Why AI investments need UX to pay off
AI pays off when people trust it and fold it into their daily work. That depends on UX: showing what the AI did and why, letting users correct or override it, fitting it into existing tools and measuring outcomes. Investing in AI without investing in UX usually produces impressive demos that few people use.
The adoption gap
Companies are spending heavily on AI, but many projects stall after the pilot. The common reason is not the model; it is that the people meant to use it do not trust it, do not understand it or cannot fit it into how they work. That is a design problem, and it is why our Agentic AI & Automation work always includes UX.
Why AI projects stall after the pilot
Many organisations follow the same pattern with AI. A pilot is launched with enthusiasm, early demos impress leadership, and a small group of users tries the tool. Then usage plateaus. People return to their old ways of working, the tool is quietly used by a handful of enthusiasts, and the business case that justified the investment never materialises. When teams investigate, the cause is rarely the model itself. It is usually that the AI does not fit how people work, its outputs are hard to trust or check, or nobody designed the experience around real tasks.
| Reason for low adoption | What users experience | Design response |
|---|---|---|
| Unclear value | Not sure when or why to use it | Show a clear first win for a real task |
| Low trust | Cannot tell if the answer is right | Show sources and confidence |
| Poor fit | Have to leave their usual tools | Embed AI in existing workflows |
| No control | Cannot correct or undo results | Editable outputs and easy undo |
| Inconsistent quality | Good answers one day, poor the next | Clear limits and feedback loops |
Five UX principles for AI products
- Transparency: Show what the AI did, the sources it used and how confident it is.
- Control: Let users edit, approve, undo or switch the AI off.
- Fit: Embed AI in existing tools such as CRM, ERP and email instead of a new app.
- Feedback: Make it easy to flag wrong answers and see that feedback is used.
- Clear limits: Say what the AI cannot do, so users do not over-rely on it.
Design patterns for trustworthy AI
Good AI experiences share a few recurring patterns. They explain what the AI is doing in plain language, cite the documents or records behind an answer, and show when the AI is uncertain. They keep the human in charge of important decisions by presenting AI output as a suggestion that can be accepted, edited or rejected. And they make it easy to give feedback, so the system and the team behind it can improve over time.
- Source citations: Link each answer to the records it came from.
- Confidence cues: Signal when results may be unreliable.
- Suggestion, not decision: Let users approve or edit before acting.
- Undo and history: Make every AI action reversible.
- Feedback buttons: Collect corrections in the moment.
- Clear onboarding: Explain what the AI can and cannot do.
AI with and without UX
| AI without UX | AI with UX | |
|---|---|---|
| First use | Confusing, unclear value | Guided, with a clear first win |
| Trust | Black box | Sources and reasoning shown |
| Errors | Users give up | Users correct and continue |
| Adoption | Pilot only | Part of daily work |
Designing around real workflows
The most successful AI features appear exactly where people already work. Instead of asking a sales team to open a separate assistant, an AI summary can appear at the top of each account in the CRM. Instead of a new reporting tool, an AI explanation can sit beside the existing dashboard chart. Mapping the current workflow first, then deciding where AI should appear and what it should produce, avoids building impressive tools that nobody opens. Our post on AI in SaaS products shows how this works in software products.
How to get started
- Pick one painful task: Choose a frequent, measurable task where AI can save real time.
- Map the workflow: Understand where users work today and where AI should appear.
- Prototype and test: Test the AI experience with real users before scaling the model.
- Measure and expand: Track time saved and accuracy, then move to the next task.
Clear dashboards and reliable APIs are often what make AI usable inside existing systems. For a broader plan, our Technology Advisory team helps decide where AI should go first.
Measuring return on AI investment
Measure AI by outcomes rather than activity. Useful measures include time saved per task, reduction in errors or rework, faster response times for customers, adoption among the people the tool was built for and the share of AI suggestions accepted without major edits. Compare these with a baseline taken before launch. If adoption is low, research with users usually reveals whether the problem is trust, fit or value, and each points to a different design fix.
- Time saved per task, measured against a baseline
- Adoption among target users each week
- Acceptance rate of AI suggestions
- Error or correction rates
- Customer-facing effects such as response times
Frequently asked questions
Why do AI projects fail?
Common reasons are low user adoption, poor data, unclear business goals and AI that does not fit existing workflows. UX addresses adoption and fit directly.
How do you measure AI ROI?
Track time saved per task, error rates, task completion and adoption over time, compared with a baseline before the AI was introduced.
What is human-in-the-loop design?
A design approach where people review, approve or correct AI output, keeping humans in control of important decisions.
Who should design AI features?
A mixed team: product designers and researchers to shape the experience, engineers and data specialists to build it, and domain experts to judge output quality.