We Called It an AI Project. It Was a Process Overhaul.

The AI part took a few weeks. We’re still fixing everything around it.

I asked for that project, argued for it, got it approved, and then spent months explaining the timeline. Almost none of it went the way I pitched it, and what went wrong wasn’t the AI.

TL;DR

• What I asked for wasn’t software. It was a person who knew how to get quality work out of AI.
• The AI piece came together in a few weeks. The steps before and after it are still being rebuilt.
• Our team had been quietly working around a broken process for years. AI can’t work around anything, so every gap showed up at once.
• Building something that didn’t exist was more work. Changing something that already existed was harder. Those aren’t the same problem.
• I went in thinking about the tool. I should have gone in thinking about the process the tool was landing in.
• Looking back, this wasn’t an AI project. It was a process overhaul with an AI tool in the middle of it.

The ask was a person, not a tool

Our content operation had outgrown the way we ran it. Each marketing manager owned strategy for their own area, which meant nobody owned it across the site. We had a long list of sections that needed an overhaul, organic growth to keep up with, and day-to-day demand that ticked up every quarter. The math stopped working. Too much content for the staff and the traditional write-it-by-hand model to absorb.

So I asked for a resource: a content strategist who could do the strategy work and knew how to get quality output from AI tools. Not a license, not a platform. A person.

I led with capacity, because capacity was the problem the organization already agreed it had. Nobody needed convincing that we were falling behind. And the risk of standing still was easy to describe: the backlog keeps growing, and eventually only the most urgent work gets done. Nobody in the room found that acceptable.

Marketing leadership approved it. What they expected in return was straightforward, and fair: write copy faster, publish pages faster, and not at the expense of quality.

That’s not what happened first.

The engine took a few weeks

The strategist we hired built the AI side of this in a few weeks. Not from nothing. Most of the raw material already existed, scattered across the organization: a writing style guide, brand-approved names for things, various forms for collecting information from stakeholders and clinical experts. The job was gathering it, organizing it, and getting it into a structured shape the AI could actually use. Then feeding it good examples, running test drafts past our communications team, and going back and forth with them, refining the output and writing rules until the drafts were good enough to use.

A few weeks. That’s the whole AI story.

The AI was the easy part. The process it landed in turned out to be the project.

Everything before that step, and everything after it, is what’s taken months. That’s still going.

Our team had been quietly covering for us

The strategist started with the inputs, and that’s where the project turned into something else.

To draft anything at the depth and quality we require, you need the facts. Real ones: what the service line actually does, what makes it different, what the physicians would say if you asked them. So she went looking for those facts, and found that our process for gathering them was mostly a person sending out interview questions and hoping. When answers were slow, someone chased. Sometimes a different someone. No templates, no structure, no clear owner. Answers came back partial. Questions went unanswered. Facts were missing. In some cases we were asking the same person the same question more than once, which is its own kind of damage: every duplicate ask spends a little of a busy physician’s patience, and the next request gets answered slower, or not at all.

This had been true for years.

Our team had been absorbing it, and they were good at it. The copy that shipped was good. It got there through revision: draft, review, someone spots what’s missing or wrong, back it goes, and around again. Each pass caught something real. By the end the page was accurate and it was on brand.

So nobody downstream ever saw a quality problem, because by the time they saw anything there wasn’t one. What they also never saw was the count. A page that should have taken two or three passes was taking six or seven, and no line item anywhere in our process said so. A process that quietly costs you four extra rounds doesn’t look broken. It looks busy.

That’s where I expect the return to show up, and it isn’t the one I pitched. Not a faster first draft. A first draft that’s close enough to right that it doesn’t need seven trips around the building.

Getting there means the inputs have to be complete, though, and that’s the part AI won’t do for you. Feed a machine partial information and it hands back a partial draft, confidently written, at whatever length you asked for. The difference is that the gap stops getting absorbed by the next reviewer and starts showing up in front of you on the first pass, where you can fix what caused it.

So we hadn’t dropped AI into a working process and watched it speed up. We’d dropped AI into a process that had never really worked, and it took a measurement nobody had asked for.

None of this is a new idea, which is the humbling part. Michael Hammer made this argument in Harvard Business Review in 1990, and the title was the whole thesis: don’t automate, obliterate. Put technology on top of a broken process and all you’ve built is a faster broken process. I knew that. I’ve quoted things like it in meetings. I just didn’t think it applied here, because this time the technology was the interesting part.

Building new is work. Changing what exists is harder.

Two big pieces of work came out of this, and comparing them taught me more than anything else in the project.

The first was building the collection of rules, examples, style, and facts that the AI works from. That was a lot of work. It also went fast, because none of it existed before. There was nothing to argue with. No established way of doing it, no team with a stake in the old version, nobody who had to change how they spend Tuesday.

The second was fixing how we gather facts in the first place. Less work by volume. Much harder. Because that process did exist, people were used to it, and it ran through marketing managers, clinical experts, and our operations team. Changing it means changing what a lot of people do and how they think about their part in it.

We have the fix designed: better questionnaires, and a tool to handle the outreach, the follow-ups, and the routing of answers so a human isn’t chasing people by hand. We haven’t finished putting it in place.

We also added a rule that I think matters more than the tooling: drafting doesn’t start until the facts are in. The strategist pushed hardest for that one, and she was right. At the quality and depth we’ve committed to, AI can’t write around missing information any more than it can invent a physician’s opinion. What made the rule stickable was fixing the collection process first. Holding the line only works when you’re confident the facts are coming.

One side effect I didn’t plan for has turned into the part I’d defend hardest. Our facts largely didn’t exist in writing. They lived in the heads of marketing managers and experts inside the service lines. Getting them written down and organized is work we’d never have prioritized on its own, and we’d have been better off with it years ago. That asset is ours whether or not AI ever drafts another word.

Before you go further, do the arithmetic on your own. Take whatever you’re planning and split the effort.

The bill nobody puts in the business case

Now the money.

We thought we might free up budget by pulling back from a third-party agency writing copy for us. What happened is that somebody still needed to edit, so the agency stayed on in that role. We moved spend from writing to editing. No net savings yet. We’re still working on it.

Then the project generated a second staffing need. Fixing the fact-gathering process is process work, not strategy work. Could the strategist do it? Yes. Is that the best use of someone who should be doing content strategy across the whole site? No. So we’re bringing in a second resource on a temporary basis first, partly to make sure we’ve got the right person and partly to prove the role is needed before I go ask for it permanently.

I’m not embarrassed by any of that. There are always curveballs you didn’t plan for. The question isn’t whether they show up, it’s what you do when they do. Should we stall the whole effort rather than add help that keeps it moving? Of course not. It’s a contained risk with a clear read at the end of it. I expect a few more surprises before this is finished, and each one gets weighed the same way.

Meanwhile, leadership still wants the curve to go up, and they’re frustrated it hasn’t gone up yet. That’s reasonable. They approved something on a promise. I’ve stayed transparent about where we are and what’s in the way, and I think they still trust where it’s headed. When people ask me how it’s going, I tell them it’s going about as I expected. This is like most technology implementations I’ve been part of. It looks simple from the outside. You start with “all we need to do is just,” and then you get into the weeds and find the nuance. Are we stuck? No. Was it easy? No. If it were easy, we’d have finished it months ago.

What I’d do differently

I’d assess the process before I bought anything or hired my AI strategist.

I was focused on the tool: the technology, and what it could produce. Some of that is just the seat I sit in. I’m the technology leader, not the operations leader, and evaluating the capability is the part of this that’s genuinely my job. But new technology forces process change. That isn’t a subtle lesson, it’s most of my career, and I’ve said it out loud to other people more than once. I just didn’t carry it into this one, because the AI was the interesting part.

I didn’t need a full change plan written in advance. But I could have spent a week walking our existing process end to end and marking the spots that were going to get hit hardest. That alone would have changed how I set expectations, which is the piece I’d most want back.

So my advice, if you’re about to do this: don’t oversell the outcome. Manage expectations up, and manage your own. There’s gold in these hills and we’re going to find it. We won’t strike it every time. Make the case, show the upside, name the risks, and say plainly how much effort the change is going to take. That last part is the one people skip, and it’s the one that costs you credibility six weeks in.

The productivity gain is hard enough to prove when it does arrive, so don’t promise a version of it you haven’t tested. And if the quality of what you get back depends on the quality of what you put in, then the real project was always the input.

So was it an AI project?

We call something an AI project because there’s AI in it. But look at what this actually was. A fragmented content strategy got centralized. A fact-gathering process that had been broken for years got rebuilt. Institutional knowledge that lived in people’s heads got written down. And one step in the middle of the process is now drafted by a machine instead of a person.

If I build a house and install a smart thermostat, is that an AI project, or did I build a house?

I’d still make the same case tomorrow. Demand keeps growing and budgets keep shrinking, and every healthcare marketing team I know is in that same squeeze. Efficiency is the only play left, and AI is how we get it at scale. That part I’m certain about.

I’m just done calling these things AI projects. Ours was a process overhaul that happened to put AI in the middle. The AI showed up on time. The rest of it is what we’re still building, and most of it needed building anyway.

If you’ve run one of these, I’d like to know what yours turned out to be underneath. I doubt many of them were AI projects either.