WhatsApp SLA on Black Friday: median, p90 and queue

Key takeaways

A useful SLA is the time to the first human response; the automated greeting does not count.
The average hides the queue: the median shows the typical case and the p90 shows the worst case that is still frequent.
Queue is the number of conversations where the last message is from the customer; read it every hour on peak days.
According to the Harvard Business Review (2011), responding within 1 hour gives about 7 times more chance of qualifying the lead than taking longer.
If the p90 exceeds the target for 2 consecutive hours, reduce the budget of the ad sets generating conversations until the queue fits the team's capacity.

What a WhatsApp SLA is and why the average misleads

An SLA (service level agreement) on WhatsApp is the time target the company commits to for giving the first human response to whoever wrote in. The average misleads because a few instant replies compensate, on paper, for many long waits. Nine replies in 1 minute and one in 91 minutes give an average of 10 minutes. The tenth lead, who waited an hour and a half, has probably already bought from a competitor. And whoever has not received any reply yet is not in the average: it only sees those who were served.

Three definitions need to be written down before measuring:

First human response: the first message sent by a person on the team, not by the greeting bot. If an AI runs the conversation, measure "first AI response" and "time until a human takes over" separately.
Clock: elapsed time or business hours. A message at 11 p.m. answered at 9 a.m. has a 10-hour wait under the first and 1 hour under the second. Pick one and state it in the report.
Queue: conversations where the last message is from the customer and the company has not replied yet.

How to measure: median, p90, open queue and longest wait

Use four numbers, calculated over all conversations started in the period, and publish all four together.

Metric	How to calculate	Good range on Black Friday	Warning sign
Median first human response	Sort the times and take the middle value	Up to 5 minutes during business hours	Above 15 minutes
p90 first response	Sort the times and take the value below which 90% fall	Up to 30 minutes	Above 1 hour
Open queue	Count conversations whose last message is from the customer	Lower than the team's capacity for the next hour	Growing for 2 consecutive hours
Current longest wait	Among the conversations in the queue, the oldest one	Below the p90 target	Above 24 hours (loses the 72-hour window)

The ranges are reference targets, not market benchmarks. Quick definition: p90 is the time below which 90% of responses fall. If the p90 is 40 minutes, one in every ten people waited longer than that: it is the frequent bad case, the one that generates complaints and lost sales.

Complement it with the percentage within target: answered within the SLA ÷ conversations started × 100. The denominator is conversations started: whoever is still waiting counts as outside the target. Below 80% during business hours, the operation is undersized.

What is the impact of delay on the sale?

Delay reduces the chance of contact and of qualification. The Harvard Business Review (2011), in an audit of 2,241 US companies, found that those who respond within 1 hour have about 7 times more chance of qualifying the lead than those who take longer; 24% of companies took more than 24 hours and 23% never responded. The caveat: it is an old study, from another market and done by phone; read it as an order of magnitude, not as a forecast for your account. The Brazilian context reinforces the urgency: according to Opinion Box (2025), 61% of Brazilians open WhatsApp several times a day and 60% have already bought through the app; whoever is writing has the phone in hand.

There is also a documented cost. According to Meta's WhatsApp Business Platform pricing documentation (2025), if the company responds within 24 hours to a lead who arrived through a Click-to-WhatsApp ad, a free 72-hour window opens, counted from the response; outside it, template messages are charged. A 24-hour queue loses the hot lead and makes follow-up more expensive.

Hypothetical example: reading the queue on a Black Friday morning

A hypothetical example, with made-up numbers to show the calculation. An electronics store runs Click-to-WhatsApp and receives 120 new conversations between 9 a.m. and 1 p.m. on Friday. The team has 3 agents.

Volume per hour: 120 ÷ 4 = 30 new conversations per hour.
Capacity: each agent gives a first response to 8 conversations per hour while handling the previous ones. Capacity = 3 × 8 = 24 per hour.
Deficit: 30 − 24 = 6 conversations per hour go unanswered. Over 4 hours, accumulated queue = 24 conversations.
Median: across the 96 conversations handled, the middle time is 5 minutes.
p90: sorting the 96 times, the value with 90% below it is 38 minutes, because from 11 a.m. onward responses fell behind.
Percentage within the 15-minute target: 70 answered on time ÷ 120 started = 58%. The 24 in the queue count as outside the target.
Longest wait: the oldest conversation in the queue arrived at 11:10 a.m. and at 1 p.m. was still unanswered: 1 hour and 50 minutes.

The correct reading: the 5-minute median says half were served well; the 38-minute p90 and the 58% within target say the operation overflowed. Decision: reduce the budget of the conversation ad sets by about 20% (from 30 to 24 conversations per hour, matching capacity) until the queue clears, or bring in a fourth agent (capacity of 32 per hour); the first is the only one the media buyer controls alone.

Decision rules: when to hold back budget and when to scale

Agree with the client, before the campaign, on an objective trigger and an action for each situation; without that, the media buyer watches the queue grow with no authorization to act.

If the p90 exceeds the target for 2 consecutive hours, reduce the budget of the ad sets generating conversations until the queue fits the next hour's capacity. Reducing is better than pausing: pausing and restarting resets the ad set's learning.
If the open queue exceeds the capacity of the next 2 hours, do not scale any ad set, even with a low CPL. Low CPL with a full queue is a cheap lead nobody answers.
If the median is good and the p90 is bad, the problem is distribution: someone is overloaded or a shift is uncovered. Redistribute before touching the media.
Only scale when median, p90 and queue are all within target at the same time.

How to size the team for the forecast demand is covered in how to size customer service for Black Friday. The calculation of how much budget the operation can absorb is in Meta Ads budget on Black Friday.

How to connect the WhatsApp queue to the media decision

The queue needs to be on the same screen as the spend, or the media buyer decides in the dark. That requires the conversation identified by source (campaign, ad set and ad), a record of who replied and when, and a response time report by source. A CRM with attribution and service analysis, such as RASTRO, shows the team's response time and follow-up per conversation and per source, and the decision stops being a hunch.

Always separate three groups in the report: answered within target, outside target and unanswered. "Unanswered" is not zero: it is the most expensive part of the campaign. That breakdown goes into the report that helps the client decide. If part of the first response is handled by AI, measure the AI by the same criterion; the limits are in AI selling with control over the offer.

Common mistakes when measuring SLA on Black Friday

Counting the automated greeting as a response. It zeroes the time in the report and makes the queue invisible. The customer knows it was a bot and keeps waiting.
Excluding unanswered conversations from the calculation. Average and median look great because they only measure those who were served. Use conversations started as the denominator.
Treating the queue as the client's problem only. The queue is the campaign's ceiling; ignoring it delivers "results" that never turn into sales. Slow replies on WhatsApp cost sales.

Frequently asked questions

Sources

What is the ideal WhatsApp SLA for Black Friday?

There is no universal number. A practical reference for anyone advertising with Click-to-WhatsApp is a median of up to 5 minutes and a p90 of up to 30 minutes during business hours, with the queue cleared at the end of each shift. Match the target to the team's real capacity.

Does an automated welcome message count as the first response?

No, for SLA purposes. The automated greeting confirms receipt, but it does not resolve the question or move the sale forward. Measure the first human response (or the first useful AI response, when the AI runs the conversation) and log the greeting separately.

How do I calculate the p90 in a spreadsheet?

List the first response times of all answered conversations, sort them from lowest to highest and take the value at the position corresponding to 90% of the list. The PERCENTILE function with 0.9 does this. Unanswered conversations count as "outside target" in the percentage, not in the p90.

Should I pause the campaign when the queue overflows?

Prefer reducing the budget of the ad sets generating conversations. Pausing and restarting resets learning and can make the restart more expensive. Reduce until the queue fits the next hour's capacity and scale back up when the longest wait is below the target.