John: Portfolio companies are looking to reduce headcount dependency, increase decision velocity and capture efficiency gains through automation, agent tools and internal AI copilots, especially across product, customer operations, and revenue teams. As someone who’s developed moonshot transformation projects, how are you approaching this shift towards agent workflows?

Craig: I generally start with an example, an analogy. Let’s go back a hundred years. If we had a production workflow for a publishing business, we would have a team of people conducting research. They would give that to a team of people who would do the writing. Then it would go to a whole bunch of people who would do sub-editing, and then to graphic design and then to print. Then there would be an entire team that handles marketing, and another team that handles distribution – including people physically going out on the street and selling that publication. This is a workflow full of agents.

“What we may be heading toward, in some sectors, is a workflow where the agents are almost exclusively machine.” But where we are now is a hybrid workflow where we have machines augmenting each of those stages of activity and making them easier, which is reducing headcount. So, if you take that same publishing business today, it won’t have a team of 20 researchers. They’re not going to have a team of 40 typists writing up content. The numbers are going to be much smaller in terms of humans, plus they’re going to have a much higher targeted output. Rather than having one newspaper a day, you’re producing 240 pieces of content every hour.

When I counsel companies trying to navigate this augmented hybrid space, I start by asking them which bits can be wholly machine, which need to be wholly human, and where is there a role for both? A great example can be seen in biotech. Consider assessing tumor markers or tumor visibility on scans. When a human reviews slide after slide for abnormalities, they are about 82% effective. When a machine does it on its own, it’s about 86 to 88% effective. But when you have a human and a machine look for abnormalities together, they’re 92% effective.

John:  I like your example because it celebrates an area where the combined capability gets better outcomes than ever before. And that’s really where I see AI taking us. I think many companies that were early adopters of AI reduced staff or hired too quickly in response to the immediate impact of an agent. Too many businesses are viewing AI as only a human replacement, when they should be thinking of AI as a means to achieve outcomes that may not have been possible before.

Admittedly, I do see product leaders changing the benchmarks and bringing in the context of what’s now possible in partnership with their technology counterparts.

So how do you balance the short-term costs of implementing that automation against the long-term efficiency gains, especially when it comes to positioning this to PE, who would have a compressed timeframe?

Craig: So it’s a really good question. Every project I’ve worked on in the past year and a half has now paid for itself within 12 months. A favourite example of a recent project would be one that cost $1.5 million to build but then immediately took $1.8 million out of payroll upon launch. That’s the kind of impact we’ve been seeing with well-planned and rapidly executed experiments. 

So we’re no longer looking at the typical multi-year ROI. Instead, we’re looking at scenarios where, if we do this and reduce this headcount, you will see this in savings within a year.

John: There’s a lot of service design potential in crafting new processes and examining all the people and roles within the system. This will require a deep understanding of each department’s tasks and processes, as well as transitioning some teams to new, agent-oriented processes.

Which functions — product, customer operations, revenue — where are you seeing the most dramatic automation wins?

Craig: Operations. In reverse order, Product has the fewest automatable pieces because it generally uses a human’s brain more. And then you go through Revenue because rev-ops is still very focused on the human interaction piece – a human buys from another human. But that will change. We’re going to move to a machine buying from a machine in months, not years. 

Next, we have Customer Success – the number of repeatable automatable tasks that a customer success team does for this customer. And then we have Operations, where the opportunities to eliminate replication are obviously huge.

John: What companies do you feel are still struggling to implement automation most effectively?

Craig: I’ve spent a lot of time thinking about this, and the link is not geographic, it’s not sectoral, and it’s not based on size. It’s attitudinal. It’s companies that have the ability to say, “we are not going to have our Kodak or Blockbuster moment and we’re scared, that’s fine. We’re going to experiment, we’re going to innovate, we’re going to try things, we’re going to iteratively build.” 

If this is the first time they’re doing that, they struggle. But if they’ve tried new things and been open to experimentation, they’re in a much better position now for this than anyone else, because the team and culture are more open to change.

The companies where the desks have all been positioned the same way for 20 years, and processes have been in place forever because “that’s just how we do it.” They are absolutely screwed.

John:  Folks are more comfortable with change, but that doesn’t always mean they’re risk-tolerant. What’s nice with starting with operations for AI implementation is that there are many low-hanging fruit areas to get started in, and you can build up to higher levels of risk willingness over time.

How do you handle the cultural resistance that can come with this type of activity?

Craig: It’s by far the hardest part. And I’ve got a binary personality. I’m black and white, that’s it. I’m not immune to it, but I am slightly more numb to the gray area of staff in between, which makes me a good moonshot leader because I know that, by the end of the process, all these people will love this. But at the beginning, they will hate it. And that journey is not linear. It’s a whole chaotic mess in between.

John: Yes, there’s a lot of very appropriate fear of being replaced, especially in the U.S., where the job market is already difficult for applicants of all levels. No one likes to talk about needing fewer people, which is why I like to focus on moving the goal post further, combining experience and AI to discover those combined outcomes. 

There’s also the very practical automation of processes, so I think there will be some very strong employee headwinds for some time. But we’re already seeing the earliest adopters, like Shopify, revert their hiring strategies after making early mistakes that leaned too heavily into AI.

Craig: It needs really strong leadership. It needs really good comms, real handholding across the team, demonstrable leadership, and first-class change management. If you’re big enough to have a people department, this needs to be their obsessive focus – making sure people know what’s happening and why it’s happening. The repetition of these messages is all really important because a lot of people are going to genuinely dislike this. 

John: What is your framework for deciding what should be automated versus what requires human exclusive judgment?

Craig: I’ve been playing with this, I’ve been asked this a few times recently, and I thought, “I don’t really have a framework. It’s always bespoke.” And then I thought, “Well, that doesn’t sound like me. I’m a product person. I must have done this. I just haven’t packaged it yet.” So, I’ve just finished writing an article about this.

I think there is a framework, and it does have guardrails and parameters in it. It is fluid as well. There’s a bit of flexibility that needs to be incorporated, but ultimately, for me, it comes down to quality, replicability, and explainability. Where you can make a decision in an automated way or you can run a process through an automation workflow, as long as it’s always explainable, you can always replicate it, and it produces as high-quality an output as if a human had done it. And then you wrap it up with “it’s faster, it’s cheaper, it’s better.” They’re my three minimums. That’s my smell test of “yes, that is automatable.”

Ultimately, that takes you from an instinctive “I think that’s right,” to “It matches this framework,” You then have to build an evaluation, experiment and build something to test it. The results of that test are when you decide whether you’ll do this now. Or not.