The Bottom of the J-Curve: Building AI Judgement Before Buying More Licences with Andy Lambert

Jim Clark: Hi, and welcome to another aibl conversation. Today I am joined by Andy Lambert, co-founder of AcademyAI. Andy, I have been looking forward to this one, because your business tackles something we talk about a lot at aibl. That never-ending problem of firms that bought AI tools, and can see people are using them, but are they actually getting much value from them? And if not, why not? I know this is an area that is very much up for debate, so I am looking forward to chewing the cud with you. Thanks for joining us today.

Andy Lambert: Absolute pleasure, lovely to see you Jim.

Jim Clark: Lovely to see you too. So, in the recent past, you built ContentCal, a social media planning and publishing platform. Then you sold it to Adobe, and spent four years there as it became part of Adobe Express. My question is, what did you see across that period that made AcademyAI the thing you wanted to start next?

Andy Lambert: Good question. We were acquired in 2021, which was already early days, the GPT-2 type of era. We were already experimenting with AI copywriting in our social media products, so we had a bit of a taste for it.

As we got into building Adobe Express, what was fascinating to me was going from, I have only ever been in startups before, to working in one of the world’s best tech organisations. It blew my mind in terms of how you have to think about building product at that scale. When you have hundreds of millions of users in 180 countries, you cannot take decisions lightly.

The bit that started to really interest me was that as AI came to the tip of our tongue, and as everyone in product and engineering started to get more hands-on with the tools, it obviously affected how much of the code we were shipping, how we evaluated product releases, and how we wrote PRDs. Anyone who is a product manager will know the terminology.

What was very interesting to me is that you saw a huge split in the team. Between those who felt like they were living in the future, and those who really just did not want to use AI on anything, because it felt like it was impacting their craft. I think we are seeing a very similar split in general society.

What that creates is a fascinating challenge in how you grow and scale an organisation, because it creates really lumpy adoption. And that means efficiency only happens in pockets of a business.

There is no shade on Adobe here at all, because it seems like every other organisation faces the same challenge. They can only really try to solve it by doing a few more lunch and learns, or having a Google Sheet where people share little workflows that have worked for them. We were just stuck in the era of personal productivity, and that personal productivity was not relating to organisational change. I thought, there has got to be a better way. As most startup stories start with that question, and that was enticing enough for me to think I am going to go and have a try with the next startup in this space.

Jim Clark: You said this is a problem a lot of organisations face, and your experience at Adobe is obviously a big enterprise. I know AcademyAI is focused very much on the mid-market. So were they telling you they were struggling with the same problem specifically? I am sure you had a lot of conversations in the run up to starting this. Or was there a slightly different lilt to it?

Andy Lambert: I did what any other startup founder would. Whilst I saw that challenge at a macro level in an organisation like Adobe, I wondered if it existed elsewhere. That led me into a six month long research project, as a bit of a side hustle.

Building a startup is always just like a big science experiment. You are investigating different things, testing little hypotheses here and there. And it turned out, certainly by the beginning of this year, that a lot of mid-market organisations were feeling this problem even more than the larger organisations.

That led me down a path where it was not just about capability and training. It is about how we actually share best practice across a team. How do we codify ways of working so that everyone can, I call it standing on the shoulders of giants. If someone has a way of working that is working well in a department over there, how do we share that best practice? Rather than just having a Friday Teams meeting to go through it, there is a better way of sharing the workflows that work across the organisation.

That led me into evaluating different parts of our product, which we can cover later if we need to. But it felt like the problem was not just training related. People are missing some of the core understanding, and they are also missing the fact that they have no way of truly learning from others. Peer groups have a massive role to play in this, learning through inspiration rather than just doing another training session. And, as we have seen from so many other folk, finding that room to experiment with these new approaches to how you work.

I have spoken to so many CEOs of mid-market organisations over the last six months where they are saying, we need to do AI. I have used ChatGPT at the weekend and I have seen the future. Everyone needs to be on it.

Unsurprisingly, that does not work. What we really need is a structured way of building capability across the organisation. I really went deep on a big research rabbit hole.

Jim Clark: Sounds like it. Months in a dark room.

Andy Lambert: Correct. Deep on how we build up capability, how, once we have got to a baseline, we then share best practice, and how we build an organisational culture around changing with AI. That gets us away from the situation I started this conversation with, where people have found little productivity hacks. That is fine, but it is the entry point into something much more meaningful, which is organisation-wide change. That is the bit I am really interested in.

Jim Clark: You have these leaders in the market who have hurtled straight into adopting AI for their team. The popular framing that has come out of that is that everyone has access, or at least those that invested have access, but nobody is seeing a return. So what is your reading of that? Where did you first notice that gap in your conversations?

Andy Lambert: That example I shared a moment ago. I will keep the name private, I am not going to name and shame, but a CEO approached me saying, we need to get my business using AI. I now use Claude and everyone needs to use it.

I think we have seen that in organisations too. Lots of enthusiasm from the top, but without deep thought into the fact that this is not just a tool or a technology. This is something that will change how our business operates. So what I have seen is folk not really sitting with that enough. They do not sit with the difficult challenge of, okay, of course we can roll out the tools and everyone can have them. How do we think about this pragmatically? How do we think about what the most critical areas of my business are that I could change, that would allow me to work in a way I had never considered possible before?

That is the reason we are seeing this. People have bought the licences and they just have not had a first-order priority look at what they really want to tackle with it. They are expecting people to solve their own challenges. Or, as I often see, they focus on efficiency gains first. We have got this process, it takes a while.

I have one particular client in mind who said, we need to improve our finance processes, because the monthly reconciliations are taking too long. Fine, you can solve it. But what becomes true once you solve it? What does that new capacity actually allow? Just because we have got people being more efficient, what does that mean?

What I have been pleased about, as a side note, is that I have not spoken to any leaders who are trying to drive efficiency gains in order to remove people. They are trying to make their business work more effectively. Fine. I just feel like we are missing a bit of ambition here.

I think this is also endemic of businesses in the UK versus the US. When I speak to businesses in the US, they are more opportunistic. What becomes true? What can I do more of? What new revenue streams can I open up because of AI? Where can I innovate? AI gives an organisation an opportunity to innovate in a way it never has before. So what becomes true with this new technology?

Right now we are still at personal productivity and efficiency. Fine, we can solve for that. But the biggest thing is that most businesses are thinking too small, which is why they are not seeing the return from their investments. Quite simply.

Jim Clark: That rings true with an aibl conversation we had a few weeks ago. Mid-market firms tend to bank the efficiency gains as savings, rather than putting those savings forward into future investment. Which alludes to what you are saying about the US experience being a little more opportunistic in that respect.

Andy Lambert: Exactly that.

Jim Clark: Going back to the thinking around AcademyAI. You have said it is about using it well, and about the fundamentals AI amplifies. Hopefully I get this right. Judgement, critical thinking, and analysis. Those are things I know companies have a hard time teaching, at least in the formal sense. So what makes AI capability more teachable than judgement itself? That is a big question when I think about it, but hopefully we can answer it.

Andy Lambert: I will try. The words judgement, creativity, and critical thinking are all broad categories that get thrown around a lot, which is fine. And they are right. Someone with good judgement is better placed than someone without, and AI amplifies that good judgement.

But judgement is actually a basket of things, a set of attributes that tree up to having good judgement. It breaks down into a few classes. The first is framing. Can they spot whether this is an opportunity for AI? Is this a problem I need to solve, or is it an opportunity? Have I explored the surface area of possibility? I call that framing, being able to spot opportunities.

Number two is what I call specification. I had an unfair advantage working at Adobe, quite frankly, because working in product at Adobe means you operate with a degree of rigour I had never experienced before in my life. When you are building products for that many people in that many countries, you do not get stuff wrong. And I get stuff wrong a lot. So you need to make sure your thinking is watertight. I am a startup guy, I have always been move fast and break things. That is not how things go down in a megacorp like that.

So specification means writing incredibly detailed PRDs, product requirement documents. I actually think that is one of the best skills I have learned, and one everyone who works with AI should have. It gives you a clear understanding of the jobs to be done. It gives you a research mindset, to make sure you have stress tested the problem with a scientific method mentality.

Just because someone said that process over there does not work very well and we should improve it, rather than saying okay, cool, I will go and improve it, let us critically think about how we improve this. What would need to be true for this to happen? What other evidence do we have that this is an important problem to solve? And then what are the success criteria? How will we evaluate outputs? What is our first order of priority that we need to address? Because of course, you are not going to solve everything in one go.

I put all of that into a bucket of specification. This is how AcademyAI scores people. We have a six axis framework, which we can probably touch on later. Two of those are framing and specification. Establishing whether that is a problem to solve, surface area of possibility, exploration, good questioning, the ability to dig deeper into a problem, defining what success looks like, and evaluating success criteria.

For all of those, you have basically got to think a bit more like an engineer or a product person. Even though loads of folk will not have gone through good product management training. But honestly, it has given me such a ridiculously advantageous view when I think about AI problems. You can see the matrix. You can think three steps ahead and look around the corner. It does not mean I am special. It just means I have been taught by some people who were really, really quite smart.

It is now incumbent on all of us within an organisation, even when the decisions seem really small. Think about a scenario where you want to create a little agent to triage your emails. That seems simple, right? I get loads of emails, so do you. Everyone will face this problem. How could I reduce my email volume? It seems a sensible problem to solve, you would think.

But rather than going onto the tools and saying, hey Grok, or hey Copilot agent, I want you to triage my emails, think about it. It will dutifully go off and do that. But then you will probably come back and say, it has triaged my emails and it is all wrong. Hold on. Wind it back a few steps.

So when I say triage emails, what do I actually mean? And what does success look like? Does success look like inbox zero? Does it look like what is high priority? How should I organise things that come through today? What do I need to organise today versus tomorrow? How will it know my decision-making framework for what is a priority, what is for later, what is junk, or whatever? You see what I mean? For the simplest of tasks, I want something to organise my emails, unless you think deeply about it, the output is going to be terrible.

That means you are then wasting a whole bunch of tokens trying to fix and refine it. You will probably get there in the end, refining backwards and forwards until you get the output you want. But you have spent a whole load of tokens, and ultimately it will have been a frustrating experience. Whereas we could have actually created a better thing.

All of that is wrapped up in that simple question of what good judgement is. You can train for that and you can help people. You just need to break down what good judgement is into its component parts, and help people learn that.

The way we think about it at AcademyAI is through micro learning, building the mental muscle. Because that is all it is, mental muscle. Everything we are talking about is habit formation. So each time AI tempts us by saying it can solve all our problems, it is, no, am I critically thinking about this?

That is the way AcademyAI trains people, through always-on micro learning, Duolingo style. How can I build in 10 minutes, or five minutes a day, that little habit-forming mentality that will get me to critically question stuff?

Because quite frankly, Jim, this is one of the few hills I will die on these days. As humans we are really going to need to get good at questioning this kind of stuff, and at that critical thinking. It is teachable, but all of it comes under judgement. And judgement is very nuanced. That was a long answer, but hopefully it makes sense.

Jim Clark: I like that answer. I see a lot of that from my own experience. If you get the right frameworks and the context rules in place, you can do a lot better, scale up beyond what you did before, and get great results.

My question is linked to that. If you have been through the mill, if you have had that enterprise style Adobe judgement beaten into you and you take the tests, but you barely use AI, could you still score quite well?

Andy Lambert: I think so. Again, this is all through the way we have built this. It is probably worthwhile quickly naming out the six axes we have defined. I say we, but again, standing on the shoulders of giants here, because this is research with the Turing Institute and Skills England that we have used to inform it. So these are the six axes of what I deem AI capability, which covers judgement.

Number one always starts with literacy. Knowing what it is we are working with, and also knowing its flaws.

Number two, safe and responsible use. How do we use this safely and responsibly? I do not think I need to explain that, it is simple stuff, but of course it is getting very important right now. That feeds into judgement too. How do I use this correctly, so I am not just uploading an export from my CRM into a free chat tool, which I have noticed a few CEOs do and have admitted to.

Number three, framing. Is this actually a problem to solve? Is this something that is going to stand the test of time? In that framing piece, I have seen this at so many organisations. People go off and build stuff, and I do not think we are really thinking about this enough. The amount of waste across an organisation of agents that people created but never used again, never ended up maintaining, because they did not solve a big enough problem or never got to work well enough. People build stuff because it feels good. Dopamine hit, I built something great, look what I have done. But all you are doing is creating a load of waste across the organisation.

Also, what happens when people leave? They have built an agent that everyone is relying on. Someone leaves, what happens with the agent? Can someone still manage it? All of those complexities. That is what framing is, what it should look like.

Specification I have already chatted about, so I will not go through it. That is the PRD style of really writing a doc of what it is we are trying to solve, the jobs to be done, the success criteria.

Number five, application. This is working with the tools. Understanding when we just need to use a chatbot, and when we need to orchestrate agents to get it done. Knowing the right tool, task, and model for the job. Super important.

And then number six, evaluation and reflection. The reason I call it evaluation and reflection is that evaluation is valuing outputs, which is very important. It is the new bottleneck in the business that is emerging. Everyone is creating, and then managers are saying, my word, I have got a tsunami of things to review. That is why we are not seeing so many productivity gains. Organisations are just moving the bottleneck elsewhere in the business.

And then reflection, because as individuals and humans, this is probably going to be a terrible statement but I am going to go anyway, humans are inherently lazy. We all find ways to not do stuff. This is again a problem you are going to see in many organisations. AI tools are going to save people time, great. But they are going to defer a huge amount of thinking. So much thinking is going to be deferred into AI tools, and I have already seen and been witness to that inside organisations I work with. We call it work slop, most generally, which is a funny title. But it is going to be a hugely costly problem for businesses if they do not get on top of it.

So that also means evaluation and reflection. Can we reflect on our own usage of that? That is what we built into AcademyAI, through micro learning. Even if you are not using these tools, these are things you can build over time. These are skills, mental muscle that we need to develop.

And over millennia, these are, well, not unfortunately, these are human skills. The reason I say unfortunately is that we should have learned these in school, because this is the critical stuff we really needed to learn. This happens to make quite good entrepreneurs, people who have all of these skills, the ability to spot opportunities, to evaluate success criteria, and so on. But it means it is much more taxing now to be an employee, because we need much greater mental strength around all of these components. Our jobs are changing underneath our feet.

Jim Clark: I have a question on that, and thank you, we moved into my next question. Are these six skills or capabilities universal? Say you have a sales team and a finance team. Would they need those equally? Would safe use not matter more for a finance person than in a sales context? Do you emphasise them differently for different teams?

Andy Lambert: It is a good question. The six axes stay the same, but the role specificness, and I do not even know if that is a word, but we are going to go with it.

The way our product works, and I do not want to get too producty, is role specific. So naturally, how it handles safe and responsible use questions for a sales director is different to how it handles them for an accountant, for example. But the six axes stay the same.

Because with safe and responsible use, again, you do not want a sales director who is thinking, I just need some time, I have got my CRM here, I wish something could sort my CRM better. ChatGPT, here is my customer data, can you find me some insights out of it? Whilst most of us would probably think that will not happen, unfortunately it really does. So the six axes all stay the same, but to your point, they are weighted differently in the product for sure.

Jim Clark: You mentioned before that these capabilities were based on frameworks from the Turing Institute. You were in a dark room for six months in the run up, thinking this through with research. So you know the assessment is valid. But how does a high score predict better work in someone’s actual job? Have you measured that?

Andy Lambert: It is a good question. Yes, but it is still early days, so I will caveat that I have still got a huge amount to learn.

Two examples. Number one is your typical anecdotal example. Take a sales director, and again being very careful and definitely not naming names, a sales director who had never touched AI tools, who went through the fundamentals programme of AcademyAI. I checked in with them a few months after. How are you working now?

Considering the baseline was zero, they talked me through all the ways they had connected different data, and how they had made sure to invite their IT team into the discussion so the data connectors were set up correctly. I thought, this is really good. I am seeing a perfect example of someone who has thought deeply about it.

They were very kind, and I got them to debrief me on the process they went through, and the mental model they applied to building this new way of creating proposals for clients. I am sure everyone listening to this has seen similar examples in their business. But the way he explained he had gone through the process, I thought, he has truly thought about this. End to end. Thought about safe and responsible use, brought the right team members in. So anecdotally, I can see it is working.

The other component, which is more evidence based, is what you can see in AcademyAI as an org owner or admin. CEOs are typically the buyers in our case, or people leading AI adoption. They can see how capability overlays with time saving, because we measure that through an MCP as well, which we will not go into detail about. And also token usage.

So you can start to see how a capability score maps with time saved, aka efficiency. Time saving is still a baseline, it is not the end goal, but it is better than nothing. And then token usage. I think that is really cool, because then we start to throw into question whether some businesses see token usage as a good thing. Probably less so now, we are starting to consider that more deeply. But when you start to overlay those data points on each other, some interesting stuff starts to come out.

Jim Clark: Eventually, and this is probably linked to another question later, it is quite difficult to measure, and maybe a bit political, but potentially better quality outputs and stronger manager ratings could form the basis of some of that.

Andy Lambert: Definitely. High quality output is definitely harder to validate defensively.

Jim Clark: Subjective, is it not?

Andy Lambert: Very much so. But you are absolutely right, Jim, because that is a North Star. We are still at the early stages of this. You just feel like something is going to have to be the glue that sits underneath AI adoption and makes sense of all of this. Whether that is AcademyAI or something else. We cannot just point to a few training sessions, or have a few anecdotes. We need real hard evidence around how our business is getting better, and a platform that underpins the whole thing. Which of course starts with building the capability of our people. So have we got it all solved? Definitely not. But it is a fascinating old journey, that is for sure.

Jim Clark: I am sure if you had it all solved, people would be handing over all their money immediately.

Andy Lambert: Maybe at some point.

Jim Clark: Terribly tricky. So we have talked about the capabilities, and I want to talk a bit more about taking the assessment. How do you design against people getting good at the assessment, rather than being good at the work?

Andy Lambert: A few things. Our assessments are what we would call micro learning. If anyone has used Duolingo, it is similar to that. No piece of courseware, whether it is a video, written, an audio guide, or flashcards, is going to take longer than five minutes to get through. The point is that you are not going to sit through an hour and a half of this, and you have no time in your day. So how can you make a little bit of time every day for this?

This is a five minutes a day type of thing, in the same way that we need to stop thinking about training as a one-off event. It is the same as getting fit. Getting fit does not mean going to the gym once, doing a massively long session, and never turning up again. It is about how we build this into a habit. So that is the first one, form a habit.

Number two is how you stop people just gaming it. The best thing you can do is have a scoring rubric that underpins free type or spoken word. So people will write their answers in, or speak them. I prefer just to speak to the computer these days, and then have the scoring framework underneath it.

The way we have built the scoring rubric means people do not need to parrot the things they have learned in the courseware or the bit of learning. They need to show they have thought it through. Some of our questions are, talk me through the steps you would follow to achieve this end goal. So we are testing how people think about multi-step processes, seeing around the corner.

So through good question design you can do this, and by minimising the amount of multiple choice or true or false. There is some of that, and even with all our true or false stuff, and sometimes matching the pairs because it is a bit more lightweight, there is always a free type justification.

Our model is not just pass or fail. There is a spectrum, where we allow people to pass if they have demonstrated they have understood the learning. But we have an excel level, where someone has demonstrated they have understood the concept and also shown how they can apply it. Again, that is through free type, or we have a sandbox so people can try their prompts. Prompts are only part of it, it is mostly about the thinking. We want people to chat to the computer and say, here is how I would consider this. That is what we score against the six axis framework, to see that they have shown really good specification here.

And we nudge them along. If someone is a little bit lower on framing, or on safe and responsible use, they get a question tailored specifically to the area they are weaker on. So it is like a personal trainer saying, okay, you are pretty good in that component, but you are a little bit weaker here, your form is not quite right, let us help you with that.

Because it feels personalised, and it is not the generic e-learning nonsense we have all been subjected to at some point in our careers, people feel invested in doing it. They think, you are a product manager, here is what you are going to get better at. I am in. Because I do not feel like I am doing this to be compliant. I am doing it for me. It is different.

Jim Clark: Going back to something we covered a little. How comfortable would you be with employers using the results from the assessment to direct the right person to the right project, or even using the results in a performance review? Where do you draw the line on that?

Andy Lambert: That is actually one of the biggest things I think about the most. I built a social media platform before, which is fine, but it feels less critical than this.

Like anything, you cannot take decisions off one data point in a silo. I would want it to be rigorous enough that people consider this as part of a broader picture of an individual. Would I want anyone making a decision on the direction of someone’s career based on this? No. Because you should never make a decision based on one single data point. It is going to need a collection of data points to make a holistic decision.

But for this to work, and for us to stand here and say AI capability matters deeply, then it has to be strong. That is why we spent six months in a dark room researching this and battle testing it with people. So it is strong, it stands up to the test. But again, we should never use single data points to evaluate the success of someone’s career, or where someone’s career goes. That is the way I would caveat it.

Jim Clark: Good answer. As a researcher, I completely agree with you. So, in mid-market firms, a business usually does not have a dedicated L&D function. Who owns this internally? And how would you advise the owners, assuming you have tracked them down, to look at AI training overall?

Andy Lambert: This is always a toughie. CEOs have been our customers to date, and occasionally, hopefully, they have appointed an AI steering group or AI champions, and it is going to sit there. Ideally, someone will have been appointed as head of AI adoption, and they are our people for sure. Because that means the CEO has bought into it, and knows this needs to be owned by someone. It does fall into an awkward gap of change, L&D, and HR.

So that is the best scenario. The best scenario is always where we are brought in by a CEO who wants to take that business somewhere, and knows this is not a quick fix. It is not like you buy AcademyAI and buy Copilot or whatever, and go, right, cool, we are done. This is part of a larger change programme. That also requires a whole bunch of other things to be true in an organisation. A steering committee to drive it, and change champions, because we need people.

Yes, we are a good tech company, but you cannot not have people. You have had a load of great guests on this show who have done wonderful things in their organisations too. But sometimes it just comes down to sitting with people and helping them along. So whilst we have a platform to underpin it, there are a few components to that.

Really, what this relates to is that our buyers are the CEOs who have realised they have a significant opportunity to steal a march on their competitors, and that buying the tools is never going to be enough. Because this is part of change, and they do not buy us for AI training.

I actually never really want to be associated with AI training, quite frankly, because training is a one and done, and it always seems such a low value thing. I hate that, because if I was not doing this I probably would have been a teacher. I just think people undervalue training. So I want to move away from that name, and think about capability building.

Building our organisation’s and our people’s capability is not only the most important thing we need to do right now, given everything you are seeing in the media, all the fear of job displacement, which I do not buy any of by the way. But really, if we are trying to build our organisation, it starts with our people. We need to build our people’s capability, and it needs to be part of a broader change project, not a silo. I have actually forgotten the question you asked, Jim. I went off on a little tangent.

Jim Clark: I think you have covered it off. Where should it be owned? It should be in some sort of governance structure that does not just land with one person. Different functional leads, potentially different departments. So I think you have made a very good point. It is broader change, and it is not something somebody should be working on alone, because it is never going to have the effect.

There are links to my final question, because this leads very nicely into it. A CEO is reading about you and thinking, yeah, we bought the licences, usage looks fine, I do not see any problems. What am I missing? What would you tell them?

Andy Lambert: I would tell them that they are at the bottom of a J-curve, which is currently ROI negative. You bought the licences, you are probably seeing nothing.

I have a few examples of people saying, thanks for my Copilot licence, because I can now summarise my emails better. Deloitte said it for me the other day. We are now spending a billion pounds of our own money in the UK, with 25,000 workers surveyed so far, a billion of our own money bringing our own AI tools to work. 63% of us now use AI tools at work. So it is like, this market is done, we are all using AI, great, our adoption is good.

But what are the three things we are doing? We are summarising emails, we are doing meeting summaries. I think there was something else, but equally transactional. We have the best technology we have ever been gifted with, and yet this is what we are doing? Come on.

That is why I say bottom of the J-curve. You have bought Copilot, whatever the tools are, and you should currently expect it all to be ROI negative. If you do nothing, expect it to continue to be ROI negative.

The only chance you have of hitting the baseline of ROI neutral is when you start to turn those personal productivity improvements into organisation-wide efficiency. Which is fine. And only then do you see ROI positivity, when you start to think about broader change with AI.

The only way you are going to get on that trajectory is by inspiring your team to start thinking more deeply about this kind of stuff. So you can really start to help those opportunities bubble up within the organisation. Because as soon as you start pointing people in the direction of how they should think about it, next time you see that, think more deeply about it, then suddenly those sparks of ideas and inspiration start to trigger across the organisation. And then we can start to do some pretty magical things. That is when we can finally see the opportunity of AI. It takes a village, yes. But hopefully capability building is an utterly fundamental and foundational part of getting there.

Jim Clark: Great answer, and it caps it off well. It links back to everything we were saying earlier. They should be looking at quality of work, time saved, fewer mistakes, workflows changing, and not letting those smaller agentic pilots fall by the wayside with people wasting their time in the first place.

Andy, it has been an absolute pleasure, and I think this is going to make for a great video. Thank you for your time.

Andy Lambert: Thanks, Jim. Appreciate it.

Why do companies see no return after buying AI licences?

Because they stopped at personal productivity. Lambert points out that summarising emails and writing meeting notes are transactional uses of the most capable technology most businesses have ever been given. Adoption looks healthy on a usage dashboard while the return stays flat. His diagnosis is that the licences were bought without a first-order look at what the business actually wanted to tackle, so people were left to solve their own problems in isolation. The result is what he calls lumpy adoption, where efficiency only ever appears in pockets, and never compounds into anything the organisation can measure.

What does Andy Lambert mean by the AI J-curve?

It describes where most businesses sit after buying tools. You have paid for licences and you are seeing very little, so the return is negative. Lambert is blunt that if you do nothing further, it stays negative. You only reach break-even when personal productivity improvements are turned into organisation-wide efficiency, which means sharing what works rather than letting individuals keep private workarounds. Positive return comes later still, once the business starts thinking about broader change. His view is that the only way onto that trajectory is getting people to think more deeply about where AI should be applied.

Can AI judgement actually be taught?

Lambert argues yes, but only once you stop treating judgement as a single quality. It is a basket of attributes, and two of them are teachable directly. Framing is the ability to spot whether something is a problem worth solving, and whether you have explored the surface area of what is possible. Specification is defining the job to be done, stress testing whether it matters, and setting success criteria before you build. He compares it to writing product requirement documents, a discipline he learned at Adobe. AcademyAI trains it as a habit through short daily sessions rather than one-off events.

What are the six axes of AI capability?

Literacy, meaning knowing what the technology is and where it fails. Safe and responsible use, which covers not pasting a CRM export into a free chat tool. Framing, meaning deciding whether this is a problem worth solving at all. Specification, meaning defining the job and the success criteria. Application, meaning knowing when a chatbot will do and when you need to orchestrate agents. Evaluation and reflection, covering both judging outputs and reviewing your own use. Lambert says the six stay constant across roles, but are weighted differently for a sales director than for an accountant.

Why does asking an AI agent to triage your emails go wrong?

Because the request sounds simple, so nobody defines it. Lambert uses it as his standard example. Ask an agent to triage your inbox and it will dutifully do so, then you look at the result and it is all wrong. The questions that should have come first are what triage means here, what success looks like, whether that is inbox zero or a priority order, and how the agent is meant to know your decision-making framework for what is urgent, what can wait, and what is junk. Skip that and you burn tokens refining backwards and forwards towards a worse outcome.

What is the new bottleneck AI creates inside a business?

Review. Lambert describes managers facing a tsunami of material to check, because everyone below them is now producing far more. The constraint has moved rather than disappeared, which is part of why the productivity gains are not showing up in the numbers. He pairs it with a second problem: thinking gets deferred to the tools, and the output quality drops in ways that are expensive to fix later. He also flags waste. Agents get built for the satisfaction of building something, then go unused and unmaintained, and nobody knows who owns them once the builder leaves.

Who should own AI capability building in a mid-market business?

Not one person, and ideally not a function it falls into by default. Lambert says it sits in an awkward gap between change, L&D, and HR, which is why it often goes unowned. His preferred setup is a CEO who has bought into it, a named head of AI adoption, and a steering group with change champions behind them. He is also against calling it training, which he sees as a one-off, low-value event. He frames it as capability building instead, forming part of a wider change programme rather than sitting in a silo.