AI and trust: principles for automating responsibly

23 October 2025

We’ve supported a lot of teams who have automated part of their core work. The boom in AI-backed tooling has accelerated this: both for internal tools (checking documents, ingesting data) and for commercial services too. Whether people are working, shopping or seeking support, machines are starting to do more and more things people once did for themselves.

But automating work means paying acute attention to trust. There are few user experiences quite as frustrating as the ‘computer says no’ chatbot. Do users (customers and staff) have faith in the quality, accuracy and safety of the services you provide?

If you’re leading this kind of change, you'll get a long way if the foundations are solid. On top of that, a few more pointed principles will set the tone for the work you want to do:

1. Start by automating work people find boring

You don’t need magic. The dull day-to-day stuff is a rich seam.

Large organisations accrue all sorts of cruft. Teams that re-key data from one service to another, or batch-process applications before sifting them. This stuff is ripe for automation. There will be clear timesinks you can prioritise.

Why start here? First off, it’s where the errors are at. You might even have data here that can inform your risk appetite. You’ll probably get better-than-human performance quickly, proving the value of the change. That’s especially important in a regulated environment. And you’ll learn a lot about the nuances of AI and its relationship to your specific context.

2. Respect your risk

It should be obvious, but let’s state it anyway: you will not build an infallible machine. No-one has.

Your motivations for automating a service might feel unique, but they’re often a blend of:

wanting to scale without hiring more people
wanting to shrink the size of a team
wanting to refocus your team on other work

Your specific blend of these will define your tolerance for risk. How certain does your document sorting algorithm need to be that a paper file is of Type A or Type B? At what point do you want to put exceptions in front of people?

Being honest about your risk appetite means you can build appropriate guardrails. Your teams will need to develop parallel services that evaluate the performance of LLMs, before you “go live” and while the service is being used. Those tests will need to be robust enough to validate the performance of new models, updated prompts and changing inputs. Without them, you’re flying blind. (PD network member Dave Guarino has written about doing this).

Where you set your tolerances, how you establish failsafes and fallbacks, those follow directly on from the purpose. They set the tone for how you establish trust in your use of AI.

3. Break problems into pieces

Replacing complex services with a black box that “fixes” multiple problems is a recipe for disaster: just ask anyone who’s trying to untangle themselves from an ERP instance.

Break the process into small chunks. Build solutions that solve discrete problems. Repeat.

To give an example from my old life (running the product team at a legal tech start-up) this meant processing documents in stages. We separated the functions that ingested data from those that acted on it. Bundling those activities would have made it harder to analyse how they performed and update them.

The time will come when AI tools can be pointed at an outcome and reliably replace systemic processes. That time might be soon. But for now ‘small pieces, loosely joined’ is the order of the day.

4. Build with users

The people who do these jobs now should be part of the team automating the work. They’ll know inside and out what the traps and bugbears are. They will make it clear where there’s more value to be found. They will show you where they have to step in and give the rest of your systems a shove.

They’re also the people who should be part of ongoing validation of the output of your automated services. Whether it’s spot-checks or plain old troubleshooting, get them to draw the map: they know the territory.

At Public Digital we support businesses and governments as they use AI to transform. Find out more about the work we’re doing, or get in touch to talk about the challenges you’re facing.

A human-centred approach to Ai

At Public Digital, we take a human-centred approach to technology. We use the best practices of the internet era to steer bold, long-lasting change in organisations.