AI investment creates value when it improves a complete piece of work—not just the speed of one task. Before buying more tools or assigning more training, inspect the bottleneck, review effort, quality and what happens to any time released.
Three things to take with you.
- Measure an outcome, not the number of prompts or licences.
- Include checking, corrections and handoffs in the work.
- Treat a pilot as a local test, not proof for every role.
Adoption is a starting point, not the outcome
A team can use an AI assistant every day while customers still wait just as long for an answer. Drafting may be faster, but a queue for approval, unclear ownership or missing information can leave the end-to-end process unchanged. More activity is not the same thing as more useful work.
Start with a specific unit of work: a resolved customer request, an approved proposal or a completed internal brief. Ask what acceptable quality looks like and where the work waits. If nobody owns the complete journey, local efficiency can disappear between teams.
What the research supports—and what it does not
Brynjolfsson, Li and Raymond’s Generative AI at Work study examines AI assistance in customer support. It reports productivity improvements that differ across workers. That supports a narrower conclusion than “AI makes every team more productive”: the task, implementation and worker context matter.
The study is not a forecast for your HR function or organisation. SOA’s practical interpretation is to run a bounded workflow test before treating tool adoption as a business result.
Source context: Brynjolfsson, Li & Raymond — Generative AI at Work
Look for the constraint before prescribing the solution
| What you observe | Question to investigate | A useful next test |
|---|---|---|
| Drafts are quicker; completion is not | Where is work waiting? | Map approval and handoff time. |
| More output; more corrections | Has checking become the bottleneck? | Count review time and rework. |
| Some people benefit; others do not | Are tasks and starting skills comparable? | Compare similar work, not just averages. |
| Time is saved; value is unclear | What happens to the released capacity? | Agree how that time will be used. |
Design a small test with an honest baseline
Select a routine, low-risk workflow with an accountable owner and approved tools. Record a baseline before introducing the change. Define the quality threshold in advance so that faster but less reliable output is not counted as success.
During the test, keep the task mix visible. Record time spent creating, checking, correcting and handing off work. Ask employees where the tool adds friction; do not turn a learning experiment into individual performance surveillance. A small pilot can identify promising patterns, but it does not establish that the same result will hold everywhere.
- Outcome: what completed work should improve?
- Guardrail: what quality or safety condition must not worsen?
- Owner: who decides whether the output is usable?
- Review: when will we stop, adapt or expand the test?
Bring the right question to the HR conversation
Instead of asking why employees are not using AI enough, ask where work is getting stuck and what evidence would change your view. A tool issue, a capability issue and an operating-model issue require different responses.
If the constraint is practical manager capability, involve the learning team. If it is access, data or approval design, involve the relevant process and technology owners. HR can help connect those perspectives without becoming the sole owner of every AI outcome.
A little more clarity.
Does low productivity mean employees need more AI training?
Not necessarily. Check workflow design, access, quality requirements and decision ownership before deciding whether training is the right response.
What should we measure in an AI pilot?
Measure completed work, quality, total effort including review and correction, and the use of any released capacity. Choose measures appropriate to the task.
Sources & editorial notes
This resource combines cited source context with practical editorial guidance from Stories of Asia. Examples and suggested actions are not claims of measured client outcomes.
- Brynjolfsson, Li & Raymond — Generative AI at Work ↗
Research in a customer-support setting; not a universal estimate for all jobs.
Published by Stories of Asia. How we work with sources and corrections ↗
Take the question into a real conversation.
Explore SOA gatherings for HR leaders and peers working through AI, people and the workplace.
Explore SOA Events ↗