Design human–AI workflows

“Adopting AI” often starts with tools: buy licences, run a training session and hope people find their own uses. A few months later the keen users are very good at it, everyone else is back to the old way, and the process itself has not changed.

For AI to enter real work, start by taking the process apart: what are the steps, what information does each need, which parts are repetitive and which need judgement and accountability? Once that is clear, the places where AI helps become obvious.

You may be seeing this

• The team spends a lot of time sorting, summarizing and reformatting.
• Everyone uses AI differently, and quality varies widely.
• Managers worry about AI mistakes and are unsure who should check.
• Automation was tried, but it breaks on exceptions and work goes back to manual.

How we approach it

• Pick one process: start with work that is frequent, rule-based and low-risk.
• Break it down: map the current way of working, with the input, output and time of each step.
• Divide the work: mark each step “AI drafts”, “person checks” or “person decides”, and write down the criteria.
• Set checkpoints: add checks at key steps, and hand back to a person when the AI is unsure or meets an exception.
• Pilot and adjust: run it at small scale, record why things were changed or sent back, then adjust.

What you get

• Process maps: before and after.
• A human–AI division of work: who does each step, and on what basis.
• Prompts and templates, so the same task is done the same way.
• Checkpoints and rules for exceptions.
• Pilot records and recommendations for next steps.

How to tell it works

Compare before and after the pilot: time spent on the same work, how often outputs are sent back, how consistent they are. Also watch whether the team keeps using it; that often says more than any single number.

Be clear about who does what first; then AI can really help.