream is easier.