Most people building AI workflows ask the wrong first question. They ask: “How accurate is the model?” When the answer comes back at 95%, or even 99%, they feel reasonably confident.
The math disagrees.
If each step in a workflow has a 95% chance of success, a 10-step workflow succeeds just 60% of the time. Stretch it to 20 steps, which most real automation does, and you’re down to 36%. Run it for 50 steps and your success rate is under 8%.
That’s not a bug. It’s arithmetic.
Why the numbers feel counterintuitive
Ninety-five percent accuracy sounds high. In most contexts, it is. A decision that’s right 19 times out of 20 is a good decision. But sequential workflows don’t work on single decisions. They compound them.
Each step in an agentic workflow depends on the last one being correct. Errors don’t cancel out. They stack. And they interact in ways that make degradation worse than simple multiplication suggests.
Michael Hannecke’s published post-mortem of a production AI system documented exactly this. A 10-step workflow at 85% per-step accuracy produced a 20% overall success rate. A 20-step workflow at 95% accuracy, which is meaningfully better accuracy per step, produced a 36% success rate. More steps, better accuracy per step, worse outcomes.
The compound error problem isn’t about model quality. It’s about workflow architecture.
AI Strategy Insight
The Compound Error Problem
Per-step accuracy sounds reassuring. Workflow success rates tell a different story.
Overall success rate for a 20-step workflow at 95% per-step accuracy
agentic AI pilot failure rate, Q1 2026
cumulative failure rate at 100 steps with 99% accuracy
Sources: Gartner 2026, Six Sigma Agent (arxiv Jan 2026), Patronus AI / Business Insider
The self-conditioning effect
There’s a second problem that doesn’t appear in simple models, and it makes things worse.
When a language model processes context containing its own errors, it becomes more likely to compound those errors in subsequent steps. The model conditions on what it’s seen. If step 3 produced flawed output, step 4 is operating on corrupted context. Step 5 is operating on corrupted context plus whatever step 4 introduced. And so on.
Researchers documented this self-conditioning effect in multi-step LLM pipelines in early 2026. The degradation isn’t linear. It accelerates. The 36% success rate at 20 steps with 95% accuracy is already a conservative estimate; in practice, the failure curve is steeper.
DeepMind’s CEO Demis Hassabis described compounding AI errors as “compound interest in reverse.” An apt description. The mechanism that makes compound interest extraordinary (time working for you on every increment) is exactly what makes compound errors catastrophic in agentic systems.
What the industry data shows
The failure rates aren’t theoretical. Gartner’s Q1 2026 data put the agentic AI pilot failure rate at 78%. Over 40% of agentic AI projects are expected to be cancelled by 2027. These aren’t projects run by underfunded teams with bad intentions. They’re well-resourced attempts to deploy autonomous AI that ran into the compound error problem without the architecture to handle it.
Research published by Patronus AI put the specific maths in front of the industry clearly. A 1% error rate per action compounds to a 63% cumulative failure rate by step 100. ScaleAI found that a 20% error rate per action means a 5-step task succeeds only 32% of the time, despite each individual step being correct four times out of five.
The gap between how these numbers feel and what they produce is where most agentic AI investment disappears.
Why interactive AI success doesn’t transfer
There’s a pattern that comes up repeatedly with businesses progressing beyond individual AI use. They’ve had real results with ChatGPT, Claude, or Copilot in collaborative, human-in-the-loop mode. The AI suggests, the human corrects, the effective accuracy approaches 99%. It works well.
They then try to automate. Remove the human, chain the steps, deploy at scale. And things break down in ways that are hard to diagnose because no single step is obviously failing.
The reason isn’t that they’ve done something wrong in the automation. It’s that they’ve changed the fundamental architecture. Human-in-the-loop interaction is effectively a single-step system with human correction applied continuously. Automation is a multi-step system where errors compound before any human sees them.
What worked at the individual, interactive level hits a structural wall when you remove the correction mechanism. The tool hasn’t changed. The architecture has.
Building systems that account for this
The compound error problem isn’t unsolvable. But it requires a different design philosophy than simply building accuracy into individual steps.
Three approaches that work in practice:
Shorter chains with review gates. Agentic workflows with human review at 3-5 step intervals maintain significantly higher overall accuracy than long chains reviewed only at the end. The review cost is real. The alternative is a 36% success rate on automations you thought were reliable.
Structured output validation at each step. Building schema validation at each step, before passing output forward, catches errors before they compound. It adds complexity. It also means a 95% accurate step doesn’t silently corrupt everything downstream.
Scoped context passing. Passing only the relevant context forward at each step, rather than the full accumulated history, reduces the self-conditioning effect. A step that doesn’t know about errors in steps 2 and 3 can’t condition on them.
None of these requires a research breakthrough or a better base model. They’re engineering decisions. And they’re the decisions that separate agentic AI projects that ship from those that get cancelled.
The right question to ask
If your team is at the point of moving from interactive AI use to autonomous workflows, the first question isn’t “which model should we use?” It’s “how many steps does this workflow actually have, and what’s our architecture for handling compound errors?”
The businesses extracting real value from agentic AI aren’t the ones with the most accurate models. They’re the ones who’ve designed their workflows to account for the mathematics of compounding.
If your team is ready to move beyond individual AI use into structured autonomous workflows, the Advisors Edge programme covers agentic workflow design, error architecture, and the governance that keeps autonomous systems reliable. For businesses further along that need hands-on support alongside the leadership team, strategic advisory is the right starting point.