The Data Huddle by Yuki is a podcast about modern data leaders at global companies. The people who make data work every day, but rarely get the spotlight. Honest conversations about their careers, wins, mistakes, and how they see the future of AI and data.
What happens to a data platform when the whole organisation decides, all at once, that it is late to AI? In this episode, Dana Bilder-Morozova sits down with Małgorzata McCarthy, Technology Head of Analytics, AI and Data Platform at ManpowerGroup, for a conversation about the year of uncontrolled spending, why mature companies still overspend on compute, what separates a working agent from a demo, and the one role she does not think AI can automate.
About the Guest
Małgorzata McCarthy runs analytics, AI and the data platform at a staffing organisation operating across multiple countries and continents – which means her daily problem is not technology, it is agreement. She has spent her career inside enterprises large enough that the same word means different things in different markets, and she is unusually direct about what that does to an AI roadmap.
She is also the rare senior technologist who will tell you plainly when she does not know something. Asked to predict the next wave, she declined, called her own answer “wishful thinking”, and then gave the most useful answer in the episode anyway. This was, incidentally, her first podcast.
Episode Highlights
The Consistency Problem Nobody Has Actually Solved
Ask what the data looks like at that scale and she reframes the question immediately.
“There is the question of what data you have versus what data you would like to have.”
At a company spanning countries and continents, the assumption is that one business produces one consistent set of definitions. It does not.
“You would think, in a staffing organisation, you would know who the candidate is, and the candidate definition would be the same in every country. If you work in real estate, you would think you know what a property is, and the property is understood the same way in every country, and then the process around it would be the same. And that’s typically the biggest challenge that I see.”
What that produces is a split she describes as permanent rather than temporary.
“On one hand we have a powerful set of data, a major set that is really unused, an unrealised asset. But then on the other hand you always have this ambiguity. That’s the problem that is forever there. It has not been, in any organisation that I’ve been to, perfectly well defined.”
And the platform itself? She is refreshingly unsentimental: “The platform is just a tool, a means.” Worth sitting with, because almost every AI programme that stalls in year two was sold in year one as a platform decision rather than a definition problem.
“I Would Want AI to Be an Enabler, Rather Than the Other Way Around”
Asked what AI is doing to her own function, she starts with a principle: you cannot credibly ship something you have not used yourself. She uses AI daily in her own life, which is exactly why she knows where it stops helping.
“I know where the flows are, where it helps me, where it actually makes my life more difficult, or actually brings down my role to a tester or a validator.”
She sees the same pattern inside engineering teams, and she names it precisely.
“We see now, especially on LinkedIn or actually Reddit, everywhere, data engineers saying that they’re starting to get burned out in their work because they become validators and testers of what code was spat out by somebody else.”
Her prescription inverts the current arrangement.
“Maybe those journeys need to be changed so that the testing actually stays within the AI realm rather than the other way around. I would want AI to be an enabler for us and making sure that we get rid of the tedious tasks, rather than the other way around.”
This is a sharper critique than the usual productivity debate. The claim is not that AI fails to speed teams up. It is that the speed-up was distributed badly: the machine took the writing, the human kept the checking, and checking was always the part nobody wanted.
Hype, Disillusionment, and the Return of Boring Governance
On what comes after agents, she refuses the prediction and describes the cycle instead.
“What’s really coming, it’s hype after hype. It’s always the excitement first before we see some sort of disillusionment, and then we kind of stabilise.”
Her hope is that the industry has already passed the disillusionment and is ready to ask why the last wave underdelivered.
“There’s gonna be more focus on solid data foundations, a lot more focus on data governance, on master data management, on process alignment, maybe a transformation of the processes to fit that AI agentic journey. Before we just create another shiny prototype and bank a lot of money on something that is due to fail.”
Then, unprompted: “That’s not really a prediction, that’s just wishful thinking.” Which is honest, and also a fair description of how governance usually arrives – late, and after the invoice.
What Works With Agents, and What Stays a Demo
Her test for whether an agent project survives contact with production is a single diagnostic, applied to the first sentence of the project brief.
“If the problem statement starts with ‘how are we using AI agents, how are we implementing agentic workflow’ – typically that doesn’t work. I can’t name a single case where such an approach would work beyond a demo or a PR level, so just marketing level.”
Invert the opening sentence and the odds change.
“If you start with a realistic problem that you want to solve, and that problem is tied to the business strategy, and that is very well planned, analysed and measured – this has a chance to work.”
Even then, she lists what still sinks it: foundations that are not there, inconsistent data, undefined processes, and people who are not ready. Her clearest line on the last one: “You cannot really put an agentic workflow in an area where there is no single process, so you don’t really know what you’re addressing.”
“We Did Have This Year of Uncontrolled Spending”
On whether enterprise thinking about compute cost has changed in 2026, she does not hedge.
“It definitely has. One thing is, we definitely seen a massive jump in consumption.”
And the consequence, stated without euphemism.
“We did have this year of uncontrolled spending. We’ve seen articles of companies spending massive amounts of money on their compute, on their credits, on their consumption in general.”
What interests her is where the correction lands. Not on whether to adopt AI, but on how the business case gets written.
“It’s not enough just assessing the business value that you’re gonna bring when you implement AI. It’s now planning of also the cost factor of this. Return on investment should have the investment factor as well, and how much it’s gonna cost you to do it.”
And a line that deserves to be taped to a wall in most engineering orgs:
“You don’t always have to go for the best if the cheaper option is fit for purpose.”
Why Mature Companies Still Overspend
We put the question to her directly: if these organisations have real strategy and real engineering discipline, why does compute spend still run away from them? Her answer is not technical at all.
“I would blame the hype. The hype gives you the sensation of ‘we are late’.”
The mechanism she describes is social rather than financial.
“When you have leadership being flooded with all those success stories, companies tend to be more aggressive with their investments and lesser operational, boring governance. So it starts with the spend.”
Her analogy is uncomfortably good.
“Social media always shows you the best scenario of anybody’s life – just successes, just travel, just great food, just holidays. And that’s kind of what we see also in the corporate world, where corporations are showing just their successes, how much they’ve won with the use of AI, how much they managed to save. So it’s the same hype being created in an enterprise corporate world, and companies join thinking ‘we are too late, we have to invest now’.”
She is careful not to overstate it. The awareness of credit control is there now, and so are the controls themselves.
“Tools give us more capabilities, more ways of managing the spend. More visibility, having alerts, having quotas, having the ability to set up benchmarks. Those abilities did not exist at the very start. When we got the AI credits first, it was just crazy – and you usually reflect on this after you’ve spent all the credits.”
Then the part most teams skip: “The tools are there, it’s just to reach out to them and start using them.” There is a version of overspending that is a tooling gap. There is another version that is simply nobody opening the page.
“Not Really Selling New Capabilities, But Selling More Credits”
Małgorzata has attended the last three Snowflake Summits, which gives her something rarer than an opinion: a trend line.
In 2024 the focus was still new engineering capability, with AI present but not dominant. In 2025 came the Microsoft partnership, integration patterns, and the prospect of the warehouse serving transactional workloads, alongside a lot of data and AI governance. This year, in her reading, the centre of gravity moved.
“This year it feels like it was heavily concentrated on AI. The compute, and the enabling journeys of data engineers and data analysts, speeding up the way of delivery.”
She is explicit that this is not a complaint about the product – “that’s not a criticism; how many new features can you get? It’s still a data warehouse that we’re talking about.” It is an observation about what is being sold.
“Not really selling new capabilities, but selling more credits. I mean, that’s really what it is. If we all jump on those new capabilities, we may end up with huge spikes – huge spikes in how much we’re spending to deliver.”
The Role AI Cannot Automate
Asked what makes AI workloads different from traditional analytics, she declines the narrow framing. It is not only compute and not only cost. For AI to be a value creator rather than a line item, the prerequisites have to exist first: standards, processes, repositories, and a governance model on top. In her words, that is “the one factor that is required to avoid turning an engineer into a tester.”
But her answer to which role becomes critical is the one we did not expect.
“In my opinion, the critical role for this to actually be possible is a business analyst. Everything starts with a business analyst. That’s the gap between the delivery and the actual business, the business process, business need, or the business strategy. And I think that’s something that cannot be automated by AI.”
It is a genuinely contrarian call in a market convinced that the surviving roles are the most technical ones. The logic holds, though: an agent can write the query, but somebody still has to know which question the business is actually asking, and that job is translation, not production.
Build or Buy: A Polite Disagreement
She has had this argument at conferences more than once, and she takes the unfashionable side.
“If something is already built out there, why would you waste your time and rebuild it? Just buy it, save time, get the gain fast. And that’s like a definite answer – you should always do that. I’ve heard that, and I have to say I disagree with the statement.”
For smaller organisations, buy is usually right. For a large one, the question is what counts as competitive advantage and what is simply back-office plumbing everybody has. The risk she names is the one procurement decks rarely mention.
“If I’m compromising my data set by letting somebody else take it and maybe build their product on it, optimise their product on it and then sell it to my competitor – that’s the moment when you think, do I really want to compromise it?”
Her conclusion is a both/and, with one caveat about starting on an off-the-shelf tool to get moving quickly: “It will never be customised just for you. It will always be a universal solution to serve other companies as well.”
A Note for Anyone Selling Into This Stack
She reads most of the vendor outreach that lands in her inbox, or at least the first line. She also told us exactly where she stops reading.
“I’m becoming allergic to AI-generated messages that do not even put enough care to address me as a female. If I’m addressed as a man, that’s the end of my reading. Or if it says ‘he or she’ – that’s to me already the end of the conversation.”
What does work is unglamorous and entirely human: a reference to something she has actually done, a conversation at a conference, or luck in hitting a topic that is live for her right now. “You always have to think that you’re speaking to a human.”
The Bottom Line
The story of the last twelve months is not that AI failed. It is that the enthusiasm arrived before the governance, and the gap between the two was paid for in credits. Małgorzata McCarthy’s account of why is the most plausible one we have heard: leadership saw a feed of other people’s successes, concluded the company was late, and funded urgency instead of discipline. The controls that would have caught it – alerts, quotas, benchmarks – mostly exist now. Someone has to open them.
The rest follows from where you start. A project that begins with “how do we use agents” produces a demo. A project that begins with a business problem tied to strategy has a chance. The cheaper option is often fit for purpose. The role that survives may not be the most technical one in the room. And if you are building any of it on top of data where “candidate” means four different things in four countries, the agent is not your problem yet.
The Data Huddle by Yuki – honest conversations with the data leaders building the future.


