You have a small platform team, a handful of people with permanent administrator rights, and a nagging feeling that this is more power than anyone needs all day. The fix is a rollout of time-bound elevation, and it fits into one working month if you take it in order. This is a step-by-step AccessOps how-to: nine steps, each with a concrete output, written for teams that would rather do the work than debate it.
Step 1: Decide what counts as privileged
Open a blank page and list every right that could change production or expose data in bulk: cloud owner roles, database superusers, root on servers, repository admin, secrets-manager admin, identity-provider admin. Do not worry about completeness yet. Aim for the top ten. The output is a list you can show to a colleague in two minutes.
Step 2: Sort the list into tiers
Rank by how much damage one misuse could do. The Ops Atlas, whose AccessOps reference entry inspired this plan, uses four tiers with matching ceilings: an hour or two for the top tier, two to four hours below it, four to eight hours for the third, and standing access with an annual review for the lowest. Copy the idea, then adjust the numbers to your risk appetite. The output is a table with three columns: right, tier, maximum window.
Step 3: List who holds each right today
Export the real memberships rather than trusting memory. For each privileged right, write down every person and every automation that holds it. You will likely find names you did not expect. Count how many are standing at the top two tiers, because that figure is your baseline. The Atlas example reduced standing top-tier accounts from thirty-four to nine over its sample period, which gives you a feel for what progress looks like.
Step 4: Choose the mechanism per system
Different platforms offer different ways to make access temporary, and you should check their limits before promising anything. Microsoft’s directory-privilege service lets activation last from one to twenty-four hours. Teleport allows a configured cap of up to fourteen days on a request. On plain Linux servers, the picture is more mixed: the Atlas notes that the newer Rust rewrite of sudo, which has become the default in recent Ubuntu releases, lacks time-window rules, timeouts and input-output logging, while the original sudo needs extra options to stop users from escaping into a shell. The output of this step is one line per system stating how you will make a grant expire.
Step 5: Write the request form
Forms feel bureaucratic until you see what the alternative looks like at two in the morning, with a chat thread full of half-remembered permissions.
Keep it to five fields: the requester, the target right, the reason, the end time and the approver. Make “reason” a free-text box and insist on a ticket or incident number where one exists. Blanket approvals by a single manager lead to stock justifications, so plan for routine low-tier requests to go to an on-call colleague or a rule, and reserve human sign-off for the top tier.
Step 6: Set up the emergency door before closing the others
This step comes before you remove any standing rights. Create at least two emergency accounts that live outside your normal directory, protect them with sign-in methods that resist phishing, and set an alert for any use. Test them. Microsoft asks customers to validate such accounts at least every ninety days, and that interval is a good default. Remember that these accounts are meant to stay enabled permanently; that is a feature, not a failure of the programme. The output is a dated test record.
Step 7: Pilot with one team for two weeks
Choose a team that is comfortable with change, ideally the one that owns the platform itself. Move their top-tier rights to request-based windows and keep a log of every request, how long approval took and whether the grant was fully used. Ask the pilot group to open a long-running session just before a grant ends, so you can find out what actually happens at expiry. The Atlas reports that a session can outlive the grant that created it, and that cloud permission changes take a few minutes to propagate, so build a small buffer into your expectations.
Step 8: Check the numbers and adjust
After two weeks, calculate these figures: how many requests were made, the typical and slowest time to approval, what fraction of granted time was used, and how many requests were approved but never used. If most windows were barely used, shorten them. If approvals are slow, add an on-call approver. For the principles behind weighing these trade-offs, our earlier primer on balancing reliability and security risk in delivery may help, and newcomers to the vocabulary can keep our long list of core DevOps terms open in another tab.
Step 9: Extend, record and review
Roll the pattern out to the next team and the next tier. Add session recording for the top tier where your tooling supports it, and agree who samples the recordings; capturing everything and watching nothing has little value. Schedule a review of any remaining standing access at least once a year, and treat that review as a backstop, not the main defence. Finally, document one page of procedure: how to request, how to approve, how to use the emergency route, and who owns the process.
A small worked example
Suppose a team of six holds permanent database superuser rights. In step 3 you discover four of them have not used the role for a month. In step 4 you pick a request flow with a two-hour limit. During the pilot, twelve requests arrive, and the log shows an average of forty minutes of actual use per two-hour grant. The conclusion writes itself: shorten the default to ninety minutes, keep a manual extension option, and remove standing rights from the four idle accounts. These numbers are illustrative, but the reasoning is exactly what the real measurements will let you do.
Mistakes to avoid along the way
- Treating the programme as complete when the tool is installed.
- Forgetting the people who approve: empty approver groups can lock an entire tenant.
- Letting infrastructure code overwrite your privileged bindings each time it runs.
- Assuming a short window stops someone who already holds a live session.
If you follow the nine steps, the end of the month brings a smaller list of standing administrators, a documented request path and a rehearsed emergency route. That is not a finished state, but it is a measurable and honest improvement, and every later refinement builds on it. Come back to the tier table every quarter, because systems and teams change, and the list that was right in January will have gaps by summer.









