Forrester: Your Old QA Playbook Can't Test an AI Agent

Forrester has published a framework arguing that traditional software testing alone cannot validate AI applications, because AI systems fail in ways conventional QA wasn’t built to catch: nondeterministic outputs, incorrect retrieval, drift over time, and confidently wrong answers delivered with the same tone as correct ones. Its recommended approach combines classical testing (functional, performance, security) with AI evaluations — “evals” — treated as an ongoing engineering discipline rather than a one-time research exercise, plus red-teaming, human review, and production telemetry tied to governance controls. The report states evals “must be developed alongside the AI-enabled application and become reusable engineering assets,” and predicts convergence between classical testing vendors adding LLM observability and AI-validation specialists adding enterprise QA capability. Forrester’s five action items for leaders: inventory AI applications by business risk, define acceptable versus unsafe behavior explicitly, build evals as living assets, connect testing disciplines into one continuous feedback loop, and assign shared product-delivery ownership for AI quality.

The framework matters for sales organizations selling AI products into the enterprise as much as for the buyers reading it. Gartner made a parallel argument in August aimed specifically at CFOs piloting finance AI agents: treat the first deployment as a governance pilot, not an ROI showcase, because unlike prior automation, agents “interpret objectives, plan steps, execute tasks and interact with multiple systems” — meaning errors propagate fast when controls are weak. Gartner’s central line was blunt: “early pilots are most likely to fail due to unclear controls, not poor technology.” Both firms are converging on the same buyer expectation, from opposite functions — testing and finance — arriving at the identical conclusion that governance and evaluation infrastructure has to exist before scale, not get retrofitted after a failure.

For AI sellers, that means procurement conversations are shifting from feature comparisons toward “show me your eval process” — a harder, slower sales cycle, but one that rewards vendors who can actually answer the question.