Latané Conant: AI Customer Service - Observe Before You Automate

A customer contacts your company because something has already gone wrong.

They search for an answer. They enter a chat. They explain the problem. They are transferred. They explain it again. Eventually, they leave—not necessarily angry enough to complete a survey or demand a manager, but frustrated enough to reconsider the relationship.

Now imagine automating that journey without understanding it first.

The technology may work perfectly while making the underlying customer problem faster, cheaper and more scalable.

That is the risk at the center of the current rush toward AI customer service.

Generative AI can clearly create operational value. Research involving more than 5,000 customer-support agents found that access to a generative AI assistant increased productivity by roughly 15% on average, with particularly strong gains among less-experienced workers. McKinsey has separately estimated that generative AI could create productivity value equivalent to 30–45% of current customer-care function costs.

But knowing that AI can improve customer service is different from knowing where it should be used.

That distinction became central to my conversation with Latané Conant, Chief Market Officer at Parloa and a four-time CMO.

Her principle is surprisingly simple:

“Let’s listen first. Let’s observe.”

The implication is much bigger.

Why Do Companies Know So Little About Their Own Customers?

Businesses have spent decades constructing systems designed to understand customers.

CRMs record activity. Surveys capture feedback. Customer-data platforms combine signals. Marketing analytics track acquisition. Support systems classify tickets.

Yet much of the richest customer information may already exist somewhere else: inside the conversations customers are having with the company.

Conant describes those interactions as part of a “dark journey”—information that exists but historically has been difficult to analyze systematically.

She argues that this creates a strange contradiction. As a marketer, she spent millions trying to understand customers and persuade them to engage with companies. Then she realized that other parts of many organizations were simultaneously designed to reduce those conversations.

AI changes the economics of that problem.

Conversational analytics can process natural-language interactions across calls, chats and other channels to identify intent, sentiment, recurring problems and customer needs. IBM describes conversational analytics specifically as extracting insights from natural-language conversations to improve customer experience and decision-making.

Instead of sampling a handful of calls, an organization can increasingly examine interaction patterns across enormous volumes of conversations.

The first use of AI, therefore, may not be answering the customer.

It may be listening to them at scale.

Observe the Friction Before Building the Agent

Imagine that thousands of customer conversations reveal repeated scheduling problems.

That is evidence for a scheduling intervention.

If customers repeatedly struggle during onboarding, that suggests another opportunity.

If they continually reach the wrong department, automation might help with routing.

Conant summarizes the logic clearly:

“That’s how we know where the opportunities are.”

This reverses a common AI-adoption sequence.

Instead of:

Technology → use case → hoped-for ROI

the sequence becomes:

Conversation → friction → use case → AI intervention → measurement

That difference matters because AI makes experimentation easier—but it also makes scaling mistakes easier.

McKinsey's work on generative AI in customer care similarly recommends identifying and prioritizing use cases according to factors such as impact, feasibility and organizational need rather than treating AI as a generic layer applied everywhere.

For leaders, the practical question changes from What can we automate? to:

What recurring customer problem have we observed strongly enough to justify automation?

Customer Effort May Reveal More Than Customer Delight

One particularly useful signal is customer effort.

How difficult was it to buy something, solve a problem, obtain support or complete a task?

Conant argues that conversational data can increasingly allow companies to derive customer-effort and satisfaction signals without creating another burden for customers by constantly asking them to complete surveys.

The idea has established roots.

A widely cited Harvard Business Review study based on more than 75,000 service interactions introduced Customer Effort Score and found that reducing customer effort could be more predictive of loyalty than attempting to “delight” customers through extraordinary service.

That does not mean every company should abandon CSAT or NPS. HBR has also cautioned that there is no universal single best customer metric.

The more useful conclusion is operational:

Different signals answer different questions.

If you want to know whether your service architecture is forcing customers through unnecessary work, effort is especially revealing.

And conversational AI may make that signal easier to detect continuously.

The Customer Conversation Is Becoming Business Intelligence

The value of these conversations extends beyond customer service.

A recurring complaint can expose a product flaw.

Questions asked before purchase can reveal a marketing problem.

Confusion during onboarding can expose poor process design.

Repeated objections can improve sales enablement.

Sentiment changes can indicate retention risk.

IBM identifies customer support, voice-of-the-customer analysis, sales and marketing optimization, personalization and customer-journey mapping among the uses of conversational analytics.

This is where Conant pushes the discussion toward the metric she ultimately cares about: customer lifetime value.

“The whole point of this all is customer lifetime value.”

The objective is not to deploy the largest number of agents.

It is to improve the economics of the customer relationship.

That means asking whether AI reduces friction, improves resolution, helps customers stay, identifies better next actions or gives human teams better information.

Technology becomes the mechanism, not the objective.

Who Watches the AI Agents?

There is another side to “observe first.”

Companies must observe not only customers, but also the AI systems interacting with them.

Conant says research discussed during the interview found significant gaps in companies' ability to monitor deployed AI agents, raising concerns about hallucinations, behavioral drift and unmanaged customer interactions. Because those figures come from research she cited during the interview, they should be treated as Parloa's reported findings rather than independent industry-wide measurements.

The broader governance issue is independently well established.

Deloitte argues that as AI agents perform more autonomous actions, organizations require stronger observability, monitoring and governance because errors can propagate at scale.

Its recent work on AI-assisted customer service also emphasizes continuous monitoring and clear escalation paths so customers do not become trapped in automated loops when human intervention is needed.

In other words, the AI customer-service strategy is incomplete if it answers only:

What should the agent do?

Leaders also need to know:

How will we know whether it is doing it well?

The Emerging Workforce May Manage AI, Not Compete With It

This shift also changes customer-service jobs.

Conant expects many frontline and service-management roles to evolve toward supervising AI agents within hybrid teams.

Evidence already suggests a complementary model can create value. The large customer-support study published in The Quarterly Journal of Economics found that AI assistance benefited less-experienced workers particularly strongly, suggesting that AI can help distribute practices associated with more experienced employees.

The managerial challenge is therefore not simply replacing labor with automation.

It is redesigning work around what machines and humans each do well.

That includes deciding when AI handles an interaction, when a person enters, how context transfers between them, how performance is measured and who is accountable when the system fails.

Start With One Customer Problem

The temptation with AI is to start with capability.

Resist it.

Start with evidence.

This week, choose one high-volume area of your customer journey and examine the conversations occurring there. Look for repeated questions, transfers, delays, confusion, complaints and workarounds.

Do not ask first, What AI agent could we build?

Ask:

Where are customers repeatedly doing unnecessary work?

That single shift—from automation first to observation first—can turn AI from an expensive experiment into a disciplined customer strategy.

And once the problem is visible, deciding what to build becomes considerably easier.