Start with comparable work
Historical data is valuable only when the old work resembles the new work. Separate categories where size, complexity, site conditions or customer requirements change the duration materially. One global average can be misleading if it mixes five-minute jobs with five-hour jobs.
Record enough context to understand the number later. A duration without task type, scope and major conditions is hard to reuse.
Look beyond the average
An average describes the centre, not reliability. If most jobs take one hour but a meaningful minority take four, planning everything at the average will create frequent misses. Median, percentiles and a simple range can reveal the spread.
For commitments, a higher percentile may be more appropriate than the mean when lateness is costly. For internal capacity modelling, expected values can still be useful because highs and lows may balance across many jobs.
Update the model
Historical numbers should not become permanent folklore. Processes, tools, staff experience and scope change. Review the data periodically and separate older conditions if they no longer represent current work.