Why human-centred design still decides whether AI work pays off

A bank adds a drafting assistant to a credit team and leaves the old checklist untouched. Staff paste text into the tool, paste it back into the form, and still chase the same approvals. The licence count rises. Cycle time barely moves.

Deloitte has put numbers on that stall. Of the more than 9,000 leaders it surveyed, only 14 percent say they're good at shaping how people and AI work together. Most organisations (59 percent) layer AI onto their old systems and processes. Organisations that redesign roles and the way people and machines work together are twice as likely to beat their return expectations for AI. Sixty-six percent of leaders say this kind of design matters, yet only 6 percent say they lead on it.

Designing people and AI together

Organisations already design how people work with people and, more and more, how machines hand work to machines. What's missing is designing for both together. That's what human-centred design looks like now. You start from the people who do the work and the outcome you want for them. Then you place the model where it helps, and keep a person wherever judgment, escalation or care still matter.

A European telco shows the gap. It added an AI "expert" to customer service without changing roles or workflow and got a 5 percent productivity lift. When most of the rollout budget went into redesigning how staff and the AI worked together, with new workflows, trust thresholds, escalation paths and training, productivity rose 30 percent as staff learned to work with the tool.

Access to AI is spreading fast. The skill of placing it inside real work isn't keeping up.

The organisations that lead on this kind of design are nearly 2.5 times more likely to report better financial results, and twice as likely to say they provide meaningful work. Those are big gaps, and you won't see them on a usage chart.

Hardwiring and softwiring

The work splits into what Deloitte calls hardwiring and softwiring. Hardwiring covers redesigned roles, decision rights and the escalation rules that say when work moves from AI to a person. Softwiring covers leadership behaviour, culture and the psychological safety that lets people question the tool, escalate, experiment and learn. Skip either one and the install looks finished while the work stays stuck.

I see the same gap in innovation programmes that treat AI as a lab demo. A demo shows a model can draft or route work. Whether a team will change how it hands work between people is a separate question. Human-centred design forces that question early, with the staff who'll live with the change in the room.

At the organisation level, design sets the ambition, the cross-functional governance and the principles that guide teams. At the role level, it picks how a person works with the model: as a coach, a thought partner, a muse or something the person supervises. Get that choice wrong and the damage is quiet. People can end up with only the hard cases, or feel micromanaged by a tool that was meant to help.

What a programme should insist on

1. Name the human and business outcomes before you pick the model.
2. Redesign the step where a person and the model meet, including who can override whom.
3. Get the people who do the work to help choose how they'll work with the model, whether it drafts, coaches, routes or acts.
4. After four weeks, measure the step you changed. Licence usage alone won't tell you much.

MetLife, for one, uses AI as a coach for call centre staff during emotionally charged calls. The person stays on the call and the model supports them. That's a design choice, and it should match the work and the worker.

A practical agenda for the redesign session

When I run this kind of session, I keep the room small and the problem live. That means eight to twelve people who touch the workflow and a sponsor with the authority to change a handoff. The facilitator's job is to stop the conversation drifting into a tool bake-off.

We spend the first hour on the current path as it actually runs, including the workarounds people already use. The second hour goes on placing the model in that path, with clear trust thresholds and an override rule in plain language. The last hour goes on the measure the team will check in four weeks, and on who owns the next experiment.

This is service design. The model is a new actor in a system of people, forms, queues and decisions. The return comes from designing the point where people meet it.

Where Collective Campus sits in this

Collective Campus has run innovation programmes with large organisations since 2015. Our AI capability work with teams started in the generative AI era, from 2022. Faster tools make sloppy process design more expensive, because now you can automate a bad handoff at scale.

If your AI programme still starts with a tool shortlist and ends with a usage dashboard, flip the order. Start with the people and the step you want to change. Our work on design thinking and AI design sprints is built to put that order back, with a live challenge and a decision the team can defend.

What to do this quarter

Pick one workflow where AI already touches the work. Map who acts, who checks and who's accountable when the output is wrong. Redesign that meeting point with the team, then check whether the step got faster or safer. Ask the people who do the work what the tool should draft and what it should never decide. Write their answers into the process before you buy another seat.

If you can't point to a changed step after four weeks, there's no return yet. Bring the design work forward, and treat the model as one actor in a service you're still redesigning with the people who run it.