Sarah Whitfield

Their call: Escalation · Graded Oct 6, 2026

A clearing-house partner was getting 429 Too Many Requests on every call during their month-end batch, with claims due the same day. Their rate limit had not changed, but their batch had doubled in size. Raising a partner's limit is not a support decision, so I opened a P1 with the platform team inside ten minutes with the request volumes attached, and meanwhile helped the partner split the batch into two windows so the urgent claims went first. Updated them every thirty minutes until the platform team raised the limit.

Reviewer Recognised it was not hers to fix and escalated fast with the evidence an engineer needs. Kept the customer moving with a workaround and regular updates. This is how a P1 should be run.

Their call: Configuration · Graded Oct 5, 2026

A newly promoted manager could see claims but not approve them. Her account had the Manager title but still held the Reviewer role, because the title change does not change the role. Confirmed with her director that approval rights were intended before granting them, assigned the Approver role, and asked her to approve a test claim while we were still talking. Suggested their admin add the role step to their promotion checklist.

Reviewer Confirmed authority before granting access, which many skip. Quick, and the checklist suggestion stops the next one.

Their call: Performance · Graded Oct 3, 2026

The dashboard took over a minute to load, but only for users in the west region. The AI triage blamed the database. I did not accept that, because the same queries were fast for everybody else. Comparing a slow request with a fast one showed the west users loading a year of claims by default: their saved filter had been reset by Tuesday's release. Restored the 30-day default for the region, confirmed load time was back under three seconds, and logged the release defect for engineering with both requests attached.

Reviewer Overruled the AI triage with evidence, which is the skill. Fixed the customer's problem now and handed the cause to engineering instead of trying to patch the release. Clear, short write-up.

Their call: Data fix · Graded Oct 1, 2026

A clinic saw every patient from Monday's import twice. Before touching anything I checked whether the duplicates were identical or only looked it, because merging two real people is far worse than leaving two copies of one. They were identical, down to the member ID, and every pair came from the same import file run twice after a timeout. Merged them with the import's own batch ID as the key, so nothing outside that batch could be touched, then sent the clinic the count before and after.

Reviewer Checked the dangerous case first and scoped the fix to the batch. Exactly the caution a data fix needs. The note to the customer had the numbers in it, which is what they will forward to their manager.

Their call: Integration · Graded Sep 25, 2026

Webhooks stopped arriving for one customer at 09:40. Their endpoint was returning 500 to every delivery, so our retries had backed off and then stopped, which is working as designed and looks like an outage from their side. Showed them the delivery log with the response codes rather than telling them it was their fault. They had deployed at 09:38. Replayed the failed deliveries once they had rolled back, and wrote up how to read the delivery log so their team can check this themselves next time.

Reviewer Showed the evidence instead of arguing, which is the hardest thing to teach. The run-book note is the reason the next person will not need us. Ready for a real desk.

Their call: Permissions · Graded Sep 16, 2026

Report exports were failing for one team only. I called it a permissions problem because the failing users were all in the same group, and I was wrong. It was a row limit on the export job, and that group simply had the largest dataset. The group was a coincidence. What I should have checked first was whether the export failed for a smaller date range, which would have ruled permissions out in one step.

Reviewer Wrong call, and the write-up says so and says what would have found it faster. That is worth more to me than a lucky right answer. Escalated at the right moment.

Their call: Data fix · Graded Sep 7, 2026

A monthly statement was showing the wrong closing balance. The customer assumed we had lost a payment, which was reasonable but not what happened. Their payment posted twice and one was reversed, and the reversal landed in the next statement period. So the balance was right and the statement was misleading. Corrected the posting date on the reversal and told them plainly what had happened, including that we had made it hard to read.

Reviewer Good separation of what went wrong from what it looked like, and the customer was told the truth without being blamed for misreading it. Timing was well inside target.

Their call: Configuration · Graded Aug 29, 2026

Customer could not sign in after our SSO change on Tuesday. Checked whether it was one user or the whole org first, because the fix is completely different either way. Two other people from the same tenant had opened tickets that morning, so it was the org. The metadata certificate on their identity provider had expired the night before. Asked them to re-upload it, confirmed with a test login on my side, then asked them to try. Logged the expiry date so we catch the next one before the customer does.

Reviewer Right instinct: one user or the whole org, asked before theorising. Slow to escalate to the identity team, which cost about fifteen minutes. Write-up is clear and I would accept it on our desk.

Check the duplicates are the same person, not two people who look alike: member ID, date of birth, and the import batch. Scope any merge to the batch ID. Tell the customer the count before and after, in numbers.

Open Integrations, then Deliveries. Filter by the customer's endpoint. A run of 5xx responses means their endpoint is failing, not ours. Retries back off and stop after the last attempt; that is working as designed. Once they have fixed their side, replay the failed deliveries from the same screen. Send them a screenshot of the response codes rather than a summary.