Ban the word efficiency from your next AI meeting

MJ has spent four years running AI inside a fintech. Her most useful advice starts with taking cost savings off the table. She has an exercise she gives leadership teams. Write down your AI initiatives. The catch is that you're not allowed to write down anything about efficiency, cost, or productivity per person.

"If you eliminate those possibilities, what are the next things that are going to come up?"

Not much, for a while. She's upfront that the second wave takes longer to arrive, and equally clear that it's the wave worth having: "those ideas will be bigger, more creative, more differentiated, and will add more value to your business."

MJ is Chief Strategy and Revenue Officer at Credibly, where she runs growth and the company's AI program. Before that she spent eight years at Ford, where she founded a mobility startup and took it external. She has a literature degree and an MBA, and she'll tell you she's a writer first and a technologist second. She came on DevLab to talk about AI adoption as a change problem, and most of what she said had very little to do with models.

Why every AI list starts with cost

Her explanation for the empty page has nothing to do with imagination. It goes back a lot further than that.

For as long as anyone has been at work, thinking has been the scarce input. "Throughout human civilization, intelligence has always been at a premium, and it's always been isolated to someone's physical brain," she says. So we built manners around rationing it. You get thirty minutes with your manager, so you bring the three things that matter and handle the rest yourself. You don't ask a senior colleague for forty options. You ask them to react to the two you already narrowed down.

Then a technology showed up with none of those limits, and we kept the manners. "It doesn't get tired. It has infinite memory. You can ask it to generate a hundred business ideas or a hundred versions of the same PowerPoint and it'll do it in minutes."

That's what she means by an abundance model, and it sits awkwardly with the way most organizations plan, because planning is the practice of allocating scarce things. Hand a planning committee a resource with no scarcity in it and you'll still get back a list of savings.

Her version of this for writing landed a little close to home for us. Drafting by hand and having AI check the grammar afterwards is "completely unnecessary". Handing it a prompt and publishing what comes back is worse: "I think that's very lazy and probably not good for writing, and we have all read too much AI crap." What she does instead sounds more like an argument. Here's what I want to say, here are my audiences, go find data that supports this, push back against me, challenge me.

"When it feels like you're playing, that's when it feels right."

The docks got automated. That wasn't the story.

Her best argument for looking past headcount came out of an executive class at Wharton, and it's about shipping containers.

When container loading was automated, the immediate math at any single port was obvious and brutal. You don't need men to lift crates any more, so what happens to the fifty guys on the dock? The ports that were built for automation became the international ports, "which changed the GDP of those countries and the per capita income of those cities, which then changed the entire global economy of what ports are now relevant and what ports are no longer relevant." Trade moved. The geopolitics moved with it.

"So yes, we as CEOs can focus on the 50 people, or you can think about what does this really mean?"

She makes no claim to know AI's version of the answer. "If I could answer that question, I would be a billionaire, but I can't." She's only clear about which question is larger, and about the stake in it: "the most valuable companies are not even built yet right now."

Generative AI is not a patch for the tools you never bought

Two requests come up over and over when she asks teams what they want, and she counts both as mistakes.

The first sounds entirely reasonable. I want this report generated every Monday, in this format, sent out automatically. "That's not generative AI. That's just automation."

Teams reach for a language model to do deterministic work because the deterministic tool never made it up the priority list, so "people use Gen AI as a band-aid for the tools they actually want." It half-works, which is the worst available outcome, because now a business process depends on something that will occasionally hand back a different answer. If you need the same output every time, this is the wrong technology and no amount of prompting fixes that.

The second mistake costs more.

Copying your best employee caps you at your best employee

The first instinct on most automation projects is to clone the best person in the room. In her world that means underwriting. "Let's take our best underwriter, let's take all of their processes, and let's make this tool do exactly that so that we can get those results. That is going to fail."

It took about a year for that to become obvious, which she puts down to the fact that nobody rethinks a workflow they still believe in until they hit a wall.

The reason replication is a ceiling: a human underwriter can hold a handful of cash flow rules in their head, so the rules get written for human memory. Here are some industries, here's what declining cash flow looks like, apply this threshold. A model has no such constraint. It can take a single file and work out what good cash flow looks like for that region, that retail category, that margin profile, then ask whether this business sits above or below it. "There's no way humans can get that granular."

Build the machine to imitate your best person and you've bought a marginally faster version of the ceiling you already had.

What ROI looks like before it looks like money

This one is for anyone whose AI program is being assessed by someone holding a spreadsheet.

Cost savings show up late, because there's a learning curve, and measuring only cost tells you very little about whether the program is working. The metric she uses is rate of change. Not the tidiest line for a board pack, but her questions are concrete: "Are you able to accomplish things faster? Are you able to accomplish new things? Are you learning from your mistakes faster? Do you have more people who want to try this thing?"

Underneath it is a view of the company as an organism with different clock speeds, where some parts move fast, some resist, and the useful question is how adaptable the whole thing is. Spend controls still matter, and she's clear that's what a finance function is there for.

She's equally clear about the cost of measuring the wrong thing, because AI doesn't only add capability. It adds complexity, and complexity is unkind to people who want simple answers: "it will make you make bad decisions faster and be really confident about it."

The bottleneck moves, it doesn't disappear

Here's the pattern she keeps seeing, and it's the one worth reading twice if you run an engineering org.

Junior developers write more code, faster, with AI. Senior developers still have to check that code, because they're the ones on the hook when it ships. So they spend their week reading a large volume of code written by a system trained on the internet, working out whether it's correct, whether it's efficient, whether it does what it appears to do. "So you may have created what you thought were efficiencies on one end of your department, and then you created burnout on the other end."

The same displacement shows up in risk, which used to belong to security and legal. Now everyone with access to a model is making calls about what data goes into it. The failure mode is undramatic: a colleague builds something useful in an unsanctioned tool because telling you would be more trouble than it's worth. Shadow IT grows best in the conditions a strict policy creates when nobody explains the reasoning behind it.

If you take AI out of your strategy, is anything left?

She keeps hearing companies describe being AI-first as the strategy itself. "What is your strategy? Our strategy is to be AI first. To what end exactly?"

Her test is that the business objective should survive having AI removed from the sentence. A better customer experience, more capacity, faster growth, whatever it is for you, that's the north star, and AI is one way of getting there sooner. Put the tool in the objective slot and you lose the ability to tell whether your decisions are any good.

She'd add one thing for anyone senior enough to set that objective: use the technology yourself, daily, including the parts where it's frustrating and confidently wrong. "If you're asking your people to use it every day, you better be using it every day." Otherwise you walk into a room and describe a vision your team already knows is undeliverable, which is what middle managers keep telling her: "my CEO or my VP came to me and was so excited about what we can do with AI, and I have no way of telling them I don't like this vision. By the way, I still have my job that I need to do and now I have a new job that you just gave me."

Leaders who arrive holding a panacea get agreement instead of information, and she's unsparing about where that ends: "at worst your people stop telling you the truth, which is the death knell to any leadership."

The posts nobody is writing

She'd like to read a company's actual working on upskilling, department by department, roles that will exist and roles that won't and the skills that carry people from one to the other. She'd like to read about institutional knowledge, which almost nobody posts about: once you automate a decision, how do you keep your senior people's expertise sharp enough to keep feeding the system? Her framing is a loop. The machine takes the volume, people take the judgment calls that need judgment, and those calls come back and shape what the machine does next.

She'd also like fewer of the other posts. "I'd love to see posts about that on LinkedIn versus, I built 10 agents and they send out 500,000 emails."

That second genre is not exactly underserved.

Worth saying she hasn't solved the loop either, and she's further along than most. Four years in, from around the time ChatGPT arrived in late 2022, and it's only in recent months that she's been able to take a model of human-AI work to her leadership team and say this is what's working. She calls the hybrid model "such a fantasy" and hard enough that most companies avoid the question, because a story about efficiency is far easier to tell.

Her own diagnosis of why so few leaders sit in that discomfort is the line to end on. Rate of change, organizational adaptability, what your industry looks like on the far side of automation: "These are much more ambiguous and gooey types of strategy questions. And a lot of leaders just don't feel like grappling with that, which is not an AI problem. That's just a deep thinking problem."

Watch the full conversation

Plenty of the episode didn't fit here. How she actually built the AI program inside Credibly, from a training on what a transformer model is through to more than a hundred documented use cases in three months. What a company should be war-gaming for the day someone deepfakes its CEO. Why she expects the strongest hires of the next decade to pair a technical skill with a humanities one. And what a failed startup taught her about which stage of a business she is built for.

Watch the full conversation with MJ on DevLab.