AI Agents, Real Work, and the Return You’re Not Getting Yet
Everyone is using AI. Almost nobody is getting return on it. There is a reason for that, and it is not the first time the story has been told.
MIT researchers tracked enterprise AI deployments in 2025. Ninety-five percent of pilots produced zero measurable P&L impact. The RAND Corporation put the AI project failure rate at 80.3 percent. Surveys of knowledge workers find that 85 percent of what people do with AI generates no measurable business value. Meanwhile, the C-suite is 31 percent more likely than the people actually doing the work to believe that widespread AI adoption has already been achieved.
I have heard this story before. The details are different. The structure is identical.
The Third Time Around
In the 1880s, factories began replacing steam engines with electric motors. Productivity barely moved for forty years. Not because electricity was a weak technology — it was transformative — but because factory managers kept the same floor layout, the same production logic, the same organizational structure designed for steam. They swapped the power source and called it modernization. The capability was real. The redesign never happened.
By the 1990s, the new technology was the personal computer. Organizations automated their paper processes. They digitized filing cabinets. They put old forms on screens. ERP systems encoded existing dysfunction — the handoffs, the approval chains, the departmental silos — into expensive software. By 1994, sixty-nine percent of Fortune 500 companies had a Business Process Reengineering initiative underway. The consulting industry built a fifty-one billion dollar business helping companies fix the mess. Seventy percent of those reengineering efforts failed. The movement’s founders issued public apologies. Michael Hammer, who coined the phrase “don’t automate, obliterate,” admitted to the Wall Street Journal that he had been “insufficiently appreciative of the human dimension.”
The economists had a name for the PC-era version of this: the Solow Paradox. Robert Solow observed in 1987 that you could see the computer age everywhere but in the productivity statistics.
Today, the quote fits AI so precisely it could have been written last week.
The root cause in each wave is the same: organizations apply new technology to existing processes instead of redesigning the processes for the new capability. McKinsey’s 2025 data quantifies the gap for AI directly — companies seeing real business impact from AI are three times more likely to have fundamentally redesigned their workflows. Not using more AI. Redesigning the company around it.
The discipline that has always navigated this correctly is lean. The Toyota Production System — documented the same year Hammer published his BPR manifesto — was built on a different premise: develop your people as problem solvers, understand your value streams end-to-end, and improve the work continuously from the inside. Toyota did not need BPR. It was the counterexample hiding in plain sight. Improvement before automation. Design before deployment.
That principle is not nostalgic. It is the operating manual for getting AI to actually work.
Enter the Agent. But What Is an Agent?
Here is where the current AI conversation gets genuinely complicated, and where the definitions matter enough to spend a moment on them.
A copilot responds to prompts. You ask, it answers. It is useful the way a fast reference librarian is useful — dramatically better than searching alone, but only producing output when you show up and ask. Most AI use today is at this level: drafting, summarizing, answering questions. The productivity is real. The business transformation is not.
An agent is different in kind, not just in degree. An agent can plan and execute. It can take a goal — not just a question — and work toward it across multiple steps, using tools, querying systems, and making decisions along the way. It can operate when you are not watching. It can hand off to other agents.
In practice, “agent” has become a marketing term. Vendors rebrand existing automation as agentic. Organizations apply large language models to tasks where a simple rule, a SQL query, or an RPA bot would be faster, cheaper, and more reliable. The industry has a name for this: agent washing. It produces the same disappointment as every other oversold technology wave — just faster.
The honest picture of where agentic AI actually is: Deloitte’s 2025 Emerging Technology Trends study found that thirty percent of surveyed organizations are exploring agentic options, thirty-eight percent are piloting solutions, fourteen percent have solutions ready to deploy, and eleven percent are actively using agentic AI in production. On the strategy side, forty-two percent are still developing their agentic roadmap, and thirty-five percent have no formal strategy at all.
Gartner has predicted that more than forty percent of agentic AI projects will fail by 2027 — not because the technology is weak, but because legacy systems were not designed to support the real-time, continuous execution that agents require.
Why the Gap Is Largest Outside Content and Code
If you have seen agents in action, you have almost certainly seen them in two settings: writing content or writing code. These are the domains where agents have made the most visible, credible progress — and for good reasons. The tasks are bounded, the output is reviewable, the cost of errors is recoverable. A coding agent that writes a function with a bug is not catastrophic. A content agent that produces a weak draft is not catastrophic.
The moment you cross into operational processes — the kind that live in ERP systems, affect financial records, touch payroll, or produce client deliverables — the requirements change entirely. Accuracy matters differently when a wrong journal entry requires correction work that takes longer than the original task. Auditability matters when professional responsibility is on the line. Reliability matters when your month-end close cannot afford a workflow that “usually works if you try it a few times.”
A 2026 benchmark that tested nineteen AI models across 101 real accounting workflows found that the best performer — the most capable model available — achieved seventy-nine percent accuracy. That means it fails one in five accounting tasks. For writing a draft blog post, seventy-nine percent is a reasonable starting point. For entering transactions, it is not a production-ready number.
The architecture that makes agents trustworthy in operational settings is different from what makes them useful in creative work. It requires deterministic execution — the agent does what was approved, in the steps that were defined, with a full audit trail. It requires process redesign that eliminates steps built around human constraints and defines explicit human judgment points. It requires the underlying systems to support real-time, continuous agent operation rather than session-based interaction.
None of that is impossible. Some of it is already working in practice. Toyota built an agent that handles vehicle tracking from pre-manufacturing through dealership delivery and is now drafting supply chain resolution emails before staff arrive in the morning. A Fortune 500 technology company built a performance review agent that assembles data analysis and structured reports from four specialized underlying agents. The examples are real. They are also still rare enough to be notable.
The Foundation Comes First
This is where I am spending my time right now.
The path to agents doing real operational work does not start with the agent. It starts with the operating system underneath — the identity, context, skills, and memory that make any agent you build smarter, more reliable, and faster to deploy. An agent built on a well-designed foundation inherits everything that foundation has learned. An agent built without one starts from scratch every time and fails at the seams when the environment changes.
I am building that foundation for NaylorCPA, and I am learning it in public. Not because the process has been clean — it has not — but because the messy middle is exactly where the useful lessons live. I have navigated a hype cycle before, in a technology with real parallels to this one. The discipline that made RPA work where it worked, and the mistakes that made it fail where it failed, are directly applicable here.
The posts that follow in this series document what I am actually building, the specific problems I have run into, and what I have learned from fixing them. Not a framework. Not a vendor pitch. A practitioner’s field notes from inside the build.
If you are an operations leader trying to figure out what is real about AI agents — what is worth your time and what is noise — these are the field notes.
This is the first post in a series on building an Agentic OS for a solo operations practice. The series covers the framework, the build process, and the lessons that only appear when you do the work.