AI exposure is an important signal, but leaders eventually need to ask: what might this mean for capacity, workload, and investment? The answer should be more rigorous than “AI will save 30%”. It should also be more honest than a single headline number.

The strongest way to estimate potential gains is to calculate them activity by activity, then roll them up to the role using the proportion of time each activity represents.

The logic is straightforward. For each activity, estimate four things:

1. Time share: What proportion of the role’s time is spent on this activity today?
2. Effective AI impact: How strong is the relevant automation or augmentation signal after weak evidence is filtered out?
3. Realistic efficiency effect: If AI is successfully adopted, what proportion of time might be reduced or redirected for this type of activity?
4. Adoption over time: How quickly will the capability be available, trusted, learned, governed, and embedded in the work?

A simple conceptual model for an augmentable activity is:

`Potential time freed = activity time share × effective augmentation factor × realistic efficiency effect × adoption factor`

The calculation is intentionally conservative. A high exposure score does not automatically mean high gains. If an activity represents only a small share of a role, its effect is limited. If the work needs substantial review, the realistic efficiency effect is lower. If systems are not integrated or users are not trained, the adoption factor delays the benefit.

Consider a document-review activity that takes 20% of a role’s time. Suppose the activity has a strong augmentation profile, a conservative 25% efficiency assumption once embedded, and 60% adoption in the first relevant period. The estimated time freed is not 25% of the job; it is 20% × 25% × 60%: 3% of the role’s time in that period. Repeat the calculation across all activities and the role-level picture emerges from many visible assumptions.

Automation requires a related but different calculation. It must account for the share of activity volume that can be automated, residual human oversight, exception handling, implementation cost, and the time required to stabilise the new process. Very few real processes reach 100% straight-through execution, and a credible business case should not assume that they will.

A good model also uses a gate or smoothing function so weak signals contribute little or nothing. This prevents a long tail of low-confidence activities from inflating the total. It is better to show a narrower, defensible range than an impressive number that cannot be explained.

Finally, time freed is not synonymous with jobs removed. It can be redeployed in several ways:

• Absorb higher volumes without proportional hiring.
• Improve service quality or speed.
• Reduce overtime, backlog, and administrative burden.
• Invest more time in judgment, customer interaction, improvement, and development.
• Create capacity for new work that AI itself makes possible.

This is the leadership value of transparent activity-level maths: it turns broad AI claims into assumptions that can be challenged, piloted, and improved. It makes the model a learning system rather than a promise.

**Next article:** Once we understand how activities are changing, we can make skill investment much more precise.