A service level agreement, or SLA, is the part of a contract in which a supplier commits to measurable standards of service and spells out what happens when it falls short. In technology, the standard is usually availability, expressed as a monthly uptime percentage, and the consequence is usually a service credit: a discount on a future invoice. SLAs exist because buyers of a service cannot inspect it the way they would inspect a delivered product. The agreement turns a vague promise of reliability into a number that can be measured, disputed and, in theory, compensated.

How an SLA is built

Most technology SLAs share the same skeleton. Google's Site Reliability Engineering book, published by O'Reilly, gives the clearest vocabulary for it. A service level indicator (SLI) is "a carefully defined quantitative measure of some aspect of the level of service that is provided," such as the share of requests that succeed. A service level objective (SLO) is "a target value or range of values for a service level that is measured by an SLI." The SLA is "an explicit or implicit contract with your users that includes consequences of meeting (or missing) the SLOs they contain."

In a written cloud SLA, that structure becomes a handful of clauses:

  • The metric. Usually a monthly uptime percentage. Google Workspace counts a period as downtime when the user interface shows "more than a five percent user error rate." Google Compute Engine counts loss of external connectivity or persistent disk access, in periods of one or more consecutive minutes.
  • The measurement window. Almost always the calendar month. Microsoft's commitment for Microsoft 365 services is measured per service, per month; Tony Redmond of Office 365 for IT Pros wrote in April 2026 that "claims cannot be made based on quarterly performance nor across multiple workloads."
  • Credit tiers. Amazon's EC2 SLA promises at least 99.99% monthly uptime for instances spread across two or more availability zones in a region. If uptime falls below that, the customer gets 10% of the monthly bill back; below 99.0%, 30%; below 95.0%, 100%.
  • Exclusions. Force majeure, customer-caused failures, third-party software and features still in preview are routinely carved out.
  • The claim process. Credits are rarely automatic. AWS requires a support case with a fixed subject line, received by the end of the second billing cycle after the incident. Google Workspace allows 30 days; Compute Engine, 60. Missing the window means the customer "will forfeit" the credit, in Google's wording.
  • The remedy cap. Nearly every hyperscaler SLA states that credits are the customer's "sole and exclusive remedy."

The arithmetic of nines

Uptime percentages hide how much downtime they permit. In a 30-day month of 43,200 minutes, a 99% commitment allows 432 minutes, or 7.2 hours, of outage before any credit is owed. At 99.9%, the allowance falls to 43.2 minutes. At 99.95% it is 21.6 minutes, and at 99.99%, about 4.3 minutes. Each additional nine cuts permitted downtime tenfold, which is why providers tie higher tiers to redundancy. Amazon offers 99.99% only for multi-zone deployments; a single EC2 instance carries a 99.5% commitment.

A worked example shows the scale of compensation. If a region were completely unavailable for 15 hours in a 31-day month of 44,640 minutes, uptime would land at roughly 98%. Under the EC2 region-level table, that earns a 30% credit on that month's EC2 charges for the affected service, applied to future bills only and not refunded.

Where SLAs sit in advertising

Advertising runs on infrastructure that carries SLAs, and on vendor relationships that increasingly carry their own.

On the sell side, Google Ad Manager's standard terms contain a service level clause for "Traditional Ad Serving" only. Google commits to use "commercially reasonable efforts" to process ad requests "at least 99% of the time," measured monthly, with a fee reduction credited to the next invoice as the "sole remedy." Programmatic and exchange demand sit outside that clause.

Analytics vendors attach SLAs to paid tiers. An archived version of Google's Analytics 360 SLA, dated July 2021, set 99.9% uptime for data collection, 99% for reporting, and a processing-latency target of four hours at 98% for properties of 2 billion hits or fewer per month. Repeated failures, in three consecutive months or four in any 12, gave the customer a termination right.

Agency relationships use the term more loosely. PPC Land's explainer on master services agreements describes an SLA as an agreement that "commits to measurable performance with defined remedies," typically appended to the MSA as a schedule. Agency metrics are operational: response times, reporting deadlines, launch turnaround.

Ad servers, demand-side platforms (DSPs) and measurement vendors mostly run on AWS, Google Cloud or Microsoft Azure, so a cloud outage cascades through the stack while each layer's SLA covers only its own contract.

Origin and evolution

SLAs predate the cloud. Fixed-line telecom operators have used them since the late 1980s, according to Wikipedia's summary of the field, covering metrics such as mean time to repair, throughput and jitter. IBM published its Web Service Level Agreement (WSLA) language specification in 2001, an attempt to make SLAs machine-readable.

The modern template arrived with public cloud. Amazon introduced an SLA for its S3 storage service effective October 1, 2007, with a 10% credit if monthly uptime fell below 99.9% and 25% below 99%, according to Data Center Knowledge. Four months later, in February 2008, an S3 outage of a little over two hours tested the model; InfoQ reported that the SLA "may not provide significant assurance." On October 23, 2008, Amazon took EC2 out of beta with "a service level commitment of 99.95% availability within a Region." Microsoft's hosted Exchange offering of the same year used a three-tier table of 25%, 50% and 100% credits at 99.9%, 99% and 95%, a structure Microsoft still follows.

Standards bodies followed. ISO/IEC 20000-1, the service management standard, reached its third edition in September 2018. ISO/IEC 19086-1, published in September 2016, set out "common cloud SLA building blocks," though it explicitly does not prescribe a standard structure.

Regulation has since moved SLAs from commercial nicety to legal requirement in some sectors. The European Union's Digital Operational Resilience Act (DORA), applicable from January 17, 2025, requires every financial entity's ICT contract to include "service level descriptions." For critical or important functions, Article 30(3)(a) demands "precise quantitative and qualitative performance targets."

Why it matters for marketers

An SLA tells a buyer how much failure a vendor considers normal. For advertisers and publishers, that number translates directly into lost impressions, missed conversions and broken reporting. When Google Ad Manager failed for roughly two and a half hours in December 2022, AdExchanger reported that the downtime came to about 0.4% of the month, inside the 99% commitment, so Google was "not bound" to compensate anyone. Krzysztof Franaszek of Adalytics called 99% "on the low end" against a typical 99.9% for hosting.

The gap between what an outage costs and what an SLA pays is the central fact marketers face. Google ad platform failures in January 2026 cut publisher revenue by 50% to 90% across two consecutive evenings as AdX match rates collapsed. Service credits, where they apply at all, are calculated on fees paid to the vendor, not on revenue lost by the customer.

Limitations and criticisms

Four objections recur. The first is the remedy cap: credits are small relative to business losses, and "sole and exclusive remedy" clauses block other claims. The second is the burden of proof. Customers must detect, document and file within tight windows; Google requires log files showing the downtime periods.

The third concerns measurement. Microsoft's 2008 formula calculated downtime in user-minutes across a whole tenant, so a two-hour outage hitting one user of 40 produced no credit. Redmond's tracking found Microsoft 365 uptime of 99.526% in the first quarter of 2026, "the lowest uptime figure since we started to record data in 2013," yet he noted that this did not mean Microsoft was "on the hook to pay out service credits."

The fourth is self-reporting. Providers usually decide what counts as an incident, through their own status pages. Google's own Search Status Dashboard records only what its operator chooses to disclose.

SLA versus adjacent terms

  • SLO. The internal target. An SLO can be stricter than the SLA so engineers react before contractual penalties apply.
  • Master services agreement (MSA). The umbrella contract covering liability, payment and data. The SLA is usually a schedule attached to it.
  • Guaranteed delivery and make-goods. In media buying, a guarantee covers impressions or ratings, and a make-good replaces shortfalls with extra inventory. It concerns audience delivery, not system availability.
  • Uptime figure. A published availability statistic is a measurement. Only an SLA attaches a consequence to it.

Recent developments

The SLA question returned today. Microsoft 365's email outage hit 19 services after Exchange Online degradation began on the afternoon of November 12, 2026, US time, with eDiscovery exports and Power BI sensitivity-label exports among the casualties. Reporting cited in that coverage, drawn from Microsoft's April 2026 volume licensing terms, put Exchange Online's commitment at 99.9% monthly, with credits of 25%, 50% and 100% at the familiar thresholds, claimable within one month after the billing month. Microsoft's licensing site lists a new edition of its online services SLA dated November 1, 2026.

The AWS US-EAST-1 failure of October 19 and 20, 2025, which lasted more than 15 hours and hit 142 services, remains the reference case for cascading dependency. Each affected company held its own SLA with AWS. None covered what end users lost.

Timeline

  • Late 1980s: Fixed-line telecom operators begin using SLAs for network services.
  • 2001: IBM publishes the Web Service Level Agreement (WSLA) language specification.
  • October 1, 2007: Amazon S3 SLA takes effect: 99.9% monthly uptime, 10% and 25% credit tiers.
  • February 2008: S3 outage of just over two hours prompts debate on cloud SLA value.
  • October 23, 2008: Amazon EC2 exits beta with a 99.95% regional availability commitment.
  • November 2008: Microsoft's hosted Exchange SLA documents 25%, 50% and 100% credit tiers.
  • September 2016: ISO/IEC 19086-1 cloud SLA framework published.
  • September 2018: ISO/IEC 20000-1 third edition published.
  • January 5, 2021: Google Ad Manager standard terms, including the 99% Traditional Ad Serving clause, last updated.
  • July 2021: Google Analytics 360 SLA version archived with 99.9% collection uptime.
  • December 2022: Google Ad Manager outage of about two and a half hours stays inside its 99% SLA.
  • January 17, 2025: DORA applies, requiring service level descriptions in financial-sector ICT contracts.
  • October 19-20, 2025: AWS US-EAST-1 outage lasts more than 15 hours.
  • January 13-15, 2026: Google ad platform failures cut publisher revenue by 50% to 90%.
  • Q1 2026: Microsoft 365 uptime falls to 99.526%, the lowest recorded since 2013.
  • August 31, 2026: Google Workspace SLA last modified.
  • October 14, 2026: Google Compute Engine SLA last modified.
  • November 1, 2026: Microsoft publishes a new edition of its online services SLA.
  • November 12, 2026: Exchange Online degradation begins, affecting 19 Microsoft 365 services.

Summary

Who: Service providers, from cloud hyperscalers such as Amazon, Google and Microsoft to ad servers, analytics vendors and agencies, write SLAs; their business customers rely on them and must claim any credits owed.

What: A service level agreement is a contractual commitment to measurable service standards, most often monthly uptime, paired with defined remedies, usually service credits capped as the sole remedy.

When: SLAs date to telecom contracts in the late 1980s; the cloud template took shape with Amazon S3 in October 2007 and EC2 in October 2008; DORA made service level descriptions mandatory for EU financial-sector ICT contracts from January 17, 2025.

Where: In cloud and SaaS terms of service, in schedules to master services agreements, in ad server and analytics contracts, and, for EU financial entities, in regulated ICT agreements.

Why: Buyers cannot inspect a service before using it, so the SLA converts a promise of reliability into a measurable number. The persistent dispute is that credits are calculated on vendor fees, not on the customer's losses, leaving most outage costs with the customer.