Insights

Field notes from inside the engine room

Long-form pieces on AI readiness, governance, and the operational architecture mid-market companies need before AI investment pays back. Built from 25+ years of operating, not from a lab.

From the Field · Field Note

You Can't Automate What You Haven't Heard

Published March 2026 · Last updated June 2026

Field notes from a diagnose phase, and the tool we built to listen better.

Most AI consulting starts in the wrong room. It starts with the technology, a demo, a platform, a list of tools looking for somewhere to be useful. The conversation jumps straight to what could we automate before anyone has earned the right to answer it.

I learned a long time ago, managing $400M+ in indirect spend and the people behind it, that you can't fix what you can't see. And the work that actually drains a mid-market business, the low-value work, is almost never written down. It lives in the heads of your most experienced people. It's the judgment call a procurement lead makes in three seconds and couldn't fully explain if you asked. It's the workaround a specialist invented years ago that the whole team now quietly depends on. That knowledge is the asset. It's also invisible. And you can't hand invisible work to an AI.

So before we recommend a single enablement, we do the unglamorous thing first: we listen.

The problem with listening at scale

Here's the catch. Listening properly is expensive. To document how a team really works, you sit with people across multiple conversations, you let them walk you through the messy reality, and you ask the same question five different ways until the actual decision logic surfaces. Then someone has to turn hours of rambling, valuable, human conversation into something structured enough to act on.

For years I did that by hand: notebooks, templates, transcripts I re-read at midnight. It worked, but it didn't scale, and the best insights were often the ones that slipped through the cracks between sessions.

So we built a tool to carry that weight. Not to replace the listening, but to make it possible to listen more, and lose less of it.

What the tool actually does

It is, deliberately, not clever for the sake of it. It does four things, in order: the same way I would.

It captures the conversations. Across multiple working sessions with a client's team, it takes in every transcript, the full, unedited reality of how the work gets done.
It flags what it can't see clearly. Where a step is vague or a decision goes unexplained, it surfaces the gap instead of papering over it, so we go back and get the real answer instead of guessing.
It documents the workflows. It turns those conversations into clean, structured records: the trigger, every step, the system touched, the decision rule, the handoff, the pain point. The institutional wisdom, finally written down.
It assesses the opportunity. It scores each step for AI potential and produces a prioritized list of enablements, ranked by value recovered, effort, and risk, with the high-stakes steps flagged to keep a human firmly in the loop.

This is Phase 1, Diagnose, and it now runs as an instrument instead of an ordeal.

The client experience

The client was a medical billing company whose senior people were doing exactly what senior people everywhere are doing: spending a third of their week on manual friction that had nothing to do with why they were hired.

Over three conversations, the tool captured how the work actually moved, not the org-chart version, the real one. What came back wasn't a tidy story. It was the truth: multiple distinct workflows, each carrying its own quiet leakage, mapped end to end.

Two things happened that I want to be precise about.

First, the team felt heard. There's a moment in every Diagnose when a long-tenured expert realizes the goal isn't to replace them, it's to write down what they know so the organization stops being one resignation away from losing it. That moment changes the room. People stop defending their work and start improving it.

Second, we ended with a decision, not a wish list. Instead of a vague sense that "we should do something with AI," the client had a ranked roadmap: here is where the leakage is, here is the time and money waiting in each one, here is what to do first, and here is what we will deliberately not automate, because a human judgment belongs there.

The principle underneath the tool

I'm wary of tools that promise to think for you. This one doesn't. It thinks with you, and it earns its keep by doing the part that doesn't scale, capturing human work faithfully, so the expensive human judgment can go where it actually matters.

That's the whole philosophy, and it's the one I'll keep building on: AI should serve people, not replace them. Diagnose before you deploy. Capture the wisdom before you automate the task. The companies that get the order right are the ones for whom AI investment finally pays back.

The ones who start in the technology room are still waiting.

Paula Taylor is CEO & Founder of AI Edge Strategy. She spent 25+ years in enterprise operations, including as Global Director of Indirect Procurement at NXP Semiconductors, where she managed $400M+ in indirect spend and a 10,000-contractor contingent workforce program.

From the Field · Field Note

The Low-Value Work Audit Nobody Runs

Published March 2026

Why the 30–40% of senior expertise spent on document review is the most under-measured cost line in mid-market ops, and what to do about it.

Every mid-market operation I've ever walked into has a number for software. A number for the people on payroll. A number for travel, for facilities, for the vendor nobody remembers hiring. What it almost never has is a number for the most expensive thing in the building, which is senior judgment spent on work that never needed it.

Here's the thing. Nobody's hiding this on purpose. It just doesn't live anywhere. There's no invoice for a procurement lead reading the same 60 page SOW for the third time. There's no line item for a director reformatting a spreadsheet an intern could have built. It's spread across calendars, buried inside jobs everyone's proud of, disguised as "that's just the work."

Let me give you a real one. Early in my career, every purchase over $5,000 needed a sourcing justification form. You filled it out by hand, uploaded it, routed it for approval, all to prove you'd actually negotiated and competitively bid what you were buying. Good control, on paper. In practice, it ate hours out of the week of the one person who'd already done the hard part, the actual negotiating, and now had to prove it in triplicate. That form is the whole problem, in miniature. Multiply it by every senior person in your building doing their own version of it, and you start to see the real size of this.

Why nobody runs the audit

Think about who'd actually own this. Finance can't see it, there's nothing to invoice. IT can't see it, the work happens in someone's head, not in a system. And the people doing it rarely complain, because to them, this is the job. They were hired for judgment, and somewhere along the way that judgment got buried under the cross checking, the reconciling, the chasing people down for signatures.

So the most expensive inefficiency in your operation is the one with no owner, no dashboard, and nobody in the room arguing for it. It just sits there and compounds.

When I was running a $400M+ indirect spend program at NXP, with a contingent workforce north of 10,000 people, the pattern never changed. My most experienced specialists were losing 30 to 40% of their week to work that didn't need their experience. Not because they were slow. Because nobody had stopped long enough to redesign the work around what they were actually good at.

Do the math on that. A team of senior people losing a third of their week isn't a footnote in a performance review. It's six or seven figures of lost time and money, every single year, sitting in plain sight. You just have to be willing to count it.

How to actually run it

You don't need a consulting army or a six month engagement. You need two weeks and the discipline to look. Here's the version I'd run first, starting Monday:

Pick one high judgment role, the one you already suspect is worst.
Have that person tag two weeks of their own work into two buckets: "only I can do this" and "someone, or something, else could do this." No tools, no science. Just honest tagging.
Take everything in the second bucket and write down how it actually runs, the trigger, the steps, the systems it touches, the decision behind it. This is the part everyone skips. It's also the part that matters.
Multiply the hours in that bucket by loaded cost. That number is your ceiling.
Rank what you found by value, effort, and risk, and decide what gets redesigned first, and just as important, what you deliberately leave alone because judgment belongs there.

That's it. What comes out the other end isn't a wish list. It's a number you can defend in a board meeting, and a short, ranked plan for getting it back.

The part nobody wants to hear

This audit is unglamorous. There's no demo, no dashboard, no ribbon cutting. That's exactly why nobody runs it, and exactly why the money is still sitting on the table.

You can't automate, redesign, or hand off work you've never measured. So start by counting what your best people shouldn't be doing anymore. The number will be bigger than you expect, and the fix was never cutting people. It's giving those people their judgment back.

What are the things that only YOU can do? Start there, then build the audit around everyone else's answer to that same question.

From the Field · Field Note

Shadow AI Is a Board Issue, Not an IT Issue

Published April 2026

Most leaders underestimate Shadow AI by an order of magnitude. Here's the disclosure framework boards should be asking for.

Right now, somewhere in your building, someone is pasting a customer contract into a free AI tool to summarize it real quick. Somebody else is dropping quarterly numbers into a chatbot to help build a deck for Monday. Nobody approved it. Nobody logged it. And most of it, you will never find out about, until you do.

That's Shadow AI. The AI your people are already using, with data you never authorized, creating exposure nobody in the room can see. When it comes up at all, it usually gets handed to IT to detect and block. Here's the thing. That's exactly why it keeps growing. Shadow AI was never an IT problem. It's a governance problem, and governance sits with the board.

Why this sits above IT

IT can manage a laptop. It cannot set the company's appetite for risk. That's the board's job, and always has been. Shadow AI is risk appetite getting set by default, one employee at a time, by people who never signed up to make that call.

I spent 25+ years managing spend and contracts that nobody outside the room ever saw the details of. A board that would never let an employee sign a six figure contract without a signature and a paper trail is, this week, letting employees feed the company's most sensitive information into tools whose terms of service nobody has actually read. That's not a hypothetical. That's Tuesday.

Blocking doesn't work

The instinct is always to ban it. Don't. People reach for these tools because they're under real pressure to keep up, and a ban just pushes the exact same behavior further into the dark, where you can see even less of it. The goal was never to stamp out Shadow AI. It's to bring it into the light: sanction the right tools, train people to use them well, and put real governance around the whole thing. That's a leadership move. A lockdown isn't.

The disclosure framework boards should actually ask for

You don't govern what you can't see. Here are the questions I'd put in front of management, and if they can't answer them cleanly, that silence is the finding.

What AI tools are actually in use across this company, sanctioned and not? If we can't produce that list, we just found our first gap.
What data has already left the building through these tools? Customer data, IP, financials, employee records. What's the real exposure?
Which decisions are now partly AI influenced, and who's accountable for them? That accountability can't be fuzzy.
What's our actual policy, and how do we know anyone follows it? A policy nobody reads is theater.
What's the enablement plan? People go looking for Shadow AI because we haven't given them a sanctioned, well supported alternative yet.
What happens when an AI tool causes real harm? Who responds, who gets notified, what's the actual playbook?

Six questions. They turn a vague sense of unease into a governed position you can defend to regulators, to customers, and to your own board.

The reframe that matters

Interestingly enough, Shadow AI isn't a sign your people are reckless. It's a sign they're trying to do their jobs in a world that moved faster than your policy did, and nobody's handed them a safe path yet. The fix is leadership. Give them the approved tools, the training, and the guardrails to use AI confidently, and in the open.

That's what governance should actually look like. Not surveillance. Not a ban. A clear, board defensible position that protects the company and trusts the people inside it.

Deloitte's own research this year, a survey of more than 3,000 IT and business leaders, found that close to eight in ten organizations still don't have mature governance over their AI agents. Only about one in five actually do. If your board hasn't asked these six questions yet, you already know which side of that number you're on.

From the Field · Field Note

Digital Associates vs. Agents vs. Copilots

Published May 2026

Three categories the market keeps blurring. Different governance, different ROI math, different operational fit. A practical decision tree.

Walk a trade show floor, or sit through three vendor demos in a week, and you'll hear "copilot," "agent," and "AI assistant" used like they're interchangeable. They're not. And here's what gets me: the blur isn't an accident. It's easier to sell when the buyer can't tell the categories apart. The real cost shows up later, when you've bought the wrong category, or you're governing it like a different one, and the whole investment quietly stalls.

After 25+ years managing people and the work they actually do, here's the distinction I use, and it has almost nothing to do with the model underneath. It comes down to one question: who's accountable for the output?

The three categories, plainly

A copilot assists a person, in real time. It drafts, it suggests, it summarizes, while a human stays in the driver's seat and does the work, just faster. The ROI math is simple: time saved per person. The governance is light, because a person reviews and owns every single output before it goes anywhere. You adopt a copilot.

An agent executes a task on its own. You give it a goal, and it works the steps end to end without anyone standing in the middle of each one. That's a bigger prize than a copilot. It's also a bigger risk. The ROI math shifts from time saved to work removed entirely, which is worth more, and costs more if it goes sideways. Governance gets heavier here. Something is now acting on its own, and you need guardrails, monitoring, and one clear owner for what it does. You deploy an agent.

A Digital Associate is an agent you manage like a person on your team. This is the category I build around, because it's the one that actually holds up in a regulated, customer facing, real money environment. It's an agent, scoped to a specific workflow, supervised, with clear expectations, and a named human accountable for it, the same way you'd manage anyone reporting to you. You can coach it, retrain it, or retire it. The ROI math is time and money recovered on a defined piece of work. The governance isn't bolted on afterward. It's built into the job description from day one. You manage a Digital Associate.

Here's the part that actually matters. The difference between a plain agent and a Digital Associate was never the technology. It's whether there's a person and a structure standing behind it. An unmanaged agent is a liability with good intentions. A Digital Associate is the same capability, with a manager.

A decision tree, before you buy anything

I'd ask these in order.

Is the work judgment heavy, where a person needs to weigh in on every single step? Then you want a copilot. Speed the expert up. Don't touch their judgment.

Is it repeatable and rule-able, the kind of thing you want off your best people's plates entirely? Then you're in autonomy territory, and one more question decides which flavor. Is it high stakes, customer facing, or regulated, the way hiring decisions are? Then build it as a Digital Associate, supervised, accountable, governed the way you'd govern a person making that same call. Screening resumes and scheduling interviews is administrative work an agent can handle. Deciding who gets the job never should be, and that line has to be designed in from the start, not discovered after something goes wrong. Is it low stakes, internal, and easy to reverse if it's wrong? A plain agent is probably enough.

Can you not yet describe the workflow precisely, the trigger, the steps, the actual decision rule? Then you're not ready to automate anything. Go document how the work really runs first. Automating a workflow you can't describe is how you scale a mess faster.

Why the precision is worth the trouble

The market blurs these three because blur sells. But precision has always been the advantage of a good operator: knowing exactly what you're deploying, what it's allowed to do, and who answers for it when it doesn't. Get the category right and both your governance and your ROI math get simpler, not harder.

That carrot cake moment, when a leader finally sees what this technology can actually do for their team, is worth having. Just make sure it happens before the purchase order, not after, when you're stuck governing an agent like a copilot or babysitting a copilot like it's an agent.

Copilots make your experts faster. Agents take work off their plate entirely. Digital Associates do both, with a manager attached. Choose on purpose. What are the things that only YOU can do? Everything else is a candidate for one of the other two.

Guides

AI Governance for Operations

Published June 2026

A practical framework for moving from Shadow AI to a governed Digital Associate model, without losing operational velocity. Includes the risk lanes, the Governor, and a 90-day sequence.

Read the guide

One email a month.

What is actually working in AI operations, from someone who has run these functions. No drip funnel, no pitch.