From automation to co-creation

Before reaching for AI, it's worth being honest about where it genuinely helps and where it doesn't.

AI is well-suited to:

• Repeated, bounded tasks that a human could do in a few seconds
• Processing large volumes of text at scale — thousands or millions of documents
• Generating images, code, and text (a capability that has expanded dramatically in the last few years)

AI is not reliable for:

• Legal and regulatory interpretation — Stanford's Human-Centred AI Institute found that one in six benchmark legal queries returns an incorrect answer
• Detecting its own mistakes — if the model was confident enough to give you a wrong answer, it's not going to flag it as wrong
• Situations where a user needs to know why an answer is what it is

This matters because the failure mode isn't just that AI gets something wrong. It's that users start treating AI outputs as a source of truth. A customer once came to us after asking ChatGPT to calculate their maximum legal rental price, then querying why our answer differed. They'd taken the ChatGPT figure as the baseline.

We removed AI-generated legal Q&A from our product entirely. Not because AI couldn't approximate the answers, but because when it got it wrong — and it did — it undermined trust in everything else we built. A product that's right 83% of the time on legal questions isn't a legal product. It's a liability.

Once you set aside the technology and ask what people are genuinely trying to achieve, three things emerge consistently:

1. Save time — eliminate manual form-filling, automate reviews, reduce the effort of checking things
2. Avoid mistakes — reduce manual data entry, ensure rules get applied consistently, flag issues before they become problems
3. Improve — understand what can be done better, spot issues earlier, act on them

These are solvable problems. AI helps with all three. But the design of how it helps is what determines whether people trust it and use it.

1. Document data extraction

The core workflow: upload a document, extract structured data from it automatically. For a lease contract, that might mean pulling the address, tenant details, rent, contract start date, and any specific clauses.

The UX pattern that makes this trustworthy is simple: show the source. When a user hovers over an extracted field, highlight the corresponding text in the original document. They can see in two seconds whether the AI read the contract correctly. This transforms a black-box output into a verifiable one.

Users who catch an error can correct it. That correction feeds back into the model through reinforcement learning — improving accuracy over time without requiring manual retraining. The human stays in the loop not as a bottleneck, but as a quality signal.

2. Automated compliance checking

Hundreds of rules and regulations apply to property management. Energy certificates expire. Rent indices update. Legal thresholds shift. No one can track all of it manually across a large portfolio.

We apply these rules automatically to the data we have about each property and surface alerts when something needs attention — not a raw list of every regulation, but a plain-language explanation: "Energy labels issued before 2021 without an energy index are invalid from 2025. This property is affected."

The AI does the matching. The UX makes the output actionable. Someone reading that alert knows exactly what to do next.

3. Improvement recommendations

For property owners, this means knowing what changes would raise an energy rating, reduce running costs, or increase rental value. The AI identifies the relevant factors. The design presents them as prioritised, specific actions rather than a dump of scores and indices.

Two examples from healthcare illustrate why explainability matters.

Mental health crisis prediction. We built a risk prediction tool with the NHS to identify patients most likely to experience a mental health crisis in the next two weeks. Clinicians received a list of high-risk patients. Some ignored it. One patient on the list had a crisis the following day and came in by ambulance. The nurse said: "If I'd known why they were on the list, I would have acted differently."

The fix was straightforward. Instead of presenting the raw output, we surfaced the four or five data points with the highest weighting in the model: the time since last contact, the number of recent unplanned interactions, the history of previous crises. The full 178-feature model was still available for anyone who wanted it. But the default view gave clinicians enough to make a judgement call.

Radiology. An AI that flags anomalies in scans isn't replacing radiologists. It's making them faster and more accurate. The value is in highlighting what might otherwise be missed — a patch that has grown, a pattern that matches known risk factors — with enough context that the clinician can assess it properly. Circling the area on the scan and noting why it was flagged takes ten seconds to read and potentially changes an outcome.

House price prediction. A predicted value of €432,000 looks precise. But a prediction is not a fact. Showing an error bar — "this property is likely priced between €370,000 and €490,000" — is less satisfying but more honest. Users who understand uncertainty make better decisions than users who treat a single number as ground truth.

The pattern across all three: AI produces an output. UX determines whether the person using it can evaluate, trust, and act on that output.

Design for systems, not screens. AI-powered products operate across data pipelines, model outputs, feedback loops, and user interventions. The design work isn't just the interface — it's how all of those pieces connect.

Design for mistakes. AI will get things wrong. Design that assumes otherwise creates brittle products that lose user trust the first time something fails. Build correction mechanisms in from the start.

Design for explainability. If the output of an AI model can't be explained to the person using it in terms they can evaluate, the product has a trust problem waiting to happen. Show sources. Surface the reasoning. Give users enough context to know whether to act.

Design for trust. Transparency about what the system does — and what it doesn't do — is not a weakness. It's what makes people willing to rely on it over time.

Saving time is a good product outcome. But the more interesting question is: what is the task that sits in the corner of someone's mind, that they keep putting off, that stresses them out just thinking about it?

Those are the tasks worth designing for. Not because eliminating them saves minutes on a timesheet, but because removing them changes how someone feels about their work.

That's what AI-powered products can do, when the design is right.