Enterprise UX in the Age of AI, Part 2: The Agent Joins the Team
Part two of a series on designing the tools behind the business. This time: Design for Real Roles and Automation First.
Part one of this series made a simple argument: AI is a multiplier, and multipliers are indifferent to what they multiply. Good UX foundations get amplified. So does chaos. If you missed it, start there.
This post takes the next two principles, and they belong together because both answer the same question: who is doing the work? For 25 years my answer was always a person in a role. That answer is changing.
Principle 3: Design for Real Roles
There is no such thing as “the user” in enterprise software.
An HR platform serves an employee requesting leave, a line manager approving it, a payroll specialist reconciling it and a legal reviewer auditing it. Same system, four different jobs, four different definitions of success. A consumer app can design for one idealised person. An enterprise tool that tries the same thing serves nobody well.
Designing for real roles means the interface reflects the job. Role-based dashboards and permissions, yes, but also terminology, defaults and layout. The screen a payroll specialist lives in for six hours should be built around their queue and their next action. And roles come with environments: a call-centre agent with three monitors and a headset works nothing like a field engineer on a phone in the rain. Design for the claims assessor with forty cases in the queue, not for a persona called “busy professional”.
The AI twist
Your newest colleague has no job title.
Agents are joining enterprise systems as operators, reading queues and moving cases along. The instinct is to treat an agent as a feature. Treat it as a role instead. Like any new starter, it needs a job description: what it can see, what it can do, when it must hand over to a person. That is role design, and the discipline you built for human roles transfers almost directly.
It cuts the other way too. The generic AI assistant, one chatbot bolted onto every screen for every user, fails the role test on day one. The summary a payroll specialist needs is not the summary a legal reviewer needs, even of the same record. Role context should shape what the model is shown, what it is allowed to say, and what it is allowed to touch.
And permissions get sharper, not looser. An agent inheriting a person’s full access is a role violation waiting for an audit. Least privilege applies to software colleagues too, and unlike human colleagues, they will act at 3am, at scale, without getting suspicious that something feels off.
Principle 4: Automation First
If a call-centre rep copies the same ID into three different fields, that is not their inefficiency. It is your design failure.
Enterprise work is full of these micro-repetitions, and automation-first design goes hunting for them: smart defaults, pre-filled fields, remembered preferences, workflow triggers. Anything a system can predict, a system should do. The aim is more judgement per hour, letting people spend their expertise on the decisions that need a human, and none of it on data wrangling.
Classic automation earned trust by being boring. Deterministic, same input, same output, tested once and trusted for years. You designed it, verified it and forgot about it.
The AI twist
Automation is becoming delegation, and delegation is a different design problem.
A workflow trigger follows steps. An agent pursues a goal. When you hand a goal to something that decides its own steps, the questions change: How do I brief it? How do I see what it did? How do I take back control halfway through? These are management questions, and the interface has to answer them. The hand-off is the new interface.
The pattern I keep coming back to is a ladder of autonomy. An agent starts by suggesting, then drafting for approval, then acting with a confirmation step, then acting and reporting back. Each rung is earned with evidence from the one below it. Nobody sensible starts a new hire on unsupervised sign-off authority. The same applies here.
Part one argued that verification is the sneakiest cost of AI. Delegation is where that bill arrives. A review step that takes longer than doing the task yourself means the delegation has failed, however impressive the agent. Design the review to be fast: show what changed, show why, link to the source, make approval one action and rollback another. Respect the task still applies. Now it includes respecting the task of supervision.
The principle underneath is unchanged: automate the repetitive, keep the judgement human. What has changed is that supervising the automation is itself becoming the judgement.
Where This Series Goes Next
The next post takes Insights Over Information and Reliable at Scale: what AI-generated insight does to trust in the numbers, and why model latency is the new page load.
These two principles share one conclusion. The org chart of your software is changing. Some of the roles you design for now will not be people, and the best enterprise UX teams are already writing job descriptions for them.
Patrick Mooney is a senior design leader based in Ireland, with more than 25 years’ experience in UX, product and enterprise software, and the founder of Dublin UX. He writes at patrickmooney.me..