22% of AI Agents That Reach Production Lose Money. None of the Reasons Are the Model.

The Number That Should Change How You Deploy

Twenty-two percent of AI agents that reach production deliver negative ROI at the twelve-month mark. That's from Forrester's The State Of Agentic AI, 2026, published this month.

Read that again. These aren't the pilots that died in the lab. Roughly 88% of pilots never reach production at all. The 22% are the survivors — the agents that got deployed, integrated into real workflows, and are actively running. And more than one in five of them is losing money a full year in.

Here's the part almost nobody is talking about. Forrester's root-cause analysis of that losing cohort names three failure modes: unclear success criteria (41%), insufficient tool or data access (33%), and drift in evaluation coverage (26%). Zero of them are about model capability.

The models work. The agents are in production. And they're still destroying value — because the organization deploying them never encoded what "success" actually means. This is the most important operational lesson of 2026, and it has nothing to do with which model you picked.

The Chase Is Easy. The Catch Is Expensive.

Forrester's own framing of the year is blunt: "Companies are chasing, few are catching." Three-quarters of enterprise leaders say they're adopting agentic AI. Only a small minority have anything running in meaningful production beyond a chatbot with a better haircut.

The technology arrived faster than anyone expected. Agents now run for hours, days, and months. What didn't arrive is organizational readiness — and readiness is the whole game once the model stops being the bottleneck.

We've watched this pattern up close. The businesses that struggle aren't the ones with the worst models. They're the ones who treated deployment as the finish line. You ship the agent, it does the task, the demo looks great — and then twelve months later finance asks what it actually returned, and nobody can answer, because nobody defined the answer before turning it on.

That gap between "it works" and "it works for what we actually needed" is where the 22% live. It's not a technology story. It's an intent story.

What the losing cohort has in common

The negative-ROI agents share a profile, and it's remarkably consistent across the data:

• No binary definition of success. The goal was "improve customer service" or "handle support tickets" — a human-readable aspiration, not an agent-actionable objective. The agent optimized for something measurable and adjacent, and nobody caught the mismatch until the numbers came in.
• Missing tool and data access. The agent couldn't reach the systems it needed to actually deliver value, so it under-performed a job it was never fully equipped to do.
• Evaluation that decayed. Prompts changed, workflows shifted, the model got swapped — and the tests that would have caught the regression didn't exist or didn't run. Quality drifted, rework piled up, and the ROI quietly went underwater.

None of these is a model problem. All of them are decisions the business either made badly or never made at all.

Success Criteria: The 41% Problem

The single largest cause of production agents losing money is that no one defined what winning looks like. Forrester puts it at 41% of the failure cohort.

This is the same trap Klarna walked into in 2024. It rolled out a customer-service agent that handled 2.3 million conversations in its first month and cut resolution times from eleven minutes to two. The projected savings were enormous. Then customers revolted over generic answers and robotic tone, and the company started rehiring the humans it had let go.

The agent wasn't broken. It was extraordinarily good at the objective it was given — resolve tickets fast — which turned out to be the wrong objective. Klarna's real intent was to build lasting customer relationships that drive lifetime value. Those are profoundly different goals, and they demand profoundly different decisions at the point of interaction.

A five-year veteran on the support team knows when to bend a policy, when to spend three extra minutes because a customer's tone says they're about to churn, and when speed is the right call versus when generosity is. She absorbed that judgment over years. The agent had a prompt. It had context. It did not have intent.

What good looks like

A well-scoped agent objective isn't a mission statement. It's a specification the agent can act on:

• The signal. What data indicates success in our context? Not "customer satisfaction" — the actual measurable proxies, and their limits.
• The authorized actions. What is the agent allowed to do to move that signal?
• The trade-offs. Speed versus thoroughness, cost versus quality — resolved explicitly, so the agent isn't guessing.
• The hard boundaries. The lines it may not cross, and the situations it must escalate instead of resolving.

What bad looks like

"Improve our support metrics." That's it. That's the whole brief. The agent picks the metric that's easiest to move — resolution time — and optimizes it into the ground while the metric that actually mattered, retention, quietly erodes. Everyone celebrates the dashboard until the churn report lands.

The difference between these two isn't the model. It's whether someone did the work of turning organizational purpose into machine-actionable parameters before deployment. That work has a name: intent engineering. We wrote about the discipline itself in why 74% of businesses see no ROI from AI — the intent gap is the through-line.

Tool And Data Access: The 33% Problem

The second failure mode is structural. A third of the losing cohort deployed agents that couldn't reach the tools or data they needed to do the job.

This sounds like a technical oversight, and sometimes it is. More often it's an architecture decision made by default rather than on purpose. The agent was built on a platform where connecting a new system means a support ticket, a re-deploy, or a workaround — so it shipped without the connections it needed, and the value it was supposed to create never materialized.

This is where the runtime your agents live on stops being a detail and starts being the whole story. An agent that can't durably hold credentials, can't reach your CRM and your ticketing system and your document store in the same session, and can't have new integrations added without tearing down and rebuilding — that agent is set up to under-deliver from day one.

What good looks like

The agent runs on persistent infrastructure with a real filesystem, durable memory, and the ability to connect to the systems it needs — securely, and without re-architecting every time the job expands. When a workflow grows, you add the tool. The agent keeps its context. Nothing gets re-cloned or reset.

What bad looks like

The agent runs in an ephemeral container that re-clones its environment every session, holds no durable state, and requires a static, hardcoded configuration to touch anything outside its sandbox. Every new data source is a project. So most of them never get connected, and the agent limps along doing a fraction of what it was scoped for. We've written before about why persistent agents beat ephemeral sessions for exactly this reason — the architecture determines the ceiling.

Evaluation Drift: The 26% Problem

The third failure mode is the quietest and, in some ways, the most dangerous. A quarter of losing agents drifted because their evaluation coverage decayed over time.

Here's the stat that makes it concrete. In Forrester's panel, agents without automated evaluations on every prompt change had a 47% rollback rate. Agents with full evaluation coverage had a 9% rollback rate. Evaluation coverage was the single strongest predictor of whether a production agent was still running twelve months later.

Think about what that means. Whether your agent survives has almost nothing to do with how smart it was on launch day. It has everything to do with whether you built the discipline to catch it when it drifted — when a prompt got tweaked, a model got upgraded, or a workflow changed and quietly broke something downstream.

Agents are not "set and forget." A long-running agent behaves less like a chatbot and more like a distributed system, and distributed systems demand monitoring, testing, and the ability to catch regressions before they cost you. Most teams aren't doing that at all.

What good looks like

Every change to an agent's behavior — a new prompt, a new model, a new tool — triggers an evaluation against a known bar before it goes live. When something drifts, you know within hours, not when the quarterly ROI review reveals the damage. And because the system is steerable, the agent can pause on ambiguity, ask a human for judgment, and resume without starting over.

What bad looks like

Someone edits a prompt to fix one edge case. It silently degrades three others. Nobody notices for six weeks because there's no test suite and no monitoring. By the time the numbers surface the regression, the rework has already eaten the agent's entire ROI. That's a rollback waiting to happen — and it's the 47% cohort in a sentence.

The Real Divide: Behavioral Safety vs. Structural Safety

Step back and the three failure modes point at one root cause. Most organizations try to make agents create value by telling them what to do and hoping they behave. That's behavioral safety, and it's brittle. Prompts drift, objectives get misread, edge cases slip through.

The alternative is structural safety: building systems where the agent can't do the wrong thing because the boundaries are encoded into the architecture, not written as suggestions in a prompt. Success criteria the agent is measured against. Tool access that's governed and scoped. Evaluation gates the agent has to pass before a change ships.

Behavioral safety is a sticky note that says "please don't." Structural safety is a locked door. The 22% who lose money almost always chose the sticky note — usually without realizing there was a choice to make.

This is what separates deploying an AI tool from running an AI coworker. A tool does what you tell it in the moment. A coworker operates inside a role with defined objectives, real access to your systems, durable memory of what happened before, and boundaries it can't cross. That's not a prompt. That's an operating layer.

What To Do Before Your Next Agent Goes Live

The 22% is avoidable. Every one of Forrester's failure modes is a decision you can make well instead of by default. Here's the pre-deployment checklist we run:

1. Write a binary success definition. Before anything ships, answer: how will we know in ninety days whether this agent made or lost money? If you can't state it as a measurable outcome, you're not ready to deploy — you're ready to plan.
2. Map every tool and data source the job actually requires. Then confirm the agent can reach all of them, securely, on the infrastructure you're using. If adding a connection is a project, fix the platform before you fix the agent.
3. Build the evaluation gate first. Define the bar the agent has to clear, and wire it to run on every prompt, model, or workflow change. This one control moved rollback rates from 47% to 9% in the data. It is the highest-return control you will build.
4. Encode the trade-offs and boundaries. Speed versus quality, when to act versus when to escalate, the lines it may not cross. Make organizational judgment explicit and machine-actionable, because the agent will not absorb it through osmosis the way a human hire does.
5. Name an owner. Forrester found organizations with a named agent owner had a 2.7x higher pilot-to-production conversion rate and were underrepresented in the losing cohort. An agent without an owner drifts. Someone has to be accountable for its ROI.

Do these five things and you're not gambling on being in the 78%. You're engineering your way out of the 22% on purpose.

FAQ

Q: Why do AI agents lose money in production even when the model works well? A: Because model quality was never the bottleneck. Forrester's 2026 root-cause analysis of negative-ROI agents found the failures came from unclear success criteria (41%), insufficient tool or data access (33%), and evaluation drift (26%) — all operational and governance issues the organization controls, not model limitations.

Q: What is the single biggest predictor of whether a production agent survives? A: Evaluation coverage. Agents with automated evaluations on every prompt change had a 9% rollback rate; those without had 47%. Whether an agent is still running twelve months later depends far more on your testing discipline than on how capable it was at launch.

Q: How is "unclear success criteria" different from just having a goal? A: A goal like "improve customer service" is a human-readable aspiration. An agent needs a machine-actionable objective: the specific signals that indicate success, the actions it's authorized to take, the trade-offs it can make, and the boundaries it can't cross. Without that translation, the agent optimizes for whatever is easiest to measure — often the wrong thing.

Q: Does switching to a better AI model fix negative ROI? A: Rarely. If the agent has no clear success definition, can't reach the data it needs, or has no evaluation coverage, a smarter model just does the wrong thing more efficiently. The fixes are structural: define success, govern access, and build evaluation gates before you deploy.

Q: What's the difference between behavioral safety and structural safety for agents? A: Behavioral safety means telling the agent what to do and hoping it complies — brittle, because prompts drift and edge cases slip through. Structural safety means building systems where the agent can't do the wrong thing because the boundaries are encoded into the architecture. The agents that lose money almost always relied on behavioral safety alone.

Run a Real Team, Not a Gamble

The businesses that stay out of the 22% aren't the ones with the best models. They're the ones who defined success, governed access, and built evaluation discipline before turning anything on. That's the difference between deploying an AI tool and running an AI coworker with a real role, durable memory, and boundaries it can't cross. If you're ready to stop using AI tools and start running a real team of AI coworkers, Associates AI Teammates gives you a 14-day free trial with no credit card required. Start your free trial at associatesai.team.