Pegasystems CEO warns against deploying thousands of AI agents

Alan Trefler, founder and CEO of Pegasystems, warned that mainstream enterprise software vendors are encouraging companies to deploy large numbers of autonomous or semi‑autonomous AI agents across corporate networks. He said the scale being discussed—”maybe over 10,000″ agents-would be unsafe for processes that require consistent, auditable outcomes.

Trefler explained his concerns in remarks about the use of agentic AI for high‑volume business functions. He argued that agents can behave unpredictably and that monitoring tools known as “control towers” cannot fully eliminate the risk of inconsistent customer treatment or unintended legal and regulatory breaches in mission‑critical workflows.

To illustrate potential harms, he pointed to a marketing campaign in South Korea that used an AI‑generated slogan and triggered public backlash that led to significant brand damage and the resignation of the company’s chief executive. He said the example shows how quickly AI can produce decisions with reputational and legal consequences when allowed to operate without strict constraints.

Industry analysts working on agentic automation say Trefler’s concerns are relevant for regulated, high‑volume tasks such as insurance eligibility checks and banking processes. Neil Ward‑Dutton, research vice‑president for agentic automation and AI technologies at IDC, described replacing deterministic business applications with loose fleets of agents as “complete madness” for those use cases. He identified lower‑risk areas-marketing campaign creation, ad copy translation and catalog checks-where agentic approaches can work alongside humans.

Several large software suppliers are promoting hybrid approaches that combine AI agents with rule‑based workflows and guardrails. Platforms are being developed to give agents constrained, plug‑in skills and to route decisions through deterministic templates. Salesforce has introduced a product called AgentForce, and other vendors are building workflow layers to limit agent behavior. A German supplier, Cumunda, uses a mix of deterministic processes with dynamic AI decisioning.

Trefler also raised economic concerns, saying the rising cost of AI compute and token usage may be unsustainable. He recalled seeing many AI company billboards and predicted a market contraction that could force some startups to exit, comparing the situation to past technology cycles that produced excess capacity and company failures.

On workforce trends, Trefler noted low‑code tooling and large language models are shifting some application design work from specialized developers to business users. He cautioned that more code generated by LLMs does not automatically make systems more reliable and said the only credible measure of enterprise AI is whether it helps save or make money, not the volume of tokens consumed.

Quotes

“I think they have already acknowledged they’re going to have thousands of agents running, maybe over 10,000, and I think that philosophy is madness,” Trefler said. Ward‑Dutton called replacing deterministic systems with roaming agents for regulated workflows “complete madness.”