At some point a customer asks, "How reliable is texting in your product?" If you answer "very," you've set yourself up. If you answer with a number you can't measure, you've set yourself up worse. A service level objective (SLO) is the honest middle: a target, a way to measure it, and a plan for what happens when you miss.

SMS is an awkward thing to put an SLO on, because part of the path belongs to you and part belongs to carriers. This guide shows how to split the two, which indicators to pick for sending, delivery and webhooks, and how an error budget turns a vague promise into a decision.

Start with who controls what

A text crosses four hops: your app, your SMS vendor's API, a messaging provider and the carrier network, then the handset. You can promise things about the first hop. You can pass along what your vendor commits to for the second. For the rest, you can measure and report, but a promise would be a guess.

That's why a single "SMS success rate" makes a bad SLO. A failed delivery to a number that was a landline, or a phone switched off, isn't an outage. A 500 from your own send endpoint is. Mixing them hides both.

HopExample failureWho owns itFit for an SLO?
Your app and queueJob stuck, bad retry loopYouYes, strictly
Vendor API5xx, timeout, slow acceptYou and your vendorYes, as a dependency
Provider and carrierFiltered, throttled, delayedCarrier networkReport, don't promise
HandsetOff, landline, opted outThe recipientNo

Pick indicators you can compute

An SLI is the measurement, an SLO the target on it. For SMS platforms, four indicators carry most of the weight.

  1. Acceptance rate. Of the sends your product tried, the share the vendor API accepted without a 5xx or timeout. Count 4xx separately: a refused send for an opted-out number is the system working.
  2. Time to accept. From the moment a user clicks send, or a job fires, to a successful API response. Track the 95th and 99th percentile, not the average.
  3. Delivery rate by cohort. Delivered over accepted, split by tenant and by carrier where you have it. This is a report, with an alert, not a promise.
  4. Webhook freshness. From the moment a status or an inbound reply reaches your endpoint to the moment your users see it. This one is fully yours.

Our post on delivery rate monitoring for platforms covers the per-tenant alert side. Here we focus on the promise.

Set targets from what you've measured

Don't pick 99.99 because it sounds good. Run a month of measurement first, then set the target a little under what you hit, and leave room.

Here's an illustration with assumptions, not a benchmark. Say your product sends 600,000 texts a month. You set an acceptance SLO of 99.9 percent over 30 days. The error budget is the remainder: 0.1 percent of 600,000, which is 600 failed accepts a month. If one bad deploy burns 450 of them in an hour, you've used three quarters of the month's budget in one incident, and that tells you to slow down releases for the rest of it.

For latency, an example target might be that 99 percent of sends are accepted within two seconds. That's a number you can check against your own logs, not a claim about carriers. If your sends run through a queue, measure from enqueue to accept, and watch queue age as its own signal.

Treat carrier-side delivery as a reported figure

Delivery depends on the number, the content and the carrier. Ready sends on registered routes, and unregistered traffic gets carrier-filtered, so registration status belongs on your dashboard next to the delivery rate. Content filters, throttling on a fresh campaign and per-second limits all show up as delayed or undelivered texts, and none of them is an outage.

So publish delivery to customers as a trend by tenant, with context, and keep the promise for what you control. Two pages help you build the explanation: what a delivery receipt does and doesn't tell you and triaging a delivery failure for support.

Webhook SLOs: freshness and loss

Status callbacks and inbound replies arrive at your endpoint, and what you do with them is yours. Two SLOs work well.

  • Ingest success. The share of webhook calls your endpoint acknowledged with a 2xx on first attempt. A target like 99.9 percent is common to aim at, and the number should come from your logs.
  • End-to-end freshness. The 95th percentile of time from receipt to the row being visible in your product.

Webhooks are delivered at least once, so duplicates and out-of-order events are normal, and your endpoint should acknowledge fast and process later. Count a duplicate that you correctly ignored as a success, not an error. See webhook deduplication, fast acknowledgement and dead letter queue replay. Replay capability is itself part of your reliability story: a webhook you can replay costs you minutes, one you can't costs you a customer.

Spend the error budget on purpose

An error budget is useful when it changes behavior. A simple policy:

  • Under 50 percent of the budget used: ship normally.
  • 50 to 100 percent: hold risky changes, fix the top cause.
  • Budget gone: freeze features that touch the send path until you're back under.

Write down what counts against the budget before the incident, not during it. Decide whether a vendor 5xx counts (it does, for acceptance), whether a blocked send for quiet hours counts (it doesn't, the rule worked), and whether a refused send for a missing consent flag counts (it doesn't, but a spike in them says your integration is wrong). Quiet-hours refusals, opt-out refusals and bad-number refusals are behavior, not failure. Keep them in a separate counter so a legitimate refusal never eats budget.

Retries matter here too. A retry that succeeds shouldn't count as a failure for the customer, but it should show up in your internal numbers, because a rising retry rate warns you before the customer sees anything. Use an idempotency key on every retry so a recovery never sends twice. See retries and failure handling.

What to tell customers

Publish a short statement, not a legal contract:

  • What you promise: texts are accepted by our system within a stated time, and webhooks reach your view within a stated time.
  • What you report: delivery trends per number and per carrier, with the reasons.
  • What you don't control: carrier filtering, handset state and the recipient's choices.

Check what your SMS vendor commits to before you copy any number into your own terms. Ready's pricing and fees are itemized on the pricing page, and the vendor due diligence checklist lists the reliability questions to ask any provider, us included.

Frequently asked questions

Should delivery rate be an SLO?

Report it, alert on it, but don't promise a figure. Most of what drives it sits with carriers and recipients.

How long should the SLO window be?

Thirty days is common. Shorter windows react quickly and are noisy, and longer ones hide a bad week.

Do opt-outs count as failures?

No. A refused send to an opted-out number is the compliance system doing its job, so count it in its own bucket.

Next step

Before you write a target, measure for a month: acceptance rate, time to accept at the 95th and 99th percentile, webhook freshness and delivery by tenant. Then set one SLO on what you control and put the rest on a report. The observability guide is a good place to start building the dashboard.