Why your agentic AI strategy is backwards (and how to fix it) – News

These days, it’s hard to keep up with the whiplash-inducing hot takes on generative artificial intelligence. One minute, it’s the savior of enterprise productivity; the next, we’re beaten down by fearmongers and their tales of AI doom and gloom.

But if we step back and look beyond the headlines, perhaps we’ll realize that both sides might be wrong. The truth often lies somewhere in between.

I’m willing to bet the reason most gen AI pilots likely fail is not because the technology is inadequate, but because it’s applied in misguided ways. Organizations are simply using the wrong kind of gen AI at the wrong time for it to work reliably in the enterprise.

Many are trying to apply the ability of AI agents to “reason” during live customer interactions – a capability known in the software business as runtime. Reasoning AI empowers AI agents to get creative, make decisions on their own and even improvise new workflows on the fly.

It feels fast, modern and exciting. But it’s also unpredictable, unscalable and nearly impossible to govern.

Consider two bank customers applying for a mortgage with an AI agent. If you’ve ever used a large language model, you know it can give different answers to the same question. While the agent might approve a loan for one customer at a preferred rate, it could also decline a different customer with a fundamentally similar financial profile for reasons known only to the LLM.

This inconsistency isn’t just unfair — it could be noncompliant, or even unlawful. If you can’t trust an LLM to give consistent answers, how can anyone expect to trust it in situations where one mistake could cost millions of dollars and inflict massive reputational damage?

Yet that’s what enterprises risk by unleashing the full creative power of reasoning AI at runtime. It’s a disaster waiting to happen.

Keep reasoning AI behind the scenes

That doesn’t mean reasoning AI isn’t viable in the enterprise. We just need to apply it when and where it belongs.

Reasoning AI is much better suited for design time,  when information technology architects and business leaders get together and plan the optimal workflows for their software to help run their business. Here, AI reasoning serves as the ultimate brainstorming partner, dreaming up new ways to solve their thorniest business problems.

A misstep or two by reasoning AI in design time is perfectly acceptable, since there are no bad ideas in a brainstorm. At this stage, creativity is a strength, not a liability. Then it’s up to a human to select the best ideas to be codified as trusted workflows that all agents – virtual, human or otherwise – will follow.

Agentic AI can still be applied at runtime with customers; it simply requires a different type of AI to ensure the predictable and reliable outcomes that enterprises demand. That’s called semantic AI.

Sibling harmony

If reasoning AI is your fun-loving brainstorming partner, semantic AI is its straightlaced sibling that gets everyone aligned around the way things are described. Its power comes from understanding context and meaning. It processes words, intent and situational nuance. This enables semantic AI to assist agents in determining exactly what a customer is trying to accomplish and then identify the most suitable workflow to achieve their goal.

The precision of semantic AI makes it ideally suited for use in live, runtime customer interactions. Because when you’re processing a mortgage, onboarding a new patient or handling a sensitive insurance claim, there’s no room for improvisation. Semantic AI follows the workflow step by step, ensuring a predictable and reliable outcome every time.

Workflows are the secret sauce for ensuring enterprises can deploy agents they can actually trust. Once semantic AI understands the customer’s intent, it searches all the approved workflows and selects the most appropriate one to complete the task. This way, the work is done exactly as expected, eliminating any risk that the agent will go rogue. As an added bonus, this approach consumes significantly less energy than constantly ‘rethinking’ the problem for millions of live customer interactions using reasoning AI.

The best of both worlds

By applying reasoning AI during design time and semantic AI during runtime, enterprises can have the best of both worlds. They can leverage agentic reasoning and creativity on initial workflow designs while ensuring predictability with semantic AI where it matters in live engagements. It’s the optimal approach for creating a sustainable, reliable and scalable agentic framework for even the most complex enterprise environments.

Despite all the confusion about AI’s practical value, I remain unwaveringly optimistic about generative AI. I believe this is the most exciting and promising time to be in the software business in my more than 40-year career.

Generative AI has the potential to help enterprises innovate in new ways, overcome tech debt, modernize legacy systems and deliver more personalized engagement. It can help organizations move beyond one-size-fits-all approaches and spam-based marketing into a new era of relevance.

But that promise won’t be realized through the half-baked approaches that are all too common today. It will be happen when enterprises embrace the discipline of trusted workflows, the power of reasoning AI at design time, and the predictability of semantic AI at runtime.

The formula for agentic success has been in front of us all along. Applying different types of AI at the most appropriate times is how we’ll move from overbaked hype to lasting transformation.

Alan Trefler is the founder and chief executive officer of Pegasystems Inc. He wrote this article for News.

Photo: Pegasystems