tracker_v7_FINAL_real.xlsx and the Analyst Who Makes Organizations Honest

Every enterprise that wants to deploy AI agents eventually hires someone like Nora Ledgerwood. Her actual title varies by engagement — "AI Readiness Analyst," "Process Formalization Lead," once, memorably, "Workflow Archaeologist" on a contract she wrote herself. Before she started preparing enterprise workflows for agent deployment, she spent eight years as a forensic accountant, tracing the distance between what financial records said happened and what actually happened. She describes the career change as "lateral."

We spoke with Ledgerwood — a composite drawn from the experiences of practitioners working in this emerging and largely unnamed discipline — over video, her background a wall of sticky notes in four colors she declined to explain.

You were a forensic accountant. How does someone go from investigating financial irregularities to preparing workflows for AI agents?

Nora: It's the same job. In forensic accounting, you look at the official record and then figure out what actually happened. Now I look at the documented process and figure out what people actually do. The gap is where I live. In forensic accounting, the gap was usually fraud. Here it's usually competence. People deviated from the process because the process was wrong, and they were quietly making things work anyway.

What does "preparing a workflow for an agent" actually involve?

Nora: The pitch is: take your existing business process, formalize it, hand it to an agent. Simple. Three months, maybe four.

What actually happens is I walk into a room with six stakeholders and ask, "Where does this workflow start?" I get four different answers. Not because anyone is confused. Because they each experience a different starting point. The person in procurement thinks it starts when the purchase request hits their queue. The person in finance thinks it starts when budget approval happens, which is three steps earlier. The requester thinks it started when they sent a Slack message to their manager two weeks ago. Technically, they're all right.

That sounds like a communication problem.

Nora: It sounds like one. But organizations run on ambiguity. That's not a criticism. Ambiguity is efficient when you have humans. A human gets a purchase request missing the cost center code, and they know — from experience, from context, from having sat next to Janet for three years — which cost center it should be. They fill it in. Nobody documents this. The work just flows.

The process-mining literature has a term for this gap. One study compared documented clinical guidelines against actual care sequences in an ICU — 187 patient cases — and found fitness scores of 0.69 and 0.66.¹ Material gaps, existing for completely legitimate reasons. Patient variability, resource constraints, expert judgment. The same thing happens in every enterprise, just with purchase orders instead of patients.

So what happens when you try to formalize that for an agent?

Nora: Everything breaks. Not the technology. The organization.

I have five questions I ask in every engagement, and each one is a landmine.

"Where does this start?" — we covered that. "Who owns this approval step?" is worse, because it surfaces the difference between the person on the org chart and the person everyone actually goes to. "What counts as done?" is the one that starts arguments. Operations says done means the system updated. Compliance says done means the audit trail is complete. The customer-facing team says done means confirmation was sent. These are three different moments in time, and the agent can't hold all three definitions simultaneously.

Then there's "which exceptions are routine?" My favorite, in a grim way. I'll find a workaround that operations has been running every single week for four years. Not in any documentation. When I surface it, we have to decide: does this become official policy — which requires approval from people who've never heard of it — or does it get eliminated, which means telling the team their workaround is now forbidden? Neither option makes friends.

You mentioned five questions. What's the fifth?

Nora: "What happens when this goes wrong?" Recovery paths. Almost never documented. When humans handle failures, the failure becomes invisible — someone fixes it, moves on, nobody writes it down. An agent needs an explicit recovery path before it encounters the failure. You can't teach it to "just figure it out" the way Janet does.

This sounds like it could get politically uncomfortable.

Nora: Could?

Every question I ask is a governance question wearing a technical costume. When I ask "who owns this step," I'm really asking "who has authority here," and sometimes the answer is that two VPs both think they do. They've coexisted peacefully for years because nobody ever forced the question. I show up with a process diagram and suddenly we're in a jurisdictional dispute.

The BPM governance research is pretty clear on this. Studies of public organizations found that none had a formalized governance model, and senior leadership didn't endorse enterprise-wide frameworks.² Process ownership is recommended by every framework and contested in every implementation.³

I had one engagement where — I can't say too much — but the "process" turned out to live entirely in a personal spreadsheet on one person's laptop. Millions of dollars of work, routed through a file called tracker_v7_FINAL_real.xlsx. And the thing is, it worked. That person was brilliant. They'd built a system that handled every exception, every edge case. But it was completely undelegable. You can't hand that to an agent because you can't even describe what it does without that person in the room, narrating.

The RPA wave went through something similar, didn't it?

Nora: Same problems, almost beat for beat. Business rules living in people's heads. Exceptions dominating real work. Ownership dissolving when automation crosses team boundaries.⁴ I sometimes feel like the latest person to rediscover an old problem, just with higher stakes.

When an RPA bot hit an exception it couldn't handle, it stopped. Full stop. When an agent hits ambiguity, it might do something plausible but wrong. That's a fundamentally different failure mode. A bot that freezes is annoying. An agent that confidently proceeds in the wrong direction is dangerous. A recent study found companies with advanced experimental AI capabilities stuck at basic deployment levels because they couldn't verify outputs — human review remained the only trusted check.⁵

What does success look like for you?

Nora: A really boring document. AWS published guidance that nailed it — in organizations where agents create visible value, "the work is defined in painful detail."⁶

That's what I produce. Painful detail. Step by step: what arrives, what happens, what done means, what happens when things go wrong. Nobody's going to put "painful detail" on a conference slide. But that's the deliverable.

Do you think most organizations are ready for this?

Nora: No. Deloitte found that only 14% of organizations have agentic solutions ready to deploy, with 42% still developing their strategy roadmap.⁷ The gap between "we want agents" and "we can describe our work precisely enough to delegate it" is enormous. More compute doesn't close it. More honesty might.

Last question. Do you enjoy this work?

Nora: …Yes?

It's strange work. I'm essentially asking organizations to be honest with themselves about how they actually operate. Most of them have never had to do that before. Humans absorbed the ambiguity. Agents can't.

Footnotes

1. Benevento et al. (2022), cited in TinyFish KB on process intelligence. COVID-19 ICU process-mining study at Uniklinik Aachen comparing guideline-derived BPMN models with actual event-log behavior. ↩

2. Santana et al., "BPM Governance: An Exploratory Study in Public Organizations." https://www.researchgate.net/publication/220921346 ↩

3. "Five Key Elements to BPM Governance," The Leonardo Blog. https://blog.leonardo.com.au/5-key-elements-to-bpm-governance ↩

4. TinyFish KB, "Historical Echoes of Automation." ↩

5. Bayer et al. (2026), "Agentic AI in Industry: Adoption Level and Deployment Barriers." https://arxiv.org/abs/2605.14675 ↩

6. "Operationalizing Agentic AI Part 1: A Stakeholder's Guide," AWS Machine Learning Blog. https://aws.amazon.com/blogs/machine-learning/operationalizing-agentic-ai-part-1-a-stakeholders-guide/ ↩

7. Deloitte, "Agentic Reality Check: Preparing for a Silicon-Based Workforce." https://www.deloitte.com/us/en/insights/topics/technology-management/tech-trends/2026/agentic-ai-strategy.html ↩