Formula to calculate the percentage of conversations that require human intervention out of the total incoming

What exactly does this number measure

The percentage of conversations that require human intervention is the portion of the total incoming conversations in a period in which at least one person on the team had to write, correct, approve or intervene in some way. It does not measure how many messages a person wrote: it measures how many conversations were not resolved without them.

It is a number distinct from volume and distinct from response time. A business can receive 300 conversations per day and need human intervention in 15%. Another can receive 80 and need intervention in 70%. The first has a capacity problem; the second has a content problem: its responses do not cover what people are asking.

That is why the formula serves two purposes at once: sizing the team and deciding what to automate or what response to prepare. If the percentage is high, more people are not always needed: often what is needed is better response material.

The formula, with its three variables

The formula is a simple division, but each variable has a decision inside it. Write it like this:

Percentage of human intervention = (Conversations with human intervention ÷ Total incoming conversations) × 100

• Conversations with human intervention: number of conversations in the period in which at least one person from the team wrote, edited or approved a reply. The conversation is counted only once, regardless of how many human messages it contains.
• Total incoming conversations: all conversations started by the customer in the period, including those resolved on their own, those left unanswered and those escalated to another channel.
• Period: the date range you are going to measure. It must be the same for both variables or the number means nothing.

The decision that generates the most discussion is what counts as 'human intervention'. If you count any message written by a person, the number goes up and stops being useful for deciding. If you count only the cases where the person resolved something no one else could resolve, the number goes down and also stops being useful. The next section sets out that criterion.

What counts as human intervention and what does not

Before calculating, set the criterion in writing and apply it the same way in all periods. If you change the criterion halfway through, the historical series breaks and you will not be able to compare months.

The 'approved without changes' row is the most debated one. If your team approves with one click and sends, that conversation did not need a person to think about the reply: it needed someone to be there. These are two different things and it is worth measuring them separately. If you count approvals as intervention, you will conclude that you need more people when in reality you need to review the approval flow.

How to collect the data without rebuilding everything by hand

1. 1Choose the period. Start with a full week that has no holidays or major promotions, so the number is comparable with the next one.
2. 2Count the total incoming conversations for the period. If your tool gives you the total conversations started by customers, use that number. If it only gives you messages, do not divide messages: build the conversation count first.
3. 3Mark each conversation where there was at least one message written or edited by a person, applying the criteria table above.
4. 4Subtract from the total the ones you already marked as intervention. That difference is the conversations resolved without a person.
5. 5Divide the conversations with intervention by the total incoming and multiply by 100.
6. 6Write down the number along with the period date and the criteria used. A percentage without a date or criteria is not useful for comparison.
7. 7Repeat the same procedure the following week, with the same criteria, and compare.

If the manual count takes you more than an hour, do it anyway the first time: it gives you the baseline. Afterward you can automate the count with tags, marking each conversation that had human intervention. The important thing is that the criteria is written down somewhere on the team, not in one person's head.

Table to apply the formula to your case

Build one row per week. The columns on the left are the ones you count; the ones on the right, the ones you calculate. Don't make up values: if you don't have a data point, leave it blank and note that it's missing, instead of estimating it.

The monthly average column is calculated by adding up the conversations from the four weeks and dividing, not by averaging the four percentages. If one week had much more volume than another, averaging percentages gives you a distorted number. It is the most common mistake when building this table.

How to read the result: four possible scenarios

The number alone says nothing. It says something when you cross it with the volume and with the cause of the intervention. These are the four cross-references that matter.

The fourth scenario is the most misleading. A low percentage can mean that the system resolves well, or it can mean that many conversations were never answered and therefore there was no human intervention. To distinguish it, look at how many incoming conversations had no answer at all: if that number is high, the low percentage is an illusion.

Frequent errors when calculating this percentage

These are the five errors that make the number useless for deciding anything.

• Dividing messages instead of conversations. A conversation with twenty human messages weighs the same as one with one. If you count messages, the number measures something else.
• Counting each human intervention instead of each conversation. If three people intervened in the same conversation, it is one conversation with intervention, not three.
• Changing the counting criteria between periods. If one month you counted approvals and the next you did not, the comparison does not exist.
• Averaging weekly percentages without weighting by volume. The week with more conversations has to weigh more.
• Measuring the percentage without measuring the unanswered conversations. A low percentage can hide a lot of people who wrote and nobody answered.

If you find any of these errors in a calculation you already did, do not correct it retroactively: note the correction and start a new series from the period in which you applied the correct criterion. Mixing series with different criteria is worse than starting from scratch.

What to do with the number once you have it

The percentage of human intervention is useful for three specific decisions. The first is about capacity: if the percentage stays the same and the volume rises, the number of people needed rises in the same proportion. The second is about content: if the percentage is high and the volume is stable, the intervention is concentrated in types of queries that could have a better prepared response. The third is about process: if the percentage is low but the response time is high, the bottleneck is not in how many conversations need a person, but in when that person is available.

For the second decision, the next step is to open the conversations with human intervention and group them by reason. A perfect taxonomy is not necessary: five or six reasons are enough to see where the work is concentrated. If a reason appears in half of the conversations with intervention, there is something to resolve there.

For the third, the missing data is not this percentage but the hourly distribution of incoming conversations. A low percentage with a distribution concentrated in a few hours still needs the same people during those hours. The two things are read together.