TalkAn SLA credit is not a refund
The remedy clause of a service level agreement is the only part of it that involves money, and it is the part nobody reads. Four published agreements were read for this essay on October 24, 2026. None of them pays a refund.
They pay a credit, in a unit the vendor chooses, against a bill that has not arrived yet, if you notice in time and can prove it.
The documents say the word themselves
The Amazon Bedrock service level agreement, last updated October 4, 2023, removes the ambiguity in one sentence. "Service Credits will not entitle you to any refund or other payment from AWS."
The mechanism is set out just above it. "We will apply any Service Credits only against future Amazon Bedrock payments otherwise due from you." A card refund appears only at the vendor's discretion.
Two further restrictions in the same paragraph narrow it again. A credit is issued "only if the credit amount for the applicable monthly billing cycle is greater than one dollar ($1 USD)".
And it is locked to the account that suffered the outage: "Service Credits may not be transferred or applied to any other account." If you stop buying, the remedy expires unused.
Atlassian's cloud service level agreement reaches the same place from a different direction. "No refunds or cash value will be given for unused Service Credits."
The percentage is applied to a different month
The credit tables look generous until you read which invoice they land on. Amazon Bedrock pays 10 per cent of the affected region's monthly charges for a month under 99.9 per cent, 25 per cent under 99.0, and 100 per cent under 95.0.
Atlassian's table for Premium plan cloud products pays 10 per cent, 25 per cent and 50 per cent across the same three bands, calculated on "the monthly fees attributed to the affected Eligible Cloud Product".
Its own worked example is the clearest statement of scale on any of these pages. Sixty downtime minutes in a 43,200 minute month gives 99.86 per cent uptime, which earns a credit of 10 per cent of that product's monthly fee.
An hour of outage, in other words, is worth about three days of subscription. For a team paying $2,000 a month for the affected product that is $200, by our arithmetic from the published table.
Atlassian then states where the money goes, and a parenthesis does the real work.
The cap is 100 per cent of the amount invoiced for the affected product "in that invoice billing period", which "is not the month in which the affected Cloud Product was unavailable".
There is a solvency condition attached as well. The credit applies "provided that Customer's account is fully paid up, without any overdue payments or disputes". A billing argument in progress suspends the remedy.
One gap is worth recording honestly. The Enterprise plan table on that page listed four uptime thresholds but carried no credit percentages beside them in the page markup read on October 24, 2026.
Nobody pays you automatically
Every agreement read here places the filing burden on the customer, and the windows are short enough to miss while you are still writing the incident report.
Atlassian requires a ticket "with all fields fully and accurately completed within fifteen (15) days after the end of the calendar month in which the alleged failure occurred".
Google Workspace gives thirty days from eligibility to create a support case, and is blunt about the consequence. "Failure to comply with this requirement will forfeit Customer's right to receive a Service Credit."
Amazon Bedrock is the most generous on time and the most demanding on form. The request must arrive "by the end of the second billing cycle after which the incident occurred".
It must also carry "the words 'SLA Credit Request' in the subject line", the billing cycle and regions claimed, and "the specific dates, times, and Availabilities for each 5-minute interval with less than 100% Availability".
Plus your own logs, "that document the errors and corroborate your claimed outage". The closing line removes any doubt about strictness: incomplete requests "will disqualify you from receiving a Service Credit".
Whose measurement counts
Asking for the credit means arguing about numbers, and two of these agreements decide that argument in advance.
Atlassian: "Atlassian's monitoring and logging infrastructure is the sole source of truth for determining whether Atlassian has met the Service Level Commitment." Your logs support the claim; they do not settle it.
The definitions then set a floor under what counts at all. An Atlassian downtime minute occurs only when the error rate in that minute exceeds five per cent. One failure in a hundred requests is not downtime.
Amazon Bedrock counts an error as "any Request that returns a 500 error code", measured over five-minute intervals.
It adds a clause that quietly protects idle accounts: "If you did not make any Requests in a given 5-minute interval, that interval is assumed to be 100% available."
Google Workspace defines downtime for a domain as a period in which the user interface "has more than a five percent user error rate", and adds that "Downtime is measured based on server side error rate".
Google pays in days, not dollars
The Google Workspace agreement is the clearest example of a credit denominated in something other than money. Its service credit is "Days of Service added to the end of the Service term".
The ceiling is fifteen days of service for all downtime in a single calendar month, for the affected service. Monetary credits appear only for customers who bought online and are billed monthly.
The agreement says so directly. "Service Credits may not be exchanged for, or converted to, monetary amounts," with that one billing exception. An annual customer is paid in subscription time.
All three agreements close the same door on anything further. Each describes the credit as the customer's sole and exclusive remedy, and Atlassian adds that credits are "Atlassian's entire liability" for the failure.
One published agreement has no number in it at all
Slack's page titled Service Level Agreement is worth reading precisely because of how short it is. On October 24, 2026 the operative text was a single sentence.
"Salesforce uses commercially reasonable efforts to make the online Slack Services available 24 hours a day, 7 days a week," excepting planned downtime with notice and anything beyond its reasonable control.
There is no uptime percentage, no credit table and no claim procedure on that page. It is an effort commitment, not a measured one, and it is published where a buyer can find it.
Atlassian's agreement has a narrower version of the same gap. The commitment covers Premium and Enterprise plan products only, and expressly excludes "sandbox instances or Free or Beta Products".
What a credit is actually for
- It is a discount, not compensation. It reduces a future invoice from the supplier that caused the outage, and three of these four agreements say it cannot become cash.
- It is small relative to the harm. An hour down earns roughly 10 per cent of one month's fee on the two tables that publish percentages.
- It expires. Fifteen days after month end at Atlassian, thirty days at Google Workspace, two billing cycles at Amazon Bedrock.
- It needs your evidence and their measurement. You supply logs and five-minute interval detail; the vendor's monitoring decides.
- It may not exist. Free tiers, sandboxes, beta products and at least one whole product line carry no measured commitment.
What follows in practice
Stop treating the credit as insurance and price it as what it is: a rebate worth a few per cent of annual spend in a bad year. Then put the real money where the real risk is.
That means a calendar reminder for the claim deadline, your own error logging because you will need it to file, and a decision about fallback made before the outage rather than during it.
The fallback question is argued in the claim that you can always switch, and the slower version of the same dependency risk in when a model you built on is retired.
The pattern of a familiar word doing work its definition cannot support is the same one described in unlimited is a pricing word, and the billing version of it in the retry you pay for twice.
Plan and price detail, with the dates each figure was read, sits in our primer on AI assistant pricing for small teams and in our subject pages for ChatGPT and Claude.
A refund returns your money. A credit returns part of it as permission to keep buying. The agreements are honest about which one they offer, in sentences most buyers never reach.
Sources, read against the live page markup on October 24, 2026: aws.amazon.com Amazon Bedrock service level agreement; atlassian.com cloud service level agreement; workspace.google.com Google Workspace service level agreement; slack.com service level agreement.