A 95% reliable AI agent gets it right 36% of the time

TL;DR: An AI agent that gets each step right 95% of the time gets the whole job right 36% of the time across twenty steps. Agent reliability isn't bought by switching models: it's built by shortening the chain, putting limits outside the prompt, and measuring against what the systems actually show rather than what the agent says it did.

The number that sounds good and isn't

A 95% success rate looks excellent. In an exam it would be a top grade; in a manufacturing process, respectable quality. In an AI agent working on its own, it is a figure that misleads — and it misleads for a purely arithmetic reason.

An agent doesn't perform an action. It performs a chain. It reads an email, checks the CRM, tests a condition, drafts a reply, updates a record, notifies someone. Each of those steps carries its own probability of going right, and those probabilities multiply.

At 95% per step:

• 5 steps: the full flow succeeds 77% of the time.
• 10 steps: 60%.
• 20 steps: 36%.

Put another way: a "highly reliable" agent handed a twenty-step process fails on two runs out of three. And it doesn't fail because the model is weak. It fails because nobody did the multiplication.

What changes when the bar goes up

The same arithmetic explains where engineering effort actually pays off. Across twenty steps:

Reliability per step	Full-flow success
95%	36%
99%	82%
99.9%	98%

Going from 95% to 99% isn't fine-tuning: it's the difference between an unusable process and a usable one. Going from 99% to 99.9% is what separates a usable process from one that can be left running unattended overnight.

Here's the catch: that improvement doesn't come from picking a smarter model. A better model adds a few points to each individual step. What moves the needle is changing the shape of the process.

What the market data says

This isn't theory. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, citing three causes: escalating costs, unclear business value, and inadequate risk controls.

None of the three is a model problem. All three are architecture and governance problems.

The same analysis adds a figure worth keeping in mind when choosing a vendor: of the thousands of suppliers currently marketing themselves as agentic, Gartner estimates only around 130 genuinely are. The rest it describes with a precise term — agent washing: rebranding existing products, AI assistants, RPA and chatbots, without substantial agentic capability.

Six layers, and only one of them is purchased

An agent in production rests on six layers. Naming them is worth the trouble, because it makes obvious where the real work sits:

1. Model — the reasoning capability. Licensed, swappable, improves without anyone lifting a finger.
2. Context — what the agent can see: which data it reaches, with what history, against which system of record.
3. Tools — what it can do: the functions and permissions exposed to it.
4. Control — the loop deciding what happens next: how it plans, how it routes and, above all, when it stops.
5. Guardrails — what it is not allowed to do, enforced outside the prompt.
6. Observability — traces and logging good enough to prove, months later, what it did and why.

The first one is bought. The other five are the responsibility of whoever deploys the agent. When an agentic project disappoints, the first layer is almost never the one that failed.

Three decisions that raise reliability without touching the model

Shorten the chain. If twenty steps yield 36%, the first question isn't how to improve each step — it's how many of those twenty steps exist only because the process was designed for a human. Splitting one long flow into two short ones with a human checkpoint in between isn't giving up on automation: it doubles the success rate.

Move the limits out of the prompt. If the only thing stopping an agent from doing something is a sentence in its instructions, that's not a guardrail — it's a suggestion. A real limit lives in permissions, in the tool itself, or in a validation that runs no matter what. The difference shows up the day it has to be explained to an auditor: "we told it not to" and "it can't" are not the same sentence.

Measure against the system, not against the agent. An agent can report that it sent an email, updated a record or escalated a case without having done any of it. The only valid check is the state of the system: is the email in the sent folder? does the record carry the new timestamp? Any measurement based on the agent's own summary measures narrative, not execution.

The metric almost nobody watches

When an agent works under human supervision, one number describes its health better than any other: the override rate — how often the reviewer changes what the agent proposed.

It reads in both directions:

• Very high — the agent isn't ready, and supervision is doing the work instead of checking it.
• Near zero — be careful, because this almost never means perfection. Usually it means review has become a rubber stamp, and the control has quietly stopped existing while still appearing on the org chart.

A healthy override rate is one that falls over time and never reaches zero.

Where to start

Three things, none of which require changing vendor:

1. Count the steps in a flow that is already automated, and do the multiplication. The resulting number usually explains, on its own, the feeling that "sometimes it works and sometimes it doesn't".
2. Check where each limit actually lives. If it's in the prompt, it isn't there.
3. Pick one success metric measured against the system, record today's baseline, and measure it again in thirty days.

A reliable agent isn't the one running the best model. It's the one with fewer steps to fail, with limits that don't depend on its good intentions, and with someone able to prove, six months later, exactly what it did.

---

Sources:

• Gartner — Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (25 Jun 2025): https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027