The rules of enterprise technology are being rewritten again, and most leaders are walking into the exact same trap that buried the ERP wave, the data warehouse wave, and the cloud wave. The question in 2026 is no longer “How do we ERP this?” or “How do we cloud this?” It is “How do we agent this?” And it is the same conversation, which means we are hurtling toward the same failure mode.
The short answer: AI process improvement means using AI to accelerate a proven improvement method, mapping the process, finding the root cause, fixing the workflow, before you decide what to automate. Skip that step and AI does not fix a broken process; it runs the dysfunction faster. The enterprises that win with AI are the ones that ruthlessly refuse to automate a process they do not yet understand.
Key Takeaways
- Automation magnifies whatever it touches. Point it at an efficient process and you gain efficiency; point it at a broken one and you industrialize the dysfunction.
- AI’s rising failure rate is a process gap, not a technology gap. Abandonment of most AI initiatives jumped from 17% to 42% in a single year.
- AI inherits your measurement system. Optimize toward a flawed KPI and the agent will destroy adjacent value to hit the number.
- Fix the workflow first. Cleaning up chaos costs linearly with humans and exponentially with agents. Get the order right.
- The AIM-IT Framework (Assess, Innovate, Model, Implement, Track) forces clarity before code: Problem and People First, Then Technology.
The Same Room, The Same Question
I am standing in the same room I have stood in for thirty years. The logos on the wall change, IBM, Ernst & Young, Johnson & Johnson, McKesson, but the executive at the table is the same, and the question is exactly the same, just reframed for a new technology cycle.
In 1995, the question was, “How do we ERP this?” In 2010, it was, “How do we cloud this?” In 2026, it is, “How do we agent this?”
It is the same conversation, which means we are hurtling toward the exact same failure mode. Bill Gates defined the philosophical spine of enterprise technology decades ago: automation applied to an efficient operation magnifies the efficiency, and automation applied to an inefficient operation magnifies the inefficiency.
AI does not break that law. It accelerates it.
The Process Gap Behind AI Failure
According to S&P Global Market Intelligence’s 2025 enterprise survey, the share of companies abandoning most of their AI initiatives jumped from 17% in 2024 to 42% in 2025. That is not a technology gap. That is a process gap.
To stop the bleeding, we have to draw hard lines between three things:
- Automating chaos: deploying AI, robotic process automation (RPA), or agents on top of a process that is undocumented, inconsistent, or measurement-blind. This locks in organizational dysfunction at machine speed.
- AI-powered process improvement: using AI to accelerate proven methodologies, Lean Six Sigma DMAIC, Kaizen, Design Thinking, First Principles, TRIZ, before deciding what to automate. This ensures the work that gets automated is the right work, executed the right way.
- The AIM-IT Framework: a five-phase methodology (Assess, Innovate, Model, Implement, Track) that forces operational clarity before automation. Its foundational first principle is Problem and People First, Then Technology.
You cannot automate your way out of a process you do not understand. AI just gives you the rope faster.
Why AI Magnifies Chaos Instead of Curing It
When leaders buy AI, they assume the intelligence of the model will somehow compensate for the dysfunction of the workflow. It will not. AI magnifies chaos for three structural reasons.
1. AI has no theory of the work. A large language model can write a SQL query, generate a contract clause, or summarize a customer call. But it has no opinion on whether that underlying process should exist in the first place. It simply makes the existing, potentially wasteful process faster.
2. AI inherits your measurement system. If your KPI is wrong, your AI will optimize toward the wrong thing with breathtaking efficiency. As Goodhart’s Law dictates: when a measure becomes a target, it ceases to be a good measure. An agent maximizing a flawed metric will destroy adjacent business value to hit its goal.
3. AI multiplies invisible handoffs. Every undocumented handoff in your current manual process becomes a prime hallucination opportunity in your agentic future. Without rigid workflow definitions, agents are excellent at confidently executing the wrong next step.
The cost of cleaning up a chaotic process scales linearly when humans do it, and exponentially when agents do it. Get the order right.
Field Story: The Healthcare RCM Replacement That Didn’t Eat Its Own Tail
Recently, a national healthcare operator with over 800 facilities set out to replace a legacy revenue cycle platform. The vision was a unified Patient Access & Revenue Cycle Management (RCM) system integrating their EHR, clinical documentation, AI documentation, and CMS portal connectivity.
The natural temptation, and exactly what every vendor in the building pitched, was to rip out the legacy system, plug in the new platform, sprinkle generative AI over the claims and prior-authorization steps, and ship it.
Instead, we ran a rigorous root-cause analysis (RCA) before touching a single line of agent design. We mapped over a dozen distinct RCM process areas end to end, patient access, registration, insurance verification, financial counseling, prior authorization, enterprise verification, documentation, charge capture, claim submission, payment posting, denial management, collections, and reporting, and catalogued more than 42 documented pain points / defects in the current state.
We then ran six 5-Whys analyses, organizing the failures into six causal loops: Data Quality, System Blindness, Handoff Fragmentation, Uncontrolled Variation, Unbounded Retries, and Measurement Void. Using First Principles and the Theory of Constraints, we located the true operational chokepoint: a missing point-of-authorization step that nobody actually owned.
Finally, we built a traceability matrix linking every defect to its root-cause group, to a functional requirement, and ultimately to a solution component. Only then did we define where AI agents would do work and where humans would keep decision ownership. The AI agent surface area shrank by roughly two-thirds from the vendor’s original proposal. The retained scope was higher-value, lower-risk, and defensibly auditable for healthcare compliance.
We were prepared to replace a 12-step revenue cycle with AI. Instead we found that three steps didn’t belong in the workflow at all, four needed redesign before any automation, and only five were ready for an agent.
That’s the difference between deploying AI and digitizing dysfunction.
Field Story: 110 AI Use Cases, Only 9 Made the Cut
The C-suite of a global edtech, serving over 100 university partners, asked the question every board is asking right now: “Where should we apply AI first?” The standard consultancy answer, “Here are fifteen pilots, pick three”, is exactly the playbook responsible for the 42% abandonment rate.
We took a different path. We identified over 120 candidate AI use cases across the enterprise, call center, recruiting, contracts, faculty operations, marketing, student success. We scored each on an effort-versus-impact matrix, and, more importantly, we stress-tested the underlying process for documentation, data quality, and ownership clarity before selecting any technology.
We killed roughly 110 of those use cases at the gate, not because the AI engineering was too hard, but because the underlying processes were not ready to be amplified. We scaled nine proofs-of-concept to production with disciplined gating. Within the first year, four went fully live, delivering measurable, hard-dollar impact. The portfolio that emerged was not a vendor demo. It was a disciplined prioritization matrix wearing a generative AI jacket.
For every AI use case you ship, kill ten, not because AI doesn’t work, but because most processes aren’t worth automating yet.
Like comedy, when it comes to operational AI execution, “timing is everything.” The things we can do with AI in 2026, was not possible in 2025.
Field Story: When the Number Itself Was Lying
A national tech infrastructure company asked us to build a new executive scorecard on Azure Databricks. This engagement proved that measurement is an active part of the process, not just an output.
A flagship IT SLA KPI that was off by nearly 50% of the true value. The natural reflex, and a software vendor’s preferred reflex, would have been to deploy an AI agent to “auto-resolve” tickets and pump the metric back into the green.
Instead, we ran a structured hypothesis elimination across people, process, and measurement. We discovered the API extract was missing a critical include=stats parameter, and the data warehouse was measuring updated_at (an auto-close function), not the true resolved_at timestamp. We closed the gap to within ±2 percentage points of the source-of-truth scorecard without altering a single step of the underlying process.
If we had built an AI remediation agent on top of that scorecard, the agent would have spent its first quarter “fixing” a problem that did not exist, burning engineering hours, generating audit risk, and eroding executive trust in the AI program before it ever delivered.
The fastest way to discredit your AI program is to deploy an agent against a broken metric. The fastest way to accelerate it is to fix the metric first.
Field Story: The $50M Process Pain Worth Automating
There is an inverse to these warnings. A well-know case study at WorkOS of Lumen Technologies, found sales teams were spending an average of four hours per customer outreach call on background research, an estimated $50 million annual drag on revenue productivity.
After building an AI Copilot grounded in clean, structured, validated customer context, that research time dropped from four hours to 15 minutes per call. The reason it succeeded where others fail is sequencing: the company quantified a specific, well-documented process pain first, then matched the exact AI capability to it. The Copilot did not automate chaos; it accelerated a well-understood, properly measured workflow.
The best AI deployments aren’t the ones with the most impressive models. They’re the ones that started with the most honest business case.
AI Automation vs. AI Process Improvement
This is not a hypothetical comparison. These are field-observed patterns across more than 30 Fortune 500 engagements, validated by the 300+ practitioners I have trained.
| Dimension | Automating Chaos | AI-Powered Process Improvement |
|---|---|---|
| Starting point | “How can we use ChatGPT?” | “What’s bleeding money in our operations?” |
| First deliverable | AI pilot demo | Process map + defect catalog + RCA |
| Methodology | Vendor playbook | Structured RCA + AIM-IT |
| Measurement | “Did the pilot work?” | “Is the metric itself trustworthy?” |
| Governance | After-the-fact | HITL/HOTL designed in the Implement phase |
| Common outcome | Pilot dies in production | 92% pilot-to-production rate (field-observed) |
| Typical timeline | 6—12 months to disappointment | 30-90 days to first measurable win |
| 2025 industry pattern | 42% abandonment rate | 15-60% productivity gains, phased |
The AIM-IT Framework: The Five-Phase Pattern That Stops Chaos
To reliably bridge legacy process discipline and agentic automation, I deploy the AIM-IT Framework. It forces clarity before code.
Assess
Define the problem in strict SMART terms. Tag your leading and lagging metrics. Map the current-state workflow. Surface the actual root causes via the 5 Whys, causal loops, and First Principles. The rule here is absolute: problem first, not technology first.
Innovate
Generate 14 to 28 alternative solutions across four lanes: traditional process improvement, AI/ML, GenAI/agentic AI, and AI-powered traditional approaches. Narrow with SCAMPER, TRIZ, and an effort-versus-impact matrix. Do not pre-commit to an AI solution if a process tweak solves the defect.
Model
Design the solution architecture by type. Build Human-in-the-Loop (HITL) and Human-on-the-Loop (HOTL) governance directly into the design, it cannot be bolted on after launch. Pilot in a strictly scoped MVP using a golden dataset. Governance is a core design constraint, not a compliance afterthought.
Implement
Execute workflow integration, establish guardrails, and build the evaluation layers. Draft standard operating procedures for both the humans and the AI. Run a three-track organizational change program covering training, communications, and executive sponsorship. The hardest engineering in this phase is human engineering.
Track
Monitor KPIs that tie back to the original Assess-phase metrics. Run continuous AI evaluations for hallucination, drift, and bias. Use tools like DSPy for ongoing prompt optimization. Feed the insights back into the Innovate phase. The loop must close, or the AI will degrade.
Problem and People First, Then Technology. That is the difference between automating chaos and improving the process.
Fix the Workflow First
Over three decades, I have watched the enterprise technology wave crest and break repeatedly, the ERP wave, the data warehouse wave, the cloud wave, and now the AI wave. Earning the IBM Golden Circle Award and the Ernst & Young Reengineering Award taught me a truth that has survived every cycle: the companies that get it right are the ones that ruthlessly refuse to automate a process they do not understand.
I have trained over 300 practitioners to hold that line. AI is the most powerful accelerant we have ever built, but it is just that, an accelerant.
Fix the workflow before you automate the chaos. Don’t change what you improve; change how fast you improve it. That is the AIM-IT discipline, and that is what separates the teams that ship from the 42% that quietly walk away.
For more on Human-First AI and enterprise transformation, visit Analytics AIML.
Saravá.
Frank “Rio” Shines
Frequently Asked Questions (FAQs)
What does “automating chaos” mean?
Automating chaos means deploying AI, RPA, or agents on top of a process that is undocumented, inconsistent, or measurement-blind. Instead of fixing the dysfunction, it locks it in and runs it faster. As Bill Gates warned, automation applied to an inefficient operation only magnifies the inefficiency.
What is the AIM-IT Framework?
The AIM-IT Framework is a five-phase methodology, Assess, Innovate, Model, Implement, Track, that forces operational clarity before automation. Its first principle is “Problem and People First, Then Technology.” It bridges decades of process discipline to modern agentic AI so teams solve the right root causes before they scale.
How does Goodhart’s Law apply to AI?
Goodhart’s Law holds that when a measure becomes a target, it ceases to be a good measure. An AI agent maximizing a flawed KPI will pursue it with breathtaking efficiency, destroying adjacent business value to hit a number. If your measurement system is wrong, AI optimizes toward the wrong thing, faster.
What is the difference between HITL and HOTL governance?
Human-in-the-Loop (HITL) keeps a person in the decision path, approving each action. Human-on-the-Loop (HOTL) lets the agent act while a human monitors and can intervene. Both must be designed into the Model phase as core constraints, not bolted on as a compliance afterthought after launch.