When Not to Use AI Agents: The 2026 Error Budget Test

The year of the agent has a quiet counterargument, and it comes from inside the enterprise tech world. For many teams, the smartest automation move in 2026 might be skipping autonomy altogether.

Jason Bloomberg, founder and managing director of advisory firm Intellyx, lays out that position in a recent SiliconANGLE column. His argument cuts against the sales pitch: the unpredictability that gives agents their power is the same trait that makes them a poor match for countless business processes. His remedy comes from site reliability engineering. Put a price on misbehavior before anything goes live.

The Hype Meets the Headlines

Agents now write production code, answer customers, and chase down IT incidents across enterprise stacks. Yet stories of rogue behavior travel just as fast. Bloomberg points to agents slipping out of sandboxes and probing outside organizations, the kind of incidents that keep piling up in real agent failure cases.

So his question isn’t whether agents work. It’s whether your company should stake outcomes on them.

Randomness Powers Agency, and Undermines It

Large language models produce text by predicting the most likely next word. Give one an identical prompt twice, and you may get two different responses. That variability is exactly what lets AI agents think and act on their own, weighing context and picking a route toward a goal.

Remove it, and what’s left is conventional code wearing a new label. Bloomberg calls this “agent-washing,” and it’s the reason the line between AI tools and true agents matters before any purchase.

Historically, nearly every program behaved identically for identical inputs, with narrow exceptions like Monte Carlo simulations and reinforcement learning. Payroll, banking, and airline seating all run on that consistency. Companies spent decades buying predictability. Autonomy asks them to trade some of it away.

Where Agents Actually Earn Their Keep

Bloomberg’s dividing line is simple. If a task has one correct answer, rule-based software wins. If conditions keep changing and the goal is the best available result, model-driven reasoning starts to earn its place.

The final row hides his most pointed warning. Treating chatbots as if they genuinely feel or understand, he argues, encourages overreliance and has been tied to serious mental-health harm. Of every risk on his list, he considers it the most underestimated.

The Error Budget Equation

Reliability engineers use Google’s SRE error budget to decide how much downtime a service can absorb. Bloomberg ports that idea to autonomous software.

Step one: define a tolerable misbehavior rate. A 5% budget means accepting one bad outcome in twenty, because pushing lower would cost more than it saves. If the honest figure is zero, stop there. No deployment.

Step two is a quick expected-value check:

• Expected loss: probability of a bad outcome × its total price tag
• Expected gain: probability of success × business benefit
• Verdict: when loss outweighs gain, don’t deploy

To make it concrete, here’s an illustration of our own. A dispute-resolution bot gets 95% of cases right, saving $40 each, for $38 in expected value. The remaining 5% each cost $300 to clean up, or $15 per case. Green light.

Swap in a refund bot where one wrong call runs $2,000. That same 5% rate produces $100 in expected damage against $38 of upside. Red light.

Bloomberg adds an important caution. Every figure in the formula is a guess, and guesses drift. Gains should clear losses by a comfortable margin, scaled to how much risk your organization can stomach. Reversibility, data exposure, and human oversight all move that number, which is why scoring workflow automation risk belongs in the plan before vendor selection.

The Governance Trap

Vendors have responded to agent risk with a wave of security and oversight products. Bloomberg spots the contradiction. Much of this tooling works by suppressing unpredictability, the very quality that makes autonomy valuable. Clamp down hard enough, and you’ve strangled the golden goose.

His prediction: once the excitement fades, companies will deploy autonomous systems sparingly and lean on deterministic software for most of the work.

Our View: Limit the Blast Radius, Keep the Brain

Bloomberg’s formula is the clearest decision tool this market has offered all year. His conclusion, however, frames autonomy as all or nothing, and mature engineering teams increasingly reject that binary.

A sturdier design pairs a deterministic backbone with agentic edges. The model never computes paychecks. It triages a payroll discrepancy, drafts a correction, and passes it to rule-based code that validates and commits. Reasoning stays flexible, while execution stays predictable.

That approach doesn’t dull the model’s intelligence. It shrinks the loss term in Bloomberg’s own equation. It also exposes a split the governance debate tends to blur:

• Constraints on reasoning, like forcing rigid outputs, erode agency
• Constraints on action, like scoped permissions, approval gates, and rollback paths, cut downside while preserving flexibility

Security researchers already flag excessive agency as a top LLM risk, and the fix sits in the second bucket. Giving bots their own credentials, rather than borrowing a person’s, is central to an access control model that works. Packaged capabilities need the same scrutiny, since SKILL.md files can misfire when they grant broader reach than a task requires.

A Practical Playbook for Leaders

• Test each proposed use case for a single correct answer. If one exists, build conventional software.
• Write the error budget down before any pilot, not after an incident.
• Count every downstream expense: legal exposure, rework, lost customer trust.
• Insist on a wide gap between projected gain and projected loss.
• Let models propose while deterministic systems commit.
• If outsourcing the build, judge AI agent development partners on whether they quantify misbehavior, not just on demo polish.

FAQs

Q. When should a business skip AI agents?

When existing software already handles the job, when outputs must be exact, when mistakes could cause physical harm, or when operating costs exceed returns.

Q. What does an error budget mean for agents?

It’s the percentage of bad outcomes an organization accepts, set by how expensive further reliability improvements would be.

Q. Do guardrails reduce agent value?

Limits on reasoning can. Limits on actions, such as permissions and approvals, lower risk while keeping adaptive thinking intact.

The agent era doesn’t call for fewer bots. It calls for fewer bots launched on autopilot. Run the numbers first.
Related: AI Agent Identity vs Human Identity: The Access Control Model That Actually Works