By James Proctor, Co-Founder and Managing Director, The Inteq Group
When AI agents behave unpredictably inside business processes, the dominant cause is not model instability. It is unspecified intent. An agent operating where outcomes are undefined, constraints are unstated, and escalation expectations exist only as tribal habit will fill those gaps with inference, and inference varies with context. The variance reads as unpredictability. It is more accurately described as improvisation, and the organization, not the model, wrote the conditions for it.
Why Do Agents Improvise?
Because the work demands more specification than the organization ever produced. Human-operated processes run for decades on ambiguity: experienced people absorb undefined goals and conflicting priorities, quietly harmonizing them through judgment and imitation. Nobody notices the ambiguity because people are so good at metabolizing it. An agent dropped into the same process receives the ambiguity raw. Asked to act where 'qualified,' 'acceptable risk,' or 'urgent' were never actually defined, it resolves each case as reasonably as inference allows, and reasonable inference is not the same thing as your intention. The agent is not hallucinating. It is improvising, because you never wrote the score.
Pilots routinely hide this, which is why the unpredictability arrives as a surprise. Pilot cases are curated, the team supervising them shares one implicit standard, and the volume is too small for variance to show a pattern. Scale the agent across the real case mix and the full spread of the organization's unstated definitions comes into play at once. What looks like degradation is exposure: the specification gap was always there, and production volume simply made it legible.
“Agents don't create ambiguity. They industrialize it.”
Why Is Blaming the Model So Attractive?
Because the alternative is admitting the specification never existed. When agent behavior varies, the convenient diagnosis is technical: the model is immature, the vendor oversold, we need a better platform. That diagnosis assigns the deficiency to something outside the organization and conveniently postpones the uncomfortable work of stating intent. I have reviewed a good number of 'unpredictable agent' escalations, and I will tell you what I almost always find underneath: three stakeholders with three different unwritten definitions of the same decision, each certain theirs was the standard. The humans disagreed for years. The disagreement was invisible because each of them was consistent with themselves. The agent, trained and instructed by all three, is consistent with none of them, and suddenly the variance has a face.
What Does This Look Like in a Real Process?
Candidate screening at staffing and recruiting firms is a clean specimen. Screening agents at such firms get accused of inconsistency on borderline candidates: the same profile advances for one role and stalls for another. Trace the inconsistency and you find that 'qualified' was never defined anywhere. One recruiter weighted recent tenure, another weighted skills adjacency, a third quietly downgraded employment gaps. Each recruiter was internally consistent, so clients experienced them as reliable. The agent inherited all three implicit standards at once, and its behavior faithfully reflects an ambiguity the firm had been selling around for years. The fix was never a better model. It was a definition, with the hard conversations that producing one requires.
How Do You Make Agent Behavior Predictable?
Specify what was always missing: the outcome the process serves, the constraints that bind it, the definitions that consequential terms actually carry, and the conditions that must reach a person. This is the same intent-layer work that leadership owes any agent-enabled process, and its absence is measurable in exactly the variance this post describes. Predictability is not a model property you purchase. It is a specification property you author.
Surfacing hidden definitional conflicts and converting them into explicit, agent-usable specifications is a core discipline in our agentic AI consulting practice, and it routinely resolves 'AI reliability' complaints without touching the technology.
