Estimate time saved by timing real tasks before and after the feature, and count review and rework as part of the new cost. A feature that drafts a fix in seconds but needs twenty minutes of checking has a real saving that is smaller than the headline suggests.
Time the baseline first
Pick one task, such as triaging a vulnerability or preparing a fix for a dependency bump. Time ten real instances done the current way. Record the median, not the average, since one slow case skews the mean. Note who did each, since seniority changes times.
Time the new way, fully
For the same type of task, time everything with the feature:
- Waiting for the output.
- Reading and verifying it.
- Editing or redoing it when it is wrong.
- Handling the cases where you discard it entirely.
Count discarded outputs. If the feature is wrong one time in four, the discards belong in the cost.
A simple comparison
| Step | Before | After |
|---|---|---|
| Gather context | Manual | Provided, still verified |
| Decide | Engineer | Engineer, with suggestion |
| Prepare change | Engineer | Draft plus edits |
| Review | Peer | Peer, perhaps closer |
| Rework | Occasional | Occasional, track it |
Fill in minutes for each step from your own timing. Do not use a vendor's figures unless you can reproduce them.
Watch for hidden costs
Review fatigue is real. If reviewers face a rising number of proposals, their attention per proposal falls. Track the accept rate and the time spent per review. A sudden jump in accept rate with a fall in review time may signal rubber stamping, not efficiency.
Count the quality side too
Time is not the only gain. Measure whether fewer issues are reopened, whether fixes land sooner, and whether urgent items get handled first. Report time and quality separately, so a quick but poor result is not mistaken for progress.
Be honest in the business case
When you share the result, give a range and state the sample size. Say what you did not measure. A modest, credible figure beats a large one that cannot be reproduced, and it protects your credibility if the numbers are later questioned. The SEC's March 2024 action over inflated AI claims is a reminder that claims about AI results are expected to be accurate.
Try it yourself
Click through it at your own pace, or open it full screen.
Walk through a fix and note which steps would still need a person, then time those in your own process.
Where Safeguard fits
Safeguard offers autonomous remediation with fix PRs, off by default with every category on Manual, so you can time the review step before extending trust. The free scan is at gold.safeguard.sh/scan.
Quick check
How many tasks should I time?
Ten per condition gives a rough median.
Should I trust a vendor's saved hours figure?
Only if you can reproduce it on your own tasks.
Sources
- SEC press release 2024-36, 18 March 2024: https://www.sec.gov/newsroom/press-releases/2024-36
Last reviewed: 19 October 2026.
