What should we ask for?

Both promises rather than one: what proportion of the time the service will be available, and how quickly somebody will engage when you report a problem. Then the definitions, since what counts as downtime and what maintenance is excluded decide what the headline figure means.

Does it cover answer quality?

No. Standard agreements address availability and response, and a model whose behaviour has shifted is neither. That gap is not an oversight so much as a genuinely hard thing to write, and it is the reason keeping your own set of checked examples remains necessary whatever you have negotiated.

Is it worth having if the remedy is small?

Usually, because the remedy is not the point. A vendor committing in writing has had to arrange something internally to meet it, and that arrangement is what you are actually buying. How readily they commit, and how precisely, is informative before anything goes wrong.

Do credits arrive automatically?

Generally not. They are usually claimed within a stated window, which means somebody has to notice, evidence it and ask in time. An agreement nobody monitors delivers its value through the vendor's incentives and effectively none through its remedy, which is worth knowing when deciding who watches.

Glossary

Service level agreement (SLA)

A service level agreement is a contractual promise about availability and response times with agreed consequences if they are missed, rather than a statement of intent.

In plain terms

A written promise that the service will be available a certain proportion of the time and that somebody will respond within a certain period, with something owed to you if it is not. The something is usually a credit against future payments rather than compensation for what the outage cost you, which is worth knowing before it is relied on as protection.

Why it matters

Because it is the difference between a vendor intending to be reliable and being obliged to be, and because most organisations discover which one they have during an outage. It also has a second function that is quietly more useful than the remedy: a vendor willing to commit in writing to a response time has to have arranged something internally to meet it, so the willingness itself is evidence about how the relationship will behave.

How it works

Availability and responsiveness are separate promises and both matter. One says how much of the time the service will be up, the other says how quickly somebody will engage once you report a problem, and they are frequently confused. A high availability figure alongside no commitment on response is a common shape, and it leaves you covered for the failure mode that is rarer.

The remedy is usually a service credit, which is a refund of part of what you paid rather than compensation for the disruption. That is standard commercial practice and it means the agreement's real value is as a statement of intent backed by a cost to the vendor, not as insurance. Reading it as insurance is what causes disappointment.

What counts as downtime is defined, and the definitions do a great deal of work. Scheduled maintenance is usually excluded, partial degradation may not count at all, and the measurement period affects how much a short outage matters. Two agreements quoting the same availability figure can differ substantially once those are read.

For AI tools the conspicuous gap is quality, and it is worth being explicit about. These agreements address whether the service responded, not whether the answer was any good, and a model that has been updated into behaving differently for your use is not an availability event. Nothing in a standard agreement covers that, which is why holding your own examples to check against remains necessary.

Claiming is usually your responsibility rather than automatic, and the window is often short. Credits are typically requested rather than applied, which means somebody has to notice the breach, gather the evidence and ask in time. An agreement nobody is watching against provides most of its value through the vendor's incentives and very little through its remedy.

What it covers, what it does not

Setting the two columns out makes it clear why an organisation can have negotiated carefully and still be unprotected against its realistic problems. Availability commitments were designed for services that either work or do not, and they transfer to AI tools accurately but incompletely: a model that has been updated and now handles your particular material less well is available throughout, responding promptly, and producing worse outcomes. No clause has been breached. The practical response is not to negotiate harder on availability, which is usually already adequate, but to recognise that the right-hand column stays with you whatever the contract says, and to keep the small set of examples that would let you notice it. That is the one protection nobody else can supply.

Seen in the wild

Asking what a vendor commits to on availability and response before an assistant becomes something a team depends on daily.
ChatGPT
Checking whether an automation platform's administered tier carries a support commitment, given unattended runs fail silently.
Make
Comparing what different providers commit to behind a single interface, where a change of provider is a routing decision.
OpenRouter

Common misconceptions

People assume

It compensates us for an outage.

In fact

It typically credits part of what you paid for the period, which is rarely related to what the disruption cost. Its value is that missing the target costs the vendor something and therefore shapes their behaviour, rather than that it makes you whole. Treating it as insurance is the commonest misreading.

People assume

It covers the quality of what we get back.

In fact

It addresses whether the service responded, not whether the answer was good. A model updated into behaving differently for your work is not an availability event and falls outside a standard agreement entirely, which is why your own checks on output remain necessary regardless of what is committed.

People assume

A high availability figure settles it.

In fact

Not without the definitions. What counts as downtime, whether partial degradation counts at all, what maintenance is excluded and over what period it is measured can make two identical-looking figures mean quite different things. The exclusions are where the substance is.

Telling them apart

Service level agreement vs Enterprise plan

Service level agreement

The specific promise about availability and response, with a consequence.

Enterprise plan

The tier at which such a promise generally becomes available at all.

The plan is where you get one; the agreement is what it actually says.

Questions

What should we ask for?: Both promises rather than one: what proportion of the time the service will be available, and how quickly somebody will engage when you report a problem. Then the definitions, since what counts as downtime and what maintenance is excluded decide what the headline figure means.
Does it cover answer quality?: No. Standard agreements address availability and response, and a model whose behaviour has shifted is neither. That gap is not an oversight so much as a genuinely hard thing to write, and it is the reason keeping your own set of checked examples remains necessary whatever you have negotiated.
Is it worth having if the remedy is small?: Usually, because the remedy is not the point. A vendor committing in writing has had to arrange something internally to meet it, and that arrangement is what you are actually buying. How readily they commit, and how precisely, is informative before anything goes wrong.
Do credits arrive automatically?: Generally not. They are usually claimed within a stated window, which means somebody has to notice, evidence it and ask in time. An agreement nobody monitors delivers its value through the vendor's incentives and effectively none through its remedy, which is worth knowing when deciding who watches.

Key takeaways

Availability and response time are two promises; check for both.
The remedy is usually a credit, not compensation for the disruption.
The exclusions and definitions decide what a headline figure means.
Nothing standard covers answer quality, so keep checking output yourself.
Credits are usually claimed rather than applied, within a short window.

Tools that use this

ChatGPT
What is committed before a team depends on an assistant daily.
Make
Whether an administered tier commits to support, given silent unattended failures.
OpenRouter
Comparing provider commitments where switching is a routing decision.