Most early AI benefits are described in tasks: draft the email, summarise the meeting, check the document, prepare the report. Those gains are real. I use them every day.
But task productivity has a ceiling. If the surrounding process remains unchanged, the organisation becomes faster at producing the same outputs through the same hand-offs. The savings are fragmented, and people often fill the released time with more activity. Useful, but not the step change I believe is possible.
The bigger opportunity begins with outcomes.
A customer does not experience the number of tasks completed across our departments. They experience whether their issue was resolved, whether the service was reliable, whether the answer was fair, and whether we kept our commitments. Regulators care about comparable outcomes: compliance, traceability, control, and the organisation’s ability to demonstrate that obligations were met.
When we start with an outcome, different questions appear. Which activities actually contribute to it? Which exist because of a historical system constraint? Where is judgement required? Where do decisions wait? Which checks add assurance, and which merely duplicate effort? What information would allow the work to flow without repeated interpretation?
This shift changes the role of AI. Instead of placing a copilot beside each task, we can redesign the flow of work. Some activities may disappear. Deterministic work may be automated. People may focus on judgement, relationships, exceptions, and improvement. Specialised agents may take responsibility for defined components, with clear escalation and verification around them.
It also changes measurement. Hours saved can be useful evidence, but they do not prove that the customer received a better outcome. A redesigned process should be judged against the dimensions that matter: quality, timeliness, safety, cost, consistency, capacity, and trust.
I am still learning how to operationalise this. It is easier to identify a task than to agree an end-to-end outcome. Outcomes cross functions, expose conflicting incentives and force conversations about ownership. That difficulty is not a reason to retreat to use cases. It is evidence that we have reached the real work.
AI may begin with tasks because that is where experimentation is easiest. Sustained value will come when we connect those experiments to the outcomes the organisation exists to deliver.