Hi, everybody. Welcome to this session on LinkedIn Live, where we're going to talk about the future of investment banking analysts and their work as a consequence of the AI tools that are coming into the industry, and actually pretty much in the industry. I want to introduce my colleague, Bogdan Dose, as well. My name is Alistair Matchett.

I started my career in banking at JP Morgan and then worked in private equity for a fund in London, and then have been in the education business since the late '90s. And I don't think in my whole career we've seen an inflection moment like this in investment banking.

The internet, to an extent, pales into, fades into a kind of low-level impact compared to, I think, AI.

Bogdan, why don't you introduce yourself? Yeah. Hi. Welcome, everyone, and thanks for joining. My name is Bogdan Dose.

I've been with Wall Street Prep now for a few months, and before that, I used to do a lot of training in the field. My background's in finance as well.

I used to work in investment banking and M&A for about four years, and then at a hedge fund for about another four years.

My background is a mix in finance and also programming. I do a lot of data science work around Python, VBA, SQL- ... Power BI, and so on. And obviously, the hot topic over the last three years has been AI.

I use pretty much all the tools, Copilot, ChatGPT, Gemini, Claude.

So it will be very great to talk about how we're seeing the shift in the industry, both on the analyst associate side, but also just overall organizations out there adopting AI in the workforce.

Great. Thanks, Bogdan. Let me run through just some slides initially, and then we can talk about some of the key issues.

And one of the things I would say is we're doing a lot of work for the really big investment banks, and we are seeing a lot of the firms doing different things about the implementation using different tools. And so we have a nice idea of actually what works best, what doesn't work well, and what the firms need to do.

So the first thing we want to talk about is that the analyst role is being reshaped, and it says here, "Firms must realize." Actually, I think that's going to happen in probably three months, because most firms have got the tools implemented, but what they haven't done is that they've not really integrated them into workflows. Or even if they have integrated them into workflows, that isn't being used by the current crop of analysts as extensively as it needs to be. Now, when that does happen, I think there'll be a really big change in productivity and potentially how many analysts that you need.

So we've seen this massive adoption of tools within the industry. So over half the firms have actually now run generative AI in production.

But what we haven't seen is we've not seen actually how it's being used to dramatically increase productivity. And I think one of the things here is where is this going to majorly impact? Well, things like creating pitch books, materials, due diligence, even valuation modeling. Those are things that are really going to be dramatically improved and increased in the kind of, the speed of this is going to be dramatically improved.

But going back to what I was saying earlier, we're not seeing the dramatic productivity jump yet.

Most of the firms that we talk to are not changing their hiring plans right now. Although 2027 is already locked in, and now they're trying to figure out what's going to happen in 2028. But because these tools are not being used in workflows, we're not seeing that massive productivity boost, and therefore firms are having a tough decision about what they do for recruiting in 2028. So I think that's one of the first things I like to talk about, Bogdan, is actually what do firms actually need to do to transition from just introducing a tool and actually getting the productivity benefits that we are all thinking will happen in the future? Yeah, it's interesting. So one of the things that you mentioned is the big shift.

Two years ago, a lot of the banks were shying away from using AI.

Over the last year, it's completely changed.

Mm-hmm. I think almost every single bank has adopted if not one, two to three tools, and they've deployed them to everybody at the organization or in pockets, for example, just for investment bankers and so on.

I think where the lag is happening right now is the tools have been deployed, everybody has access to them, but a lot of people don't really know what to do with them.

Mm-hmm.

A lot of people use ChatGPT or Claude at home.

They kind of know the personal use cases and maybe playing around with them or people coming out of school, they know how to use them- Mm ... very well. But in terms of actually applying them to the workflows that are happening on the job, there's that disconnect right now.

Yeah.

And pretty much every single organization that we've talked to, there's a handful of power users that kind of figured it out and they're using it, but there's this big chunk of people that are kind of the laggards.

They may be opening it up every once in a while, using it for summarizing some documents here and there, but they haven't really been integrating it day-to-day in their workflows.

Yeah.

Just from my own experience of using myself and also I do a lot of coaching for senior managing directors, what I've found is that even getting them to do basic stuff like making sure that you've got the model set up, so A, knows who you are, B, you give it some kind of instructions about how you work.

I mean simple things like ... getting your Microsoft Copilot to describe your tone of voice in emails. That's really helpful because then you can put it into your LLM tool and say, "When I ask for an email, make sure you use my tone of voice." And that means you don't get garbage that has those kind of em dashes in, and everyone knows- Yeah ... that comes from AI tools. So there's lots of little things like that that will suddenly make the output feel much more authentic, much more sounding like you.

And that's a minimal thing. The second thing I've also seen is when you're setting up projects or spaces, depending on which tool you're using, making sure there's really good context is very, very important. And that means you want to make sure it has access to all the prior work that you've done in files, but it also has really good instructions.

And this means you're going to get less hallucination the more context you give the tool. The more examples, the more instructions you have in the project, means you're going to get much more valid responses. And if you can automate that through things like skills, that means you're going to have the output that looks like your firm output, it's going to sound like your firm output, it will be able to refer to files through things like connectors, and that means you're going to get much, much better output that will feel exactly like the standard firm output. That takes a lot of work, and I think what we're seeing, well, I said what I'm seeing, some firms have people just dedicated, or teams dedicated to making sure that the model is set up well, it's got great context, it has access to all the internal data, not just the connectors to FactSet or CapIQ, but actually context of all the pitches the firm have done, all the kind of summaries, weekly summaries that the firm's done. All that stuff is really, really valuable, and it creates competitive advantage because you are adding things to the tool that are not generally available on the marketplace, and it means you get better outputs, it will sound and look like the firm, and it will be better referenced, and there'll be less hallucination because it will be actually referencing documents that you've given it.

And I think when we get there, it means you will get a jump because you will get output that doesn't need to be tweaked so much.

Yeah.

And even things like... Well, the other thing, sorry, I've got one more thing to say, is the use of the tools is also important.

Now, where people just go in and do a kind of basic prompt and say, "Build a five-page deck on X," that's a really bad way of using them.

You need to set them up with the context, but also then say, "I want you to give me a spine of what you would do in a five-page deck." So that can be done quite quickly, and you can iterate that before you build out the deck itself. And that gives you a much faster process because if you're prompting to generate a deck, you're using tons of token, it'll take a long time. But if you're getting a spine, then you can iterate the spine, and then when you eventually get it to produce the deck, you get something that is almost there. And I think that it's setting it up, it's the actual use of the tool.

Those kind of two things, I don't see that being there right now in the industry, and that's the big gap is this not setting up tools properly, number one, and number two, not actually using the tool properly. I don't know if you have any insights to that problem.

Yeah. I want to touch upon the valuation modeling piece there, too, that's on the slide.

Yeah.

Like that third bucket, right? Mm-hmm.

That's a very technical piece where AI still isn't quite there yet out of the gate. So we've seen lots of videos and people using these to build financial models or DCFs or LBOs, and you see people one-shotting them, where they'll just go, "Hey, build me a DCF based off of this PDF or these assumptions." And it'll go off and do it, and it looks amazing.

Mm-hmm.

It's creating a free financial statement model. Everything is linked up.

Everything looks great. But when you start digging into it, you can see holes in the accounting or holes in the assumptions and so on.

So that's where I think a lot of people are struggling right now.

People are getting told that, yeah, these tools can build whole models, but then when you ask them to do it, it's not at the level of the bank's quality of work.

Mm-hmm.

So this is the stuff where people need to shift their mindset a little bit, and they have to break apart something more complex like that into pieces.

So when you're doing something like a DCF or an LBO, you can't just do it in one shot. You have to break apart the complicated task into little components, and you teach AI how to do every one of those single components really, really, really well. You save them down to something like Skills or GPTs or whatever you're using, whatever tool, and then you can chain them, so you can reproduce the work again in the future.

One of the biggest feedbacks we're getting from analysts is that, "My work is so niche. I create a model for a telecom company.

I create a model for healthcare. I create a model for a biotech company.

They're completely different types of models, the way I'm organizing them and building them out. How can I automate this? Because there's not an easy way to automate it." The idea is you're not automating the whole model. You need to automate pieces of it, so you can build up that model a lot faster. So that mindset of it, it's almost like a computer engineer, computer scientist. You're taking a part of very complicated code. It's not really code, it's the Excel analysis, but you're saving it down into sections, so you can reproduce it again.

Hmm.

Yeah, and I think with that in mind, I've found if you have an Excel connector or you have an Excel add-in from the tool, you can do that more easily, right? Yeah.

Because you can kind of see what it's doing as it's doing it.

And whereas if you just email the tool, "Create a DCF," it will go off and do, maybe it will use your template, and maybe it will be broadly good. But as you said, if you're in a specific industry which has nuances, you kind of want to build it, as you said, you want to build out, "This is DCF," and the key financial data first, then the free cash flows, then you'll work the terminal value and get your WACC assumptions, and then do the discounting, and then do the- Enterprise value to equity bridge, and those are the steps you can go through, and you want to validate each one. It's much less time-consuming if you do that in steps than just get it to build the whole thing and then check the outputs. Like when you build models manually, right? Yeah.

You don't want to check it at the very end because it just takes five times as long to check it than if you do it as you go through it.

Do you think it will get to a point where AI will do the modeling and the analysts won't do the modeling, and they'll just rely on it because it will get so good? Do you think we'll get to that phase or not? I think we will, but with a lot of input from the analysts at the firm, building up those skills so that they're very fine-tuned to the way the company does it. I remember- Yeah ... when I was in M&A and we were working on live deals, we had all these calls where we're talking to our counterparts, like sometimes these mega deals had two or three different banks advising the client.

Mm-hmm.

And we'd go on calls and just check our comps, check our EV build-ups and so on. And we were having debates like bank to bank, analyst to analyst, like, "Oh, how come you got that multiple? No, it's this multiple." Yeah.

And this is already analysts that have been working for four or five years at the banks. They're disagreeing on how we build up to EV and so on.

So now imagine these AI tools, if you're just using them out of the gate, they've been trained up by Anthropic or OpenAI- Mm-hmm ... cookie-cutter templates. They're not going to produce the way your organization does it. So, yes, I think there's going to be a lot of input that has to come from the analyst to help build up the skills that are very customized to the organization. Once they get there, then yes, I do think that you can one-shot ask it to build- Yeah ... a more complex model with the skills that that organization has built.

Mm-hmm. And I think that takes time, right? Yeah.

Getting the models customized to be able to do that is a big lift, and it needs somebody internally, often, who is really experienced to be able to do that.

And the difficulty is, as ever in banking, those people you want to be working on transactions. You don't want them to be sitting just kind of grinding through, sending out the models.

So there's a big disconnect in actually getting people to set up models well and getting the right people to do that.

Because if you don't have the right people to do that, they're not experienced, you're not going to get a great output.

So I think that's a little tension in the industry, and I've seen a few firms where they have identified some really quite senior people just to work on this.

But I think it's few and far between.

Just in terms of, we talked about the kind of gains here, but in terms of where do you think this will start landing in terms of the timeframe? Because I think this is for the firms that we work with, this is a big issue. When is this productivity going to hit? Because that is going to be, I think, when the recruitment decisions will start to change.

What do you think about recruitment decisions, Bhaktar? My view is I do think recruitment will change, because I think if you-- I just know myself, you can get a lot done in a kind of building decks, building models, and if you know what you're doing, you can identify errors pretty quickly.

So I do think there is going to be a jump in productivity.

The question is, how long is that going to take? I've suggested it by the autumn, we'll really understand whether these tools, how productive they are, and therefore will know in the fall about recruitment decisions for 2028. But I'm interested in your view.

Yeah, it's hard to have a crystal ball on this, right? Yeah.

Because there is a lag, you're right, in terms of the hiring.

People are hiring years in advance for the new cohorts, even with interviews and so on. I remember when I was recruiting back in my days, you would kind of recruit one year or half a year in advance for the next summer's position or the next full-time position. Now people are getting poached first-year university, second year, they're already getting interviewed for a job that they're lining up to do in two years' time.

Yeah.

It's interesting from the HR perspective and what's happening in the field.

I do think things are moving really, really fast.

Where Claude was even half a year ago versus where it is now in terms of the capabilities of what it can do in Excel, it's moved significantly. I do think by the end of the year we'll have a lot of these banks start integrating much, much more efficiently some of these skills. So I think what's going to happen is a lot of the roles at the analyst level at the current banks that have been there for a while, they're going to shift over to the associate-type mindset a lot faster, right? This whole cutting down the workforce because of the tools' efficiency gains, I'm actually on the fence on that because I do think in my mind, this is more like the Excel shift that happened many, many years ago where, sure, Excel replaced a lot of the mundane manual task of accounting and calculations and whatnot, but what happened is people just had more work to do.

Yeah.

You can do more stuff. You push out more and more deals, push out more volume of analysis, right? Yeah.

So I think the size of the workforce is going to be very similar. We're still going to hire waves of people, of analysts and so on.

Mm-hmm.

But what they're doing is completely different, right? They're going to be working a lot more closely with AI to push out content.

Yeah.

And they're going to be expected to do a lot more volume of work.

Mm-hmm.

But I do think that the whole industry is going to change.

Instead of running a model with three scenarios, you can now run a model with 10, 15 scenarios with different components to it, and you can work on five deals at the same time instead of two deals at the same time, right? So, I think that's where it's going to change- Yeah ... what's happening.

Yeah, but I think I'm probably more bullish on Reduce numbers than you, but we'll see, right? Yeah.

We'll see what is going to happen. So, one of the other thing, if you kind of look at what happened with the internet and with electricity, when electricity came in, it came in and people used it and it was kind of cool, but only really had the productivity boost when they introduced small electrical engines, where in a manufacturing process you could kind of do things, have a tool that would do something really, really quickly for you.

And I think this is the same kind of idea with these knowledge worker jobs where you get the tool, it's fine.

You can have to type in and get prompt output and prompting.

But actually when it becomes kind of within the workflow that you're doing, and you can just say, "Okay, I want you to build this and go through the steps.

Let me check it," and it comes out into the beautiful output and it's checked and it's kind of in your workflow. I think that's where you'll start to get really big jumps in productivity.

Mm-hmm. We got a question- Yep ... in the chat there that happy to address, too.

So one of the questions we got, how many banks actually permit these agentic Excel add-ins, though? For what I've seen, banks are currently not doing this extensively due to data privacy concerns. So very great question.

I do think that a lot more people are going to be adopting it.

Yeah.

From the clients that we're talking to, a lot of the big global banks, they already turned it on.

What we're seeing is some banks are lagging a little bit in terms of compliance and IT policies and so on. It's the same trend we saw with adopting even the foundational models like ChatGPT and Claude.

Yeah.

So three, four years ago, a lot of the banks were building their own internal tools because of the same reason, data privacy concerns and so on.

What we found is a lot of the banks have now realized that the security concerns have kind of been addressed by these big platforms, so they're getting more comfortable with just turning it on.

And now most of the big banks have either ChatGPT turned on or- Yeah ... Claude Enterprise turned on. Where it's still kind of in the gray zone is things like using Claude Code or Claude Cowork.

Yeah. Cowork, yeah.

Those have still been disabled on most big banks for now, just because it has control over your whole platform, your whole computer, right? So that's a bigger IT risk.

Yeah.

But in terms of the plug-ins and so on, a lot of them have turned on Claude in Excel, or if you're using Copilot, Copilot has Copilot in Excel with the Claude model underneath the hood. So a lot of the banks are getting there already and they've turned these on.

Do you think, because I think Cowork, I use Cowork personally, and it's unbelievable, right? Mm-hmm.

I got it to reorganize my drive.

Yeah.

I got it to, like I said, "Look at my drive, identify how it can be organized, what things need to be aggregated, if there are any duplicates," and it kind of went through my drive and gave me recommendations, and I tell it to do it, and it just reorganized everything.

And it creates PDFs, it can watermark PDFs, it can do this whole stuff, which is just really manual, and improves it massively without very much input. So I do think, and I know it'd be interesting to see if they will ever allow Claude Cowork, but that's- Well, it's interesting. So on our buy-side client side, some of the private equity shops we talk to- Mm ... the way they've gotten around this is obviously not economical, but just to show how people are pushing, right? They've given some of their power users just secondary laptops with Claude Code turned on.

Really? Oh, wow. Wow. And basically those people will just email themselves or send themselves the information that they want to work on.

Yeah.

So the data room files or whatever they're working on, they'll put it on that secondary laptop, and now it's kind of like a sandboxed environment.

Claude Cowork will go on it and do all the data room organization and so on.

Once it's done, they email it back to themselves on their main laptop, and then upload it to the system.

Interesting.

So you're kind of sandboxing, guardrailing these models- Yeah ... so they don't see everything at your firm and all your inbox and emails and all that, but you can still work with it.

It kind of reminds me of back in the day when, I remember when I was at the bank that I used to work for.

Not everybody had a Bloomberg Terminal, so we kind of had these Bloomberg desks where- Yeah ... basically, you would just go and allocate time on it, and then people would do the work on it, then email themselves the information and go back to their desk and do the analysis. So we're kind of in that back in the '80s and '90s where- Yeah ... you allocate some time on the terminal to work on Claude Cowork.

Yeah. Yeah, I just love Claude Cowork. It's phenomenal.

I'm just going to ask, so one next thing I want to talk about is there seems to be, what I've seen is three different approaches, and some firms kind of mix these approaches, where on the far left of the screen, you've got specialist tools.

And the example of that would be Rogo, which is kind of purpose-built pretty much off the shelf. You still can create skills and things that will make the output feel like your firm, but it's pretty much off-the-shelf tool. Then at the other end of the spectrum is where you have something like Claude as your benchmark, and then you build all the connectors, and then you build the skills yourself to kind of use your formatting standards, your methodologies, et cetera.

And then in the middle is the hybrid, and that's kind of where it's kind of off the shelf, but you actually need to customize it a lot and create the kind of firm version of it.

And then middle I said that would be Heavy AI.

So it's kind of Rogo on the left, Heavy in the middle, and then, say, just Claude standard, and then you kind of do everything yourself.

So I'm just interested in which of these do you think is going to kind of be the lead in, say, two to three years? Yeah, it's interesting. I'm seeing with the very large banks, the bucket A is happening more and more.

Yeah.

And I think it's because of just the scalability of it, right? It's so hard to deploy to everybody.

Yeah.

So they're getting these specialized vendors to come in and almost act like consultants to- Mm ... bring in engineers and work closely with the firm to build out the skills and so on.

Yeah.

Smaller boutique banks or even private equity shops were kind of more on bucket C or bucket B, where they'll just adopt one major platform like Copilot or OpenAI or Anthropic, and then they'll just use their internal people to kind of help build up those skills themselves.

Yeah. Although I just think about the top three banks, I think one of them is using A, one of them is using B, and one is using C.

I'm not going to mention their names.

Yeah.

But the top three banks, they're actually going to spread across these three categories, and it'll be interesting to see.

The one that's using category A is invested in Rogo as well. So Yeah ... I think that was a kind of smart investment because the value of the firm's gone up enormously. But I've actually got, that's a mistake on the slide.

Because I'd say Hebbia is more in the middle there.

So one of the major banks is using kind of Rogo, one of the major banks is using Hebbia, kind of customized, and then one of the major banks is using Claude, I think mostly, and then building its own kind of internal agents.

But I'm kind of in two minds here, because on the one hand, I thought, okay, the specialist tools are really fantastic, but as I've started to get to use Claude more, I've seen if it's customized well, you can give it the flavor of your firm, and you can integrate all the knowledge. You'd think an investment bank has informational advantages, and that's kind of what you're pitching to clients. Why do they go to a big investment bank? Because they have advised on the last 20 deals in the sector, and that's why you choose that firm. So that means you need to have all that information from those deals in your LLM tool to be able to leverage that knowledge, because that's the thing that will give you a differentiated competitive advantage with the other firms.

And you kind of get that more with something that is custom-built than you do with something that's off the shelf. But I could be wrong.

It's just an interesting issue.

Yeah.

Let's move on to one more issue, because I'm just conscious of time. It's gone really quickly.

One of the things that I've been seeing is two things.

The new interns that are coming into the industry, one of the feedback that we've had is that they're very keen on using AI, which is great, but they don't check it. They will kind of even create emails using AI and then just push them out without really kind of going through and checking it. And so there's a quite disconnect between the skill set of analysts moving from not just building it, but actually having to really check it thoroughly, and that can be dull and boring.

The other side of that coin is the managing directors who are asking for the work to be done have got much shorter timeframe expectations of when it can get done, and that's the other side of this is the managing directors are kind of saying, "Well, I need it like in two hours." But they don't necessarily understand that actually it's going to take you a lot more than two hours to check the content thoroughly. So there's a kind of new people coming to the industry that are just leveraging AI and pushing it out, managing directors saying, "I want it like in an hour." And then the analyst is sitting there, and actually what they need to do is check it, and that takes a lot of time.

So that is a big shift. We always had to check stuff in banking, right? Mm-hmm.

But it was easier if you built it yourself, because if you built it yourself, you could kind of remember where everything was, remember the numbers- Yeah ... and it was easier to check. You have an output, and to your point, you were saying you need to be an associate.

Well, that's difficult for a new analyst to become an associate suddenly.

It's questions, how do firms actually do that? Yeah, and I'll answer that too, but I'll just take it a step further too.

Aside from just checking the work it's doing is just having a really good grasp on the company or the industry itself, too, right? So I remember when I used to build models in M&A, part of the grunt work of pulling all the data from the financials and grabbing it and feeding into the models, you become an expert on that company, on that industry.

So when you're going into meetings and people are pushing you on your assumptions and so on, you have a good understanding of why you built it the way you built it, right? Mm-hmm.

What we're finding right now is people are speeding up for the analysis so fast, like you're just grabbing all, extracting the data from the 10-Ks, the 10-Qs, building the model very fast. They have superficial understanding of what's going on in the company.

So if they get pushed in meetings and then like, "Well, I don't know.

I don't know why that's the number I used there because Claude did it for me." You can't say that in a meeting, right? So I think it's twofold. Making sure that your answers are correct, that you don't have hallucinations in there and so on, and you checked every source. But also, once you've saved that time building the model faster, you've got to spend more time now understanding what are the different scenarios, why do you have those scenarios, and so on.

So, yeah, I think part of this is educating the MDs as well.

We're seeing this. A lot of our firms are asking for more one-on-one coaching with- Yeah ... the partners, just so they understand, not necessarily for them to use the tools. A lot of the times they'll use the tools as well too, but just so they understand how are their analysts using these tools, what can they produce out of them.

Yeah.

But the other side is the new hire training programs we're doing and the summer programs is making sure all those new hires have very good grasp on the fundamentals.

Sure, AI can help speed up the work, but unless you have a good grasp on the accounting, on the fundamentals of linking up the models, you're not going to be able to catch the mistakes- Yeah ... these tools are making. So it's kind of twofold, both educating, upskilling all the new hires, but also managing upwards a little bit and kind of telling the MDs, "Hey, look, you're saving some time on creating these things, but some of that time now becomes associate time where you have to go back and check the work." Unfortunately, we're out of time, Bogdan, but we are at our half-hour end. What I would just leave with is that we are going to do more of these LinkedIn Live sessions.

We're also running a two-day AI analyst program, and it's two days because you can get so much more done in two days, where we'll literally do peer group analysis and building a three-step model, doing a valuation literally in two days using AI, and we will have Claude with a Felix connector, which has access to all the financial data. So that's coming up in July, and there's one in August in New York, I think, as well. So if you want access to that, just go onto our website. You should be able to see that in our public courses.

But thank you, Bogdan, and thanks everybody for joining us today.

I know this was short and sweet, but we know everybody's busy, and the weekend's coming up. So I hope everyone has a really great weekend. And Bogdan, thank you for your insight. I really appreciate it.

Awesome. Thank you, everyone.