Service Level Agreement

While "it's reliable" was the answer given during evaluation, that answer went unverified. The selling side sincerely believes it's reliable and the buying side has no way to confirm the belief. That it went down twice in the past six months only surfaces if asked, and asking produces two different recollections of how long.

Enterprise procurement is where this surfaced. The review document before signing has a field for uptime as a number and a field for what the customer receives when there's an outage. Two empty fields and the approval doesn't move. Sofia explaining that it's a good product doesn't fill them; a document does.

The same problem remained after an actual outage. With no agreement on what counts as a breach, every incident becomes an argument. The customer says four hours of unusable service and our records say twelve minutes, because the measurement method and the verdict criteria were never written down. Both numbers are honest; they are answers to different questions.

It consists of six items. The order varies by company; which items cause disputes when missing does not.

1. Scope — which functions of which service this applies to. Document editing and saving are covered; notification email delivery is not, for instance. Leave scope out and every function in the product inherits the same promise.
2. The metric and its formula — pin down the service level indicator (SLI), the actual measurement of service quality. Measuring the share of five-minute probes that returned success, versus the share of real customer requests that succeeded, turns the same outage into different numbers. Who does the measuring gets written down too, since a customer's own monitoring and ours will not agree on the minute.
3. The target number — written as something like 99.9 percent monthly. This only makes sense once converted into permitted downtime. Over one month, 99 percent is 7 hours 18 minutes, 99.9 percent is 43 minutes, and 99.99 percent is 4 minutes.
4. Exclusions — announced maintenance windows, problems on the customer's network, requests the customer sent above their cap, and events outside anyone's control. Without this item, two hours of announced maintenance counts as a breach.
5. Compensation for a breach — usually a discount. Written as a staircase: below 99.9 percent, 10 percent of that month's fee; below 99.0 percent, 25 percent. A total cap keeps it from exceeding the month's fee.
6. Process and deadlines — how many days the customer has to file, who renders the verdict, and when the incident report arrives. Leave this field empty and the process gets invented after an outage, while both sides are already annoyed.

Choose the number by converting to downtime first. 99.99 percent is four minutes a month, and on a system that pauses 30 seconds per deploy, eight deploys exhaust the promise.

It isn't a one-person document. Three hands are required. Engineering sets numbers it can hold, legal shapes the wording and the limits of liability, and sales brings how far the terms can be negotiated. Reverse the order and unholdable numbers end up in a contract — which is what promising 99.99 percent in a negotiation and informing engineering afterward produces.

Set the internal target before writing it. Keep the service level objective (SLO), the quality target the team commits to internally, on the inside, and write a looser number into the contract. Set the two equal and missing the internal target once is immediately a contract breach. The error budget, the allowance for failure while still meeting the target, gets computed inside that gap.

Four audiences read it. The customer's procurement lead reads the numbers and the compensation clause; their security reviewer reads the exclusions. Their legal team reads the liability cap, and our own on-call engineer reads the downtime formula. That last one is the one usually left out. When the formula in the contract differs from the value we monitor, nobody can determine whether an incident was a breach.

How the document gets handed over is also settled. Usually one standard version exists with different numbers per plan. The free plan gets no document, the team plan gets the standard version, and only enterprise contracts negotiate clauses. Once negotiated versions accumulate, the number to hold differs per contract, so what was promised to whom belongs in one place.

"We can't write 99.99. We pause 30 seconds on every deploy."

Converting the promised number into downtime shows it can't be held.

"Put announced maintenance in the exclusions. Right now that counts as a breach too."

A missing item turns normal work into a violation.

"This customer is on a negotiated version at 99.95. It differs from the standard one."

Confirming that the number to hold can vary per contract.

At Pagedeck, Theo calculates the holdable number first and Grace settles the wording and the liability cap. The standard version is 99.9 percent monthly while the internal target sits higher at 99.95. Sofia's negotiating ceiling is fixed at 99.95, so customers asking for more get answered with other terms. The six negotiated versions live in one table of per-contract commitments, and that table is the first thing opened during an incident.

The most common mistake is reading 99.9 percent as almost never down. A high percentage with a decimal point reads as stability. But converted to a month it's 43 minutes, an amount a single incident consumes whole. Pause 30 minutes migrating a database and the month has 13 minutes left. Watching only the percentage, a team keeps deploying without knowing the remaining allowance and reaches a breach at month end. In practice it gets managed in minutes remaining, not percent.

Believing a well-written document produces trust is the other one. High numbers and generous compensation look reassuring on a contract. But compensation is usually a fraction of one month's fee, and what the customer lost to the outage exceeds it. A customer refunded 10 percent of a month rarely stays satisfied. What this document does is state a floor so procurement can proceed; trust comes from a record of holding that number for twelve months. Which is why showing the record beats negotiating the number when a renewal comes up.