By now, we all know that generative AI models have their own personalities, or at least their own quirky behaviors. OpenAI’s GPT-4o model was so beloved for its warmth and easy friendliness that some users experienced genuine grief when OpenAI retired it. Other models have become known for being overly agreeable, maddeningly cautious, or perhaps a little too eager to offer life advise when all you wanted was a dinner suggestion.

But do these evolving model personalities matter within enterprise CX applications? We touched on this topic in a previous post from April where we noted:

One of the more interesting problems emerging in CX automation is that some of what vendors call agent design is, in practice, a workaround for the limitations of current models.

Anthropic has been providing valuable insights into how its model behaviors change with each new generation. With the release of Opus 5 they published a detailed guide for those using the new model within their Claude Code coding harness. While the guide is aimed at coders, the information it contains has potential ramifications for those using AI to power CX solutions.

Updating the Model Powering Your Customer-Facing AI Agent

If you or your CX software vendor decide to change the model powering your customer-facing AI agent (say from an older model to Opus 5), the ramifications probably aren’t enormous. There may be a shift in the agent’s tone or verbosity that warrants testing and adjustment, but chances are good the system will still work as intended.

Also, you don’t need a Fable 5 or Opus 5 to identify that a customer is asking about an order and pull up the shipping status. Plenty of cheaper models can do that job well, sometimes paired with conventional speech recognition and intent detection that doesn’t incur a token cost. So, there may be little point in updating the model behind your AI service agent, at least for now.

The differences start to matter once the model is running a multistep task, which is exactly what’s happening across the current wave of agentic AI use cases in CX. An agent doesn’t just answer a question. It gathers information, picks tools, and decides what happens next.

From Rules to Judgment – Models are Getting Better at Figuring Things Out on Their Own

Earlier LLMs were bad at following complex instructions, so developers compensated by piling on more rules, examples, and reminders in their prompts. When that still didn’t work, they might even throw in capital letters. The resulting prompts started to look like furniture assembly instructions: do this first, never do that, verify six conditions before proceeding, and UNDER NO CIRCUMSTANCES skip step 9B!

Anthropic is now saying this approach backfires on its newest models. In their recent post on context engineering for the Claude 5 generation, the company describes stripping more than 80 percent of Claude Code’s system prompt for Opus 5 and Fable 5, with no drop in coding performance. All those accumulated instructions and safeguards had started working against the model, forcing it to reason through directions that sometimes contradicted each other.

Anthropic sums up one of the key changes in how we should approach creating prompts for their generation 5 models:

Before, give Claude rules.

Now, let Claude use judgment.

Models, it seems, are improving at planning how to execute complex work and dealing with whatever obstacles arise. The way we instruct our CX AI agents about how to carry out workflows may need to adjust to this new reality if we want to get the most from these more mature models.

Your Agentic Workflows May Need Tweaking

Many CX vendors offer tools for building both single-purpose AI agents (for example, for detecting customer intent and responding to simple questions) as well as for constructing complex agentic workflows. For the latter, most agentic builder studios consist of a combination of generative AI-powered modules and deterministic components that spell out steps and tools in code.

A CX leader might assume a model upgrade won’t touch an existing workflow, since so much of the backend is deterministic. That’s often true, as the API call itself doesn’t wake up with a new personality. But the model deciding whether to call it might.

Take an agentic workflow responsible for handling returns. It carries out a series of tasks: look up an order, check refund eligibility, issue a refund, or escalate to a human. The refund process might be fully deterministic. But everything upstream of it might not be. The AI model might be the one judging whether it has enough information, what to authenticate before it proceeds, how to handle a tool error, and when the task actually counts as done.

If you or your vendor were to swap in a new model, that deterministic transaction won’t change at all. What changes is the judgment wrapped around it. A more capable model might infer a value the old one would have stopped to ask for. It might also run redundant checks left over from a prompt written to babysit a weaker predecessor, or push further on its own than anyone designed for.

This is where your team will need to test each piece of the process and make adjustments, possibly re-writing prompts that worked well for the older model. Anthropic is now offering a new command within Claude code, “claude doctor”, that diagnoses prompts and skills that need to be adjusted and makes recommended changes.

Wouldn’t it be great if CX vendors offered the same automated “fix it” tools when they decide to switch out the models powering their agentic capabilities?

Maintenance is Becoming Its Own Category

CX platforms will need their own version of claude doctor. Something that can scan a full agent configuration and flag conflicting instructions after older models are replaced with newer ones. The doctor will check for things like bloated prompts, redundant checks, tools that are too broad, and controls that really belong in deterministic logic instead of prose.

The healthy long-term setup separates the two layers for good: a stable definition of what a customer-service skill is supposed to do, its policies, permissions, and evaluation criteria, sitting apart from the model-specific instructions used to get a particular model to execute it. Business policy should hold steady. Prompt scaffolding shouldn’t have to.

What we want to prevent is the situation where your team learns that much of your “permanent” customer-service operation was actually just behavioral advice, written for a model that doesn’t want it anymore.

Categories: Articles