The right mix of AI and human labor in the contact center

Back to Insights

The question is not whether to use AI. It is where AI does the work well, where a person does, and how to design the split for the specific operation.

The market has settled on a comfortable answer: AI is the answer. Point it at the contact center, remove the agents, watch the cost fall. The honest read is less tidy. There is no perfect answer for every operation; it is a mix, and the work is designing the right one.

I say that as someone who now builds the evaluation content a frontier AI lab uses to measure large language models. I am not skeptical of AI. I am skeptical of deploying it without measuring where it actually holds up, because a model is only as good as the evaluation that measures it.

What AI does well today

There is real value here, and it is worth being specific about it. AI deflects repetitive, low-stakes contacts that never needed a person, from balance checks to order status. It summarizes and drafts, cutting after-call work and giving agents a head start on notes and responses. It sits beside a human as assist, surfacing the right knowledge in real time so a newer agent performs like a seasoned one. It routes and triages more accurately than a phone tree. And it makes quality visible at a scale humans cannot match, scoring every contact instead of a one-percent sample.

Those are not future promises. Deployed against the right work, they lower cost and raise consistency at the same time.

What still needs a person

The failure modes cluster in a predictable place: judgment, stakes, and the unfamiliar. A person still wins on ambiguous situations, on empathy and de-escalation, on the edge cases the model has not seen, and on decisions where being confidently wrong is expensive.

In insurance work, which is where I spend much of my time, the gap is easy to see. A model can draft a coverage summary in seconds, and it can also produce a fluent, wrong answer with total confidence. On a first-notice call from someone who just had a loss, tone and judgment are the job, not text generation. The point is not that AI cannot touch these. It is that the cost of a mistake sets how much human judgment belongs in the loop.

The over-automation trap

The most common mistake I see is optimizing the wrong number. A team sets an automation-rate or deflection target, hits it, and reports a win while true resolution and CSAT quietly fall. The customer who could not get their real problem solved by the bot calls back twice and then leaves.

I could want an orange and need an orange, but they are selling me a rose. A high deflection rate can be a rose. It looks like the thing you wanted and does not do what you needed. You always get what you pay for, and cheap automation that frustrates customers is not cheap once you count the repeat contacts, the escalations, and the churn.

Designing the mix

Start from the work and the outcome, not the tool. Measure cost per resolved contact and the quality bar the work requires, CSAT and first-contact resolution, rather than automation rate. Automation rate is an input a vendor can move; resolution is the outcome you actually care about.

From there the pattern is straightforward. Put AI where volume is high and stakes are low. Keep people where judgment and stakes are high. And use AI to make your human agents better, through assist, summarization, and quality coverage, not only to remove them. Then instrument it honestly and keep measuring, because both the model's quality and your own mix will drift over time.

None of this is new in shape. Every wave of contact center technology promised to remove the human, from the IVR to offshoring to the first generation of chatbots, and each one reshaped the work rather than erasing it. AI is a bigger wave than most, and the pattern still holds: it changes what the humans do, it does not make the humans irrelevant.

There is no perfect answer for every operation; it is a mix, designed for the specific work and measured against the outcome. If you are trying to figure out where AI belongs in your operation and where it does not, it is worth at least having the conversation.

Worth at least having the conversation.

Tell us what you are trying to fix, scale, or evaluate. We will give you an honest read on whether we can help.

Contact us