top of page
Search

Leading AI Teams When the Ground Keeps Moving

joydeepml2020
17 hours ago
6 min read

Think back to late 2022, just before ChatGPT arrived. If you were an engineer, debugging meant a search bar, a Stack Overflow thread from 2014 and a lot of patience. Designers dug through reference boards. Marketers lived in keyword tools.

Four years later, most of us start those same tasks with an AI assistant. I debug with Claude. My teams prototype on models that didn't exist when we wrote last year's roadmap. That isn't a small productivity gain. It changes how we build, how we ship and how we sell.

And it changes what leadership means.

For decades, technology organisations ran on a comfortable rhythm: build a capability, ship a major release, follow it with minor ones, repeat. Plans were measured in years. In AI, one model release can make a feature you spent two quarters building redundant overnight, or quietly remove the moat under an entire business model. Disruptive innovation isn't a chapter in a strategy textbook anymore. It's Tuesday morning's news.


Schoemaker, Heaton and Teece made the point in California Management Review years before generative AI went mainstream: in turbulent conditions, good leadership looks less like detailed planning and more like testing hypotheses about new technologies and changing markets. In 2026, for anyone leading an AI team, those hypotheses are about tech stacks, platforms, models and product ideas. And we have to test them faster than ever.


So how do you lead and scale a team in that world? I find Bennett and Lemoine's VUCA matrix a useful lens. It sorts challenges along two questions: how much do you know about the situation, and how well can you predict the results of your actions? Each quadrant asks something different of a leader. Here is how I map it to AI teams, as five things I keep coming back to.


1. Know what's really happening at three levels (complexity)

AI programmes are complex in the textbook sense. They have many interconnected parts and plenty of information, more than anyone can process casually. Before I commit to anything, I want an honest view of three layers:

  • Infrastructure. What compute, platforms and tooling do we actually have, and what can we realistically get given security, procurement and regulatory constraints?

  • Architecture. How do our systems talk to each other? Where would a model sit, and what breaks if it misbehaves?

  • Data. This is where most AI ambitions quietly die. The question isn't "how much data do we have?" It's "what is usable, consented, governed and fit for this purpose, and what simply can't be used?"

In regulated industries like banking, the honest answer to the data question often reshapes the whole roadmap. The matrix's advice for complexity fits well: restructure where needed, bring in or grow specialists, and resource the problem properly.


2. Make the build, tune or buy call deliberately (uncertainty)

Every AI leader faces the same fork. Do we train our own foundation model, fine-tune an existing one, or call an API?

This is classic uncertainty. We understand the basic cause and effect, but not what comes next: the next open-weight release, a provider's pricing change, a competitor's launch. The prescription is to invest in information.

In practice, I ask where our differentiation really lives. If it's in proprietary data and domain behaviour that no general model captures, a domain or foundation model may be worth the investment. If it's in workflow, product experience or speed to market, API calls plus strong engineering usually win, and you stay free to switch models when a better one lands. Fine-tuning sits in between. It often makes most sense once you've proven value with an API and know exactly which behaviour you need to own.

Whatever you choose, design for replaceability. The model you pick today is a hypothesis, not a marriage. Kodak is the cautionary tale here. In 1999 it forecast that digital cameras would be about 5% of the market ten years later. By 2009 they were 95%. The future arrived faster than leadership's mental model allowed. Assume it will for you too.


3. Plan capacity for experiments, not just delivery (ambiguity)

When you move into genuinely new territory, like a new product class or an agentic workflow nobody in your industry has shipped, you face ambiguity. There are no precedents, only unknown unknowns. The only reliable way forward is to experiment: form hypotheses, test them, and make sure the lessons travel beyond the team that ran them.

That has direct consequences for workforce planning. If 100% of your team's capacity is committed to roadmap delivery, you have zero capacity to learn. I would rather deliberately reserve part of our capacity for small, time-boxed experiments, each with clear kill criteria and a habit of writing up what we learned, including what failed.

It also changes who you hire. When stacks shift every few months, I look for people who learn fast and are comfortable saying "I don't know yet". I value that over years of depth in a single framework. The framework will change. Curiosity compounds.


4. Treat token cost, human cost and ROI as one equation (volatility)

Here is a conversation more AI leaders need to have openly: what does this capability actually cost, and what does it return?

AI costs are volatile. Token prices fall, then usage explodes. An agent that is cheap in a demo can get expensive at production scale once it loops, retries and calls tools. Meanwhile the human cost of building, evaluating and supervising these systems is real and often underestimated.

So I treat it as one equation: token cost plus human cost, weighed against expected ROI. Some workflows are best served by a smaller, cheaper model with a human reviewer. Some by a more capable model with no reviewer. Some don't need AI at all. The matrix's advice for volatility applies directly: build in slack and match the investment to the risk. In practice that means budget headroom for usage spikes, fallbacks when a provider has a bad day, and unit economics reviewed every quarter, not every year.


5. Treat people as people, not resources

I've saved the most important point for last.

Right now it's tempting to see AI mainly as a way to cut headcount. I think that misreads where the value is. Models are being commoditised quickly. What isn't commoditised is the judgement of people who understand your customers, your data, your controls and your culture: people who can sense change and act on it. That is the core of what Teece calls dynamic capabilities, the ability to sense, seize and transform. You can't buy that from a vendor. You have to grow it.

C.K. Prahalad's short HBR column "The Responsible Manager" has stayed with me for years. He asks managers to invest generously in developing their colleagues, and to care about fair process, because people want fairness rather than favours and want to be heard. He describes "humility in success and courage in failure" as hallmarks of a good leader. None of that matters less because we now have agents. If anything, it matters more.

There is a practical side to this too. When leaders clearly understand estimated revenue and capacity (what the team can deliver, what it will earn, and where AI genuinely changes the equation), they can plan workforce changes gradually and humanely. Many layoffs are the bill for poor capacity planning that was ignored for too long. Good forecasting won't prevent every hard decision. It can save a company from the avoidable ones, and save people from being treated as a line item that got rounded off.

The leader's real job

None of this works with a once-a-year strategy deck. The editors of CMR's special issue on VUCA argue that strategy is turning into an ongoing conversation rather than an annual plan, and that innovation has to become a way of life rather than a one-off project. For AI teams, I'd put it simply:

  • Sense constantly: infrastructure, architecture, data and the market.

  • Decide with incomplete information, and design every decision to be reversible.

  • Fund learning, not just delivery.

  • Count tokens and people in the same spreadsheet.

  • Grow your people, because they are the one capability your competitors can't download.

The VUCA world isn't going to calm down. The real question is whether we build teams that are frightened by it or fluent in it.


 
 
 

Recent Posts

See All

Comments


bottom of page