If any part of your job — any part at all — involves knowledge work, you can’t throw a rock without hitting a robot right now.
Generative technology was literally built to do this. Whether it’s doing it well is a question for another post. But the upshot is teams everywhere have been scrambling to invest in AI-ification to improve efficiency.
My team is right there with them.
A year and a half ago, I asked one of my direct reports to talk with content-developer teammates and find out what they were learning from their AI experiments, because, well, leaders are very interested in that sort of thing.
(Side quest: If you’re not familiar with the typical content workflow, it goes something like Plan, Draft, Review, Edit, Publish. The Review step can be deceptively simple, because the minute you write something, ten mandatory reviewers come out of the woodwork with conflicting opinions on what to revise. This can go on for weeks, if not months.)
Back to my intrepid direct. She made the rounds and gathered up the learnings, and they were all generally the same. The most time saved was with the Draft step, going from about a day of work to a half-day. Nothing to sneeze at and definitely worth pursuing. But Review, the longest step, was just as onerous as ever, which diluted the overall savings.
Mind you, the technology has significantly changed since then, and the results are different today. But it illustrates the struggle many teams are encountering as they automate.
No sprinkles, please
Fast forward to summer 2025, and MIT’s official and oft-quoted research clocked it, too: Many teams are not, in fact, seeing the Promised Land of AI integration, and up to 95% of enterprise projects are seeing zero return on the investment. Zero.
But what about all those visions of automated splendor? Weren’t we promised a workplace revolution? Machines and humans working in kumbaya harmony to deliver faster and better?
MIT says: “Most fail due to brittle workflows, lack of contextual learning, and misalignment with day-to-day operations.”
Translated into real-person-speak: The tools don’t match up with the way teams work, which was broken to begin with.
A colleague of mine refers to this phenomenon as sprinkling AI on top. On the surface, it looks all nice and shiny. Underneath, it’s the same old mess. And now it’s a really expensive mess.
The trouble comes when you try to copy the old system. Here’s why:
- Humans are biologically incapable of moving as fast as AI. Any process that was built solely for humans — pretty much all of them at the moment — will put a limit on AI’s efficiency. It still carries you way past what we could imagine even a year ago, but AI will always beat humans at the speed game.
- AI is intellectually incapable of thinking like humans do. It might seem like it does, but it’s only mimicking what it’s been taught. It makes mistakes. A lot. (Did you find the missing feet in my image above?) And it needs oversight. Perhaps less so as the technology improves, but at least for now, humans still need to play a big role.
The functions are complementary: a Doer Brain (AI) for speed and a Thinker Brain (humans) for quality. But if you’re trying to push both into a system designed for only one brain, it won’t work the way you intended when you made the pitch for cool AI tools. It’ll be even worse if the one-brain system is plagued by technical or operational debt. AI’s trademark efficiency will simply magnify that instead of fixing it.
Enter the matrix
A system is the interconnected set of tools, processes, and relationships that result in Doing a Thing. And the experts are in agreement that you won’t get the ROI you’re looking for until you rebuild all of it. From the ground up. No shortcuts or only redoing bits and pieces.
Gosh, though. That seems like a lot of work.
Okay. Deep breath. You don’t have to blow up your factory yet. Pick one system to start with, something straightforward.
1. Objective
First and foremost, you need a clear understanding of your system’s raison d’etre. You’re bound to have resistors, and if you can’t articulate a reason for why you’re doing this and what’s in it for them, you won’t win them over. Even if they pretend you did.
Be brutally honest. What’s your system meant to do? Is it actually doing that? What does a good result look like? Why is it a good candidate for AI-ifying? How would success move you closer to your other goals? (or *cough* the CEO’s goals)
The more specific you can be, the better, because these answers will ground the rest of your decisions. Props if you can set clear success metrics, too. These’ll tie straight back to ROI.
2. Process map
Grab your markers because it’s white board time! You and your team are gonna map out the one-brain system end to end. Each input. Each task. Each decision. Each tool. Each person or group involved. How much time between tasks and across the whole shebang. The unofficial workarounds people are using because a pain point isn’t or couldn’t be fixed.
Now take a second look at your map. Do you see obvious fluff or redundancy that you could get rid of right now without AI? Are you missing anything? Is anything already automated? Do recurring problems bog things down? Make notes on all that.
Cool-factor bonus: Record your session and turn on the AI transcriber. Dump the notes in the LLM tool of your choice, prompt it to be a systems expert, and ask for its opinion on what to improve.
3. Technology
Here’s where you need an expert who understands what AI is already capable of, and what’s coming next. They don’t necessarily have to have “engineer” in their title. But probably not Bob from Accounting who once created a macro in Excel. (Stellar use of VLOOKUP, though, Bob!)
Your expert will scour your map and figure out what’s automatable and what the best tool is for the job. For example, do you need an AI agent to do a task or would a rules-based automation work?
Cool-factor bonus: Your expert is almost certainly using AI to help them plan this out.
The idea is to end up with a blueprint for what tools you need to build or buy, and for creating the knowledge bases, agents, workflows, and instructions that make the system run.
4. People
While the tech/AI experts are busy creating the Doer brain, you’ll switch over to the Thinker brain. What are the human roles in your system? Do you have gatekeepers? Do you need planners or strategists? What about humans-in-the-loop or evaluators who are assessing quality?
Think in terms of what each role is expected to do for the new system, and not the current one. List out the responsibilities and the skills needed to accomplish them. It pays to be detailed here, too, because you’re likely to spot talent gaps.
With your lists in hand, you’re ready to set up programs that upskill or reskill the team. You’ll also want to brush up on best practices for managing big changes, such as making sure your leaders are all socializing the same objective and benefits of the new system.
While you’re at it, you should have a look at your organizational structure. It was built for a different era and might not be the best setup for supporting your mission anymore.
5. Governance
None of this is one and done. I glossed over it for the sake of brevity, but you’ll need to do a lot of testing — A LOT — before pushing your system live. Don’t skimp on it. Pressure-test the agents, cross-check for conflicts, look for the failure points where hallucinations are creeping in, and so on. Then, once the system is live, get in the habit of redoing all this on a regular basis.
You’ll also need to put operational backups in place in case, say, Claude gets a wild hair and vibe-codes your production database into oblivion or tells an unwitting customer they’re an a-hole and can eff right off. And the technology is changing so fast, you’ll have to maintain those reskilling and upskilling programs, too.
Don’t go chasing waterfalls
Speaking of the pace of change, you don’t have the luxury of doing your system overhaul Waterfall-style where you plan every step in great detail, build it over the course of a year, cut over to the new system in a single big hurrah, and give everyone pats on the back. By the time all that’s done, your system will be outdated. Again. And how much side-eye will the CEO give you when you ask for an investment do-over?
Unless it’s a relatively small system, you’ll have to do this in pieces, and you should expect your plan to be more of a living document than a roadmap. The good news is, it’s easier to test, iterate, and learn as you go. The bad news is, it will seem chaotic to your team. This is one of the reasons you need that clear end-state objective to rally around.
If you’re the lucky soul tasked with this Mission Impossible, I see you and feel your pain. And if your familiarity with AI is mostly chat tools, I’d recommend an intense cramming of free AI-for-leaders online courses (or the paid ones, if your company will reimburse). This’ll give you context on how the technology works and how others are putting it to use. I also drafted up this nifty set of free worksheets to help you collect your thoughts.
Cool-factor bonus: You’ve got the perfect opportunity yourself for some career future-proofing.
All opinions here are my own. All text is my own, too, including the em dashes. I welcome constructive comments and discussion on LinkedIn and Bluesky.