The Inversion: Designing AI Around Human Judgement with Amin Mrini, Informa

John Emmerson: Good morning and hello everybody. Welcome back to another aibl Leadership Series interview. I am joined today by Amin Mrini, Chief Commercial AI Officer at Informa, one of the world’s largest events and media businesses. Thank you very much for joining me this morning, Amin.

Amin Mrini: Thanks for having me, John.

John Emmerson: I am going to go straight into some questions. Most businesses globally are now using AI. They are using it to do the same things they have always done, but faster or cheaper. You have argued previously that the real opportunity is in what becomes possible when the cost of execution collapses. What do you mean by that, and what does it open up for mid-market businesses?

Amin Mrini: This idea of doing things more cheaply is incredibly short-termist. It is one-dimensional as a way of trying to unlock AI value. AI is a multiplying force for productivity and output per person. It does that in ways that help organisations think about new forms of demand, and new products and services.

The goal should be dollars of revenue, or ARR, per head. Not trying to minimise the heads per dollar of profit. A cost impact is a one-off by definition. A new product compounds over time.

For those thinking of AI only as a cost-saving lever, I am fairly sure we will have to go through a phase where that is the primary benefit. It is easier to measure. But in any market anywhere, there is either a set of AI-native companies or legacy providers with a better approach to AI. They are trying to unlock its value in the way I would describe it.

Being protective and thinking about costs only is a defence strategy. It will affect your market share and your long-term growth potential. Someone out there is trying to eat your lunch. They are thinking about new products, new services, new demand, and new ways of distributing content, information, and products. That is something you have to be doing as well.

John Emmerson: I think it is a really interesting point. We speak to a lot of people from different types of business across a range of sectors. There really are two distinct schools of thought here. There is what you have just described, which asks where the opportunity is. How can we use this disruptive new tech to grow and scale at multiples that would have been very difficult in the past?

Then there is the other school of thought, the one you are arguing against. How can I save some money? How can I trim headcount? How can I improve processes?

There are merits in both of those arguments. But if you have a growth mindset, you surely want to run at the opportunity this presents. Have you got any examples of really interesting things you have seen people doing in that field?

Amin Mrini: There is a very long list. Thankfully the cost saving, and frankly the AI washing, has dominated the media narrative. So you hear a lot about cutting waste in big tech. I think that is more of a correction for over-hiring than actual AI impact.

When you talk about mid-market, there are plenty of small to medium-sized companies that are, if not AI native, AI nimble. They are thinking about talent, and about developing their existing staff, in exactly those ways. How do I increase output per person?

In functions like sales and marketing, primarily, a fairly clear blueprint is starting to appear. It shows what the AI stack, and the new AI-enabled ways of working, actually look like.

John Emmerson: I agree. There are a lot of people trying very hard to create an efficient and effective AI stack. Then there are loads of other businesses just bolting any AI tool they can find onto existing workflows. What we find is that this pushes the problem somewhere else, or replicates the problems the business already had.

Have you taken the time to step back and redesign the flow around the new tech? Is that something you see people doing? Or do you see it as: let us define a problem, then find a tool that will solve it? What is the best advice you can give people?

Amin Mrini: It is what half of my job consists of. First of all, we need to ban this AI use case term. People turn up wanting to hear about what the use case is.

The task-based approach does not work. Let us brainstorm a hundred different ideas, then build discrete AI tools, which are mostly probably chatbots, and then hope for the best. Hope that we are more productive, or that I can take out 10 to 15% of costs. It just does not work.

A simple example. I can create presentations more quickly. But if the meeting in which the decision takes place is still a monthly meeting, you have not saved time. You have just given people more time to over-polish an asset, or to do other things that are not necessarily going to unlock value.

So this task-based idea of automating tasks, or finding the AI use cases, or thinking that giving a tool to everyone is a strategy for AI, is not going to lead anywhere. You might be making people slightly more efficient at one thing, in very tiny pockets. But the value is only unlocked if you start chaining tasks. That is why we start talking about workflows.

The reinvention piece, to me, comes from looking at the problem the opposite way. In the past, companies have started by thinking about what they can help humans do more quickly, or more effectively, or more cheaply. The question I ask is different. What do humans absolutely need to stay in charge of?

Once you protect that, or take that out, you delegate and find ways of giving the rest to AI. What in this chain of tasks or sequences needs to be a human decision, or a human approval? What needs to stay in the hands of someone to create an output and then hand that over? Everything else, to the extent that it makes sense for the workflow in question, can go to an agent.

Obviously there are accuracy problems, integration problems, and compliance problems. But the rest, to some degree, will be handed to an agent to perform. So look at the problem the other way around. Simply thinking about how I speed up outputs the way things were done before, with the humans who were in charge of that task before, is not where the value is at all.

John Emmerson: What would your advice be to people trying to decide where humans need to stay in this loop?

Amin Mrini: This is where AI fluency comes in. There is a base layer of judgement, and of knowing your processes. Knowing the level of risk you are willing to accept. Knowing where the approval gate is, or what your governance framework dictates. That is number one.

Then there is a good enough understanding of what AI is good at, and the extent to which it can perform the task accurately. So you need to understand the limits of reasoning in the model, the limits of accuracy, and the good enough equation you are willing to accept.

If you are reconciling invoices, where does the needle shift in terms of productivity if you can handle the 20% of hardest cases that are currently reviewed manually? What if you can do 50% or 80% of them? And if you are left with just 20%, what does your team look like after that?

So it is a combination of understanding the process, the need for accuracy, the risk profile of delegating to an agent, and a good understanding of what AI does well and reliably. That comes from experimentation. It is not something you either have or do not have. Just try it out.

John Emmerson: It changes all the time. Even in the past forty-eight hours, a new Claude model has been released that seems to be a lot better at certain things. So you are constantly reviewing what the model or tool you are using is actually capable of.

What we talk about is encouraging people to use AI for things that are repetitive and administrative, that do not require strategic thought, creative thought, or relationships. Our belief, and one of the things we say all the time, is that using the tools now at your disposal allows humans to do the things they are really good at, and machines are not so good at. For us that comes back to strategy, relationships, creativity, and socialising. Socialising has a less business focus to it. Is that something you would agree with?

Amin Mrini: On your first point, I do not think people should panic or read too much into the four to six week new model waves. They should look at capability improvement over a slightly larger window. Six months to a year, not every cycle.

Because I bet you that the absolute vast majority of your processes do not require frontier models. Unless you are in healthcare looking at a drug pipeline, unless you are in advanced engineering, unless you are Airbus, I guarantee it. First of all, that comes with a cost advantage.

It also means you do not need to worry about keeping up with the frontier, or paying for state of the art. Models are increasingly looking like an exponential curve. But on a day-to-day basis, for most of us and for most workflows, you are covered. I would spend more time thinking about how I ride the cost curve down than how I ride the capability curve up for a given workflow. That is the first point.

Then there was creativity and judgement. It is a good idea to start with what is repetitive and codified, and has the right context layer around it. That makes sure you are making the right decisions, or performing tasks in the way that you want.

As it relates to strategy and creativity, creativity is intrinsically a human attribute. So I do not know that it is even the right word. Strategy as a hands-off, tell me what I need to do, probably not. But as an accelerant for humans, yes. Same for ideation, brainstorming, and prototyping. I would not call that creativity, but it is not far from it.

Increasingly I find AI to be very capable. There are more and more things where I start with an idea, and then I am given a highly refined version of it. Something it would probably have taken me a while to find, and I might not have found it at all.

Ultimately it is not AI versus non-AI. It is whether I can comfortably outsource something to AI. Then there are things that require an AI and human pairing. I do not think there are many non-AI fields, apart from human contact and relationships.

John Emmerson: I like the way you have described it as an accelerant. That is certainly something I do myself. I use it to get an idea off the ground, or to refine an idea and move it through my thought process quicker than I would have before. I like that description.

There is a lot of noise about how people measure the return on value of any AI tools they are buying. There are figures out there from Deloitte, Forrester, McKinsey, and various other places, saying 94% of businesses are not seeing any return on value. You have said previously that without measuring the baseline first, you will never be able to prove that AI has delivered an upside. Thinking about smaller businesses, they may not have a dedicated data team, or an in-house team to do that measurement. What would your advice be to them?

Amin Mrini: In a way, it is just P&L. When you are launching a new product, you are accountable for how it performs. For how it changes the customer journey, satisfaction, financials, you name it. So there is nothing new there.

You need the ability to attribute a success metric, and to track how KPIs evolve with the introduction of AI inside a process. If you are automating content creation for a certain flow, for a certain persona, you need to measure throughput. You need to measure content interactions, views, clicks, time on site, repeat visits, and so on.

So the baseline thing is to make sure you start with a clear understanding of where your KPIs are. Otherwise, how would you ever see an uplift if you do not know what the status quo looks like?

There is nothing more sophisticated here than what anyone introducing a new product or service has had to do for years. Yes, attribution is difficult in some cases. There is the causation versus correlation problem. In my experience, it is pretty clear whether it is working or not.

You cannot do without the numbers. But I have not had a hard time getting comfortable with the numbers, even when it is a partial view, or when you cannot fully close the attribution loop. Done well, there is nothing I have done that was such a tiny incremental gain that we were not quite sure it was worth it. There is so much to be redesigned, so much to be rethought.

If you are starting with the things that drive the extra 0.01%, and you are not quite sure whether it sits within the margin of error, I would argue you are probably not starting in the right place. So it is not as complicated as you could make it sound.

John Emmerson: I agree with that. If you bring in any new tool, system, or process, you need to measure it. This is no different. Just because it is more technical in nature does not mean it has to be hard.

Amin Mrini: The AI value debate comes from an idea that is still very present. It goes like this. I have an AI strategy. I have given everyone GPT or Claude. So everyone is burning tokens. If you say everyone is going to access this, free for all, then by definition you are not quite sure what you are trying to achieve.

If your strategy is everyone be more productive, here we go, here is a chatbot, then what are you trying to shift? Everything all at once? Some intangible sense that we should be more productive, or that we could do more with less? That is impossible to measure.

The strategy of I want to rethink this workflow is different. I want to increase the throughput of my sales team. I want to bring conversion up. I want to drive pipeline growth of X%. I want to be able to manage a book of business with 20% fewer sales reps, and the other reps will be doing something else. That is something you can measure.

People have approached it from a general purpose position. Let us have fun, and trust people to sit there and reimagine their jobs by themselves as they do it. It cannot be everyone’s responsibility to rethink. Everyone now says reimagine the workflow, reimagine the workflow. It cannot be an individual contributor’s responsibility day in, day out. No one wakes up and thinks about how to rethink their job.

So it is the approach, and how AI was deployed, that is creating this fake debate around AI value. Also because people, for some reason, thought that burning tokens was a sign of pride, or something that equates to AI fluency or literacy.

The recipe is so obvious it almost pains me to have to say it. But I still hear it on some pretty big podcasts out there, presented as a stroke of genius. Very small teams. Engineering capability. Sitting down with the functions, and then embedding AI behind the scenes inside operations.

So you bypass the whole adoption challenge. You bypass the whole change management challenge. And you preserve your ability to chain and transform whole sequences. So you need technical fluency. It cannot be everyone’s responsibility. You need people accountable for that change, and people embedded within functions for that change. That is why we have the whole FDE craze and mania right now, and all the deployment shops and consultancies rebranding themselves.

So there is a clear way of unlocking value. The debate just stems from a complete lack of strategy as to how AI was deployed.

John Emmerson: That is really interesting. It is something we have spoken about a lot with other people. There are a variety of ways people are deploying it. Sometimes it sits with the CTO, sometimes with the COO, sometimes with the CHRO, treating it more as an HR adoption challenge. I think there are probably merits in all of those.

Having accountability is important when you are trying to do anything in business. That is an absolute given. But your suggestion that you break it down into segments, go into the teams, understand the challenges, and then embed it piece by piece, seems very logical to me.

Amin Mrini: Whoever it sits with needs to come with the realisation that it really cannot sit with one function. Technology is responsible for technology. So the moment an AI team sits in technology and is responsible for reimagining and rewiring processes, there is a problem. Technology is not responsible for the entire process wiring across functions in the business. Nor is HR. Nor is anyone.

First of all, the mandate, the story told around AI, the investment behind AI, and what the company is trying to unlock via AI needs to come from the CEO. That is number one.

Number two, yes, the AI function sits somewhere. Unless it is a direct CEO report, it tends to be in technology. It needs to come with the realisation that it is in technology because the engineering skill set, the product skill set, and the delivery skill set tend to be there. Fine.

But every function is responsible for the outcomes that AI engineering or re-engineering is unlocking. That function is a partner. It does not own the entirety of the business process outcomes across functions. It just cannot work like that.

So there is a level of AI fluency required in non-technical functions. There is a level of AI engineering capability required outside of technical functions as well. So there is a bit of a rewiring of the organisation needed. If not in the org chart, then at least mentally.

John Emmerson: I could not agree with you more. You have mentioned it a few times, this idea of AI fluency across the board. It is so important. How do you enable people, and help people become AI fluent? That comes back to the way you adopt it, and to training and development. The same stuff as always.

A couple more questions. You have mentioned previously that AI is squeezing the capable generalist. What do you mean by that?

Amin Mrini: I think it was the pseudo-generalist. There is a profile that is quite common in most organisations. Someone in charge of coordinating and managing, without owning a domain expertise.

When the cost of execution was high, and when domain expertise was quite siloed, you needed the manager. Someone sitting on top, coordinating, scheduling meetings, reporting, making sure the handover between those different silos was managed effectively.

Those silos are getting closer to one another, and in some cases they are merging. And execution is not scarce anymore. It does not need to be micromanaged as a rare occurrence in the business. It is no longer so hard to make things happen that I need three layers of management to make sure people know what to do and are doing it well.

Those without a really deep domain expertise, and those without the judgement to be proficient across domains, are going to lose value. So either you are a hyper deep domain expert who is able to challenge AI output, to refine AI output, and to see mistakes or inaccuracies. Or you are someone who is able to bring down the frontiers between functions.

Take a product manager who is technically fluent enough to prototype, to look at design, to look at research, and to look at launch planning. Why would you need a UI designer, a product designer, an engineer, a sales enablement specialist, and a marketing manager today to launch a product? I am pretty sure you could compress those five or six roles into two or three, if not more. That is how I have built my team.

So I think we are entering the age of the hybrid. AI augments exactly those profiles, and those are the people who will thrive. Those who have combined one deep technical domain expertise with the ability to synthesise and operate across domains.

By extension, those who are not particularly hands on, those who are not individual contributors, and those who are just the lubricant between layers of an organisation that is collapsing and will flatten out over time, will lose relevance.

John Emmerson: I heard this about ten or fifteen years ago. I was working in a sector where people were talking about the demise of middle management, which is similar to what you are talking about here. That lubricant layer. I think you could be right this time.

You have done this, so this is a good question to ask you. I imagine hiring genuine hybrids at this moment is quite hard. Is that right, or am I wrong about that?

Amin Mrini: No, it is very hard. Very hard.

John Emmerson: So how do you find them?

Amin Mrini: I wish I had the recipe. I would be more successful at it. Look, the good news is that we are about to find out. You have companies like Block, for example, who have decided they could do the same job with 40% fewer people. No middle managers. A context and information layer managed entirely with AI. That is effectively the equivalent of replacing that layer of coordination, of meetings, of information gathering and information sharing.

They are not the only ones. So we will find out pretty quickly whether the thesis is right or not. Both things can be right at once. We could see a rise of the new hybrids without losing those middle management positions for a while. We will find out.

To come back to your question, unfortunately there is nothing new there. In terms of hiring good people, there have always been hybrids out there. An AI-augmented hybrid is a much better hybrid, which is the point I am making. So having one helps you find the second, and having two helps you find the third.

The profile of the hiring team, to me, is the biggest differentiator in finding the next best person who fits the bill. Because people with that skill set want to work with people who have it, recognise it, and value it. Obviously that does not solve the first hire. N and N plus one is fine, but it does not solve the first. In most cases you will not find those people. If you do find them, they are quite rare.

So ultimately it is a lot more about upskilling and training. How do I assess performance as it relates to AI usage? How do I upskill my own team on AI usage, and turn your people into these hybrids? You will not start with a clean slate and find a full team of hybrids out there. If you find a small number of them, use them to upskill your team.

I am pretty sure that upskilling is more important than hiring, purely based on the volume and the time it takes to churn and then replace. So I do not have a magic recipe. The channels for hiring that work well are word of mouth, and being constantly out there interviewing and meeting people, whether you have open positions or not. That builds a pipeline, because you never bump into the right people when you need them. Then trust that once you create that culture and bring in that skill set, great attracts great.

John Emmerson: Great advice. We are out of time, but I am going to ask you the final question I ask everyone on these. Someone watching this is at the start of their AI adoption. They are a senior leader in a mid-market organisation, and they are not as advanced as you. What would be your single piece of advice to them about how they take on this opportunity and challenge?

Amin Mrini: It is hard for me to pick one. Start with simple but transformational goals. If you are starting from scratch, by definition there is a lot of low hanging fruit that can transform productivity. That could be in your commercial performance, or your product performance, you name it. But simple, clearly stated, and measurable goals.

You will not transform the organisation in one go. And again, giving people individual tools is not going to achieve that. So do not underestimate the engineering challenge that this represents. Invest in that technical capability.

Then embrace what is now commonly referred to as the FDE model. That means embedding a very small number of capable engineering and product people within functions, then driving the change from within that function, and thinking about AI behind the scenes. The less AI you see, the better for productivity. You are not building tools for the masses, where everyone spends their day across four or five different tools, and every task calls for a new one.

Those would be my main pieces of advice. In terms of ways of thinking about how to deconstruct the problem, and the right model for unlocking AI value, those are the ones.

John Emmerson: Wonderful. Thank you so much for your time today, Amin. We really appreciate it. Wishing you and Informa the best of luck as you use AI to advance your business over the coming months and years.

Amin Mrini: My pleasure. Thank you very much.

John Emmerson: Thank you.

Why is treating AI as a cost-saving tool a mistake?

Mrini describes cost-led AI as short-termist and one-dimensional. A cost saving is a one-off by definition. A new product or service compounds over time. His argument is that AI multiplies output per person, which lets a business create new demand rather than defend an existing margin. The measure he prefers is revenue per head, not headcount per pound of profit. He accepts that most companies will pass through a cost phase first, because savings are easier to measure. The risk is competitive. In any market, an AI-native rival or a faster-moving incumbent is already building the new products.

Why does the AI use case approach fail?

Mrini would ban the term. The pattern he describes is familiar: brainstorm a hundred ideas, build a hundred discrete tools, and hope 10 to 15% of cost disappears. Most of those tools end up being chatbots. His example is a presentation built in minutes for a meeting that still happens monthly. No time is saved. People simply over-polish the asset. Value appears when tasks are chained together, which is the point at which a business is talking about workflows rather than tasks. Automating single steps inside a process nobody has redesigned moves the bottleneck rather than removing it.

How do you decide which tasks humans should keep and which to give to AI?

Mrini inverts the usual question. Instead of asking what AI can do faster or cheaper, he asks what humans absolutely need to stay in charge of. Once those points are protected, everything else can be delegated. In practice that means naming the human decisions, the approval gates, and the outputs that must be produced by a person before handover. Answering it takes two things. First, judgement about your own processes, your risk appetite, and what your governance framework requires. Second, a reasonable grasp of where AI is accurate enough. That second part comes from experimentation, not from reading.

Do mid-market companies need frontier AI models?

Mrini is direct on this. Unless you are running a drug pipeline, working in advanced engineering, or building aircraft, the vast majority of your processes do not require frontier models. He advises against reacting to every release on a four to six week cycle. Look at capability improvement over six months to a year instead. Two benefits follow. Costs fall, and you stop paying for state of the art you do not need. His summary is worth repeating: spend more time working out how to ride the cost curve down than how to ride the capability curve up for a given workflow.

How can a smaller company measure AI ROI without a data team?

Mrini treats this as ordinary P&L discipline. When you launch a new product you are accountable for what it does to performance, satisfaction, and financials. AI is no different. The requirement is a success metric you can attribute, and a set of KPIs you track before and after. If you automate content production, measure throughput, interactions, clicks, time on site, and repeat visits. The baseline matters most. Without knowing the status quo, an uplift is impossible to prove. Attribution is imperfect in places. His view is that if a gain sits inside the margin of error, you started in the wrong place.

Who should own AI in a mid-market organisation?

The mandate sits with the CEO, in Mrini’s view. That covers the investment, the story told internally, and what the business is trying to achieve. The AI function itself usually sits in technology, because the engineering, product, and delivery skills are already there. What matters is that it does not own process outcomes across the business. Every function stays accountable for the results that re-engineering produces. He also recommends the forward deployed engineer approach: place a small number of capable engineering and product people inside a function, then drive change from within. Embedding AI behind operations sidesteps much of the adoption problem.

Which roles does AI put at risk, and which gain value?

Mrini points at what he calls the pseudo-generalist. Someone who coordinates, schedules, chases, and reports, without owning deep expertise in any domain. That role existed because execution was expensive and silos sat far apart. Both conditions are changing. Two profiles gain instead. The deep domain expert who can challenge and correct AI output, and the hybrid who works across functions, such as a product manager fluent enough to prototype, review design, and plan a launch. Hiring genuine hybrids is very hard right now. His practical answer is upskilling, because churn and replacement take longer than developing the team you have.