Security and delivery teams share an awkward habit: they commit to dates for work whose size they have never measured. Upgrading a library, wiring a scanner into a pipeline, or clearing a backlog of findings all get a number in the planning meeting, and rarely does anyone return to compare it with reality. Think of what follows as a self-check on your estimating practice that fits into one afternoon. Each item is a yes or no question, plus a note on the fix when the answer is no. The aim is simple: make guessed hours and actual hours comparable, so the next estimate can be corrected by a known factor.
Part one: are the guesses recorded?
- Is every estimate written down before work starts? If it lives only in a conversation, it will be rewritten by memory. Put the number in the ticket itself, in a dedicated field, with the date.
- Is the unit consistent? Choose ideal hours or elapsed days and stay with one. Mixed units make a ratio meaningless.
- Is the estimate owned by the person who does the work? Numbers handed down from a manager rarely reflect the real task. Where the owner differs from the doer, record both guesses, since the gap between them is itself informative.
Part two: are the actuals recorded?
- Do you capture actual hours on completion? A single field filled in when the ticket closes is enough. Reconstructing hours weeks later produces flattering numbers.
- Is waiting time separated from working time? Hours spent blocked on approvals, access or another team belong in a separate line. Otherwise you will blame skill for what was queueing.
- Are scope changes logged against the original guess? When requirements move, note the revised guess without erasing the first. Both figures feed the audit.
Part three: is the ratio calculated and read?
- Do you compute actual divided by estimate per task? Without the ratio, the data stays a pile of numbers. A result over 1 means the task ran long; under 1, it finished early.
- Do you group tasks by type? Dependency upgrades, pipeline configuration, finding triage and documentation each behave differently. Pool them and you flatten the signal. Keep the groups small in number, four to six is workable.
- Do you use medians rather than a single average? One unusual incident can distort a mean for months. A median ratio per task type is sturdier with samples under thirty.
Part four: does the ratio change the next plan?
- Is a correction factor applied to new estimates? This is the step most teams skip. If upgrade tasks have a median ratio of 1.8, a fresh upgrade guess of ten hours enters the plan as eighteen.
- Is the factor shown with its sample size? “1.8 from fourteen tasks” is trustworthy. “1.8 from three tasks” is a hint. Say which one it is.
- Is the factor revisited on a schedule? Review it quarterly. Tooling and skill change, and ratios should drift toward 1 as the team learns.
Reading your tally
Count your yes answers. Ten to twelve means you already calibrate and need only maintenance. Seven to nine means the records exist but are not feeding decisions; focus on part four. Four to six means the habit is half-formed, so start with items one and four, which cost almost nothing. Below four, begin with a single task type for one month and ignore everything else.
A sample record
To see how the pieces connect, take a team that logged eight pipeline-configuration tasks over a quarter. Guessed hours: 4, 6, 3, 8, 5, 4, 10, 6. Actual hours: 7, 9, 4, 17, 8, 6, 15, 12. The ratios are 1.75, 1.5, 1.33, 2.13, 1.6, 1.5, 1.5 and 2.0. The median sits at 1.6. The two highest ratios, the 8-hour and the 6-hour guesses that became 17 and 12, both involved a new credential store, so the team added a note: any task touching an unfamiliar secrets tool is planned at double. That is a rule born from evidence and it costs a line in the audit sheet.
Applying the factor to a real plan
Suppose next sprint contains four tasks: a dependency upgrade guessed at ten hours, a pipeline change guessed at six, a batch of finding triage guessed at eight, and a documentation update guessed at three. The team’s medians are 1.8 for upgrades, 1.6 for pipeline work, 1.3 for triage and 1.1 for documentation. Corrected figures are 18, 9.6, 10.4 and 3.3 hours, a total of about 41 against a raw 27. The team now has two honest options: commit to 41 hours of capacity, or drop one task before the sprint starts. Both are better than discovering the gap on the last Thursday.
Notice that nobody argued with the guesses. The estimators stayed free to say what they believed; the factor did the adjusting afterwards. That separation keeps the conversation calm, since no one has to defend an instinct against a spreadsheet in public.
What to do when the data refuses to settle
Some task types never produce a stable ratio. If a type ranges from 0.8 to 4.0 across fifteen samples, the problem is probably the grouping. Look at the extremes: does the high group involve a particular system, a particular dependency owner, or a certain stage of the release? Split the type along that line and re-run the numbers. If the spread stays wide even after splitting, plan that work with a range, say twelve to thirty hours, and flag it as high uncertainty so the people reading the plan know what they are looking at.
Where this fits with planning
Calibration pays off most when requirements are clear enough to estimate in the first place. If your scoping is shaky, two earlier pieces in this archive help: a detailed look at writing up web project requirements and a list of prompts to work through before a custom build begins. Better inputs make every ratio tighter.
For a wider view of how technology and learning meet in the voluntary sector, the pages of a nonprofit foundation’s website include programs, research, community and a blog, and offer some context from outside the delivery world. It plays no part in the audit itself.
Who should own the sheet
Pick one person to curate the audit sheet, ideally someone who closes tickets rather than someone who assigns them. Their job is light: check that the guess field was filled before work began, that actuals arrive on completion, and that the monthly median is recalculated. Rotating the role each quarter spreads the habit across the team and stops the sheet depending on one enthusiast. The curator should also write the one-line cause for any miss above double the guess, in plain words, because patterns hide in those notes.
Common failure modes
- Audit as blame. If ratios appear in performance reviews, people will pad guesses and the data is spoilt. Keep the audit about tasks, not individuals.
- Over-fine categories. Twenty task types means three samples each. Merge until each group has at least ten.
- Skipping the outliers. A task at five times the guess is the most valuable row in the sheet. Record its cause in plain words.
- Adjusting history. Never edit an old guess. Changing it destroys the only evidence of your bias.
Running the audit this week
Block ninety minutes. Walk through the twelve questions with two or three colleagues, mark each yes or no, and agree on the three nos that matter most. Assign a single owner to each fix, and fix a date for the next review. When the next quarter ends, repeat the count. The score is less important than the direction: more yes answers, tighter ratios, and plans that survive contact with the calendar.

