Name: The Real AI ROI Question No One Is Asking
Uploaded: 2026-10-29T15:31:16.417Z
Duration: 35 min 56 s
Description: The Real AI ROI Question No One Is Asking

Transcript for "The Real AI ROI Question No One Is Asking": Welcome, everyone, to the intelligent firm and executive briefing. My name is Brett Tassels. I lead up the product and technology teams here at Deltek focusing on the solutions for architecture, engineering, and professional services, and I will be your host today. Thank you for giving us the next ninety minutes or so. This session was really built for the people who run professional services firms like yours. It isn't a product demo. It's a candid conversation about where I... Where AI is moving the needle for firms like yours and just as important where it won't. Now let me start with a very quick story. And you hear a lot of metaphors like this, but I think this one is especially poignant for what we're gonna talk about here today. About a 130 years old, when factories first got electricity, most did the obvious things. They pulled out the big steam engine, put a big electric motor in its place, and kept everything else the same, same layout, same belts and shafts. And the productivity gains everyone expected did not show up for decades. The payoff really came when managers asked a different question. Not how do we power the factory, but now that that power can go anywhere, how should the factory work? How should it work differently? They put a motor on every machine and rebuilt the floor around the floor of work. That's when productivity took off. And a lot of firms are in the swap the engine phase with AI right now. New tools, lots of pilot projects, and a firm that still runs probably the way it did five years ago. So here's my working definition for today. An intelligent firm isn't the one using the most AI. It's the one that has changed how it operates because of it. Every firm on this call has access to the same tools, but not every firm will get the same results. So with that, here's how the next ninety minutes will run. First, Dindika Hitivalli will take on the real AI ROI question no one is asking. Then I'll moderate a panel from experiment to advantage, includes Carmen Johnson, chief financial officer and administrative officer for OR Collin Associates. She'll share what it actually took inside her firm from governance and process to the people side. And then we're also joined by Marcus Quigley, Partner for Technology and Innovation at EFCG. And he'll bring sort of the industry and research side of things. And of course, Deniker will connect it to what Deltek sees across the market. And we'll finish up with an open Q and A session. So submit your questions anytime in the q and a tab here in the webinar interface and we'll get to many of those questions as we possibly can. The session is being recorded so you can share... You can certainly share it afterwards. So with that, let's get on to our keynote. Dinaker Hitavalli is executive vice president and chief product and technology officer here at Deltek, where he leads product strategy, engineering, cloud operations, and AI innovation. Before joining Deltek in 2023, he spent twenty five years at Oracle, leading development of its cloud ERP applications, and he was building AI and machine learning in the enterprise software long before it was done. Earlier this year, he was named to the Northern Virginia Technology Council inaugural AI fifty one. Full disclosure, he is also my boss, so I'll keep this short. And what I value most about Deniker is, you know, he really never accepts the first answer. He's gonna argue today that most firms are asking the wrong question about AI ROI and he's gonna give you a better one. So please welcome Deniker Hitanalli. Over to you, Deniker. Thank you, Brett. Thanks for the warm welcome. Good morning, everybody. Really happy to be here. Again, appreciate you, you know, giving us the time to let us walk through some of our thoughts and where the industry is going. So, as Ben said, I need, product and engineering for Deltek. I've been there for three and a half years. And before that, I spent over twenty years running, ERP products at Oracle. And I'm saying this for one reason. Right? I mean, I've spent my entire career inside the operating system of professional services firms, financial services firms, you know, the project accounting, the time capture, the billing engines. So when I look at AI in this industry, I really don't see a technology question. I see, you know, and Brett kind of alluded to this. Right? I see an operating model question. And the title of this session today, says that there is this ROI question that no one is asking. And for the first few minutes of my session, I'm gonna show you where the industry is actually heading. So there's gonna be quite a bit of data that I'll walk through, to kind of prove the point. And then then I'll tell you what the actual question is. Right? What's the question that no one's asking? And then, hopefully, I'll be able to kind of show that, you know, what the firms who are actually answering the question well are doing and why I believe that these firms are gonna be difficult to catch. So before we do, all that, we first need to agree on the baseline. Right? So architecture, engineering, consulting, legal, accounting, you know, among all of these industries, you know, what is actually happening to revenue and margins, in these businesses today? So if you look at, the two charts here, right, it kind of shows two different data points, but it actually tells the same story. So on the left, you know, it shows the billing utilization across professional services. So this is from SPI Research benchmark of 509 firms, and it shows that it has declined every single year since 2021. And it now sits at 66%, which, which SPI calls as a nineteen year low. So think about what that means in a business where people are the product. One third of your capacity is not generating revenue. Now on the right side, you see operating profit for architect and engineering firms from the Delta Clarity study of 896 firms. Operating profit has climbed for a decade. Right? 2024 was the peak at 21% of net revenue. In 2025, it fell to around 17%. So that's a 22% relative decline in a single year. And if you see the graph, it's the only decline in the series since 2016 that was not caused by the pandemic. Now, obviously, you know, we we have to be careful not to read too much into, you know, the data based on one year. But the combination actually matters. Right? Utilization has been kind of eroding slowly for five years. Now margin has broken, fee pressure is real, delivery capacity is strained, and demand is uneven. So this is the baseline. So when your firm deploys AI, this is the operating model that it actually lands in. So just kind of hold on to that because, you know, we'll be kind of coming back to that point again. The market picture also is not uniform. Right? If you look at this particular kind of slide, you know, architecture and engineering is down. Professional services, kind of overall down. But accounting is up. Legal is up. Now accounting firms grew more slowly in 2025, and they became more profitable. Now legal has both revenue and profit up. So the question here is what did accounting and legal do differently? They raised prices or they raised productivity or both. And they actually did that faster than the revenue slowdown actually hit them. So this is not an AI story yet, but it tells you which levels actually move the p and l in this industry. It's price and it's productivity, not really the volume. Now this is kind of the money slide. Right? I mean, this is the slide that really, you know, motivated me to give this talk. And by the way, what, you know, what we'll kind of talk about in the slide is not unique to just your industry. Even, you know, us in technology are kind of grappling with this to some extent. So let me get a walk through this. So on the left side, you see that generative AI use among AMD firms went up from 64 to 78% in one year, in just one year. Overall, AI adoption went up from 53 to 70. So big increases in kind of AI adoption and generative AI use, you know, among the employees. So by any measure, the industry itself, you could say that the industry has actually already adopted AI. Right? The adoption problem has been solved. Now on the right side, out of the same 896 firms, 38% can point to measurable business impact. Right? Only 38%. Now McKinsey's global survey of over 1,700 organization kind of landed in the same space. Right? They found that 37% see any effect on the data. And in professional services, specifically, Thomson Reuters found that only 18% of the firms even track any returns on their AI tool investment. And finally, this... The popular kind of metric that we've all seen there. Right? The MIT figure which kind of has been widely quoted, which is, you know, 95% of organizations report zero returns on their AI pilots. And this is a little old. Right? It was last year. But But this has been one of the most frequently quoted and over quoted kind of, you know, data point that, you know, we will see. So here's what I take away from all of this. Right? If 78% of the firms are actually using generative AI and only 37, 38% can show any impact, then the technology is not the variable. Right? Because the same models, the same tools, the same vendors are available to all of them. Something else is the reason why some people get results and some don't. And the reason is this, and this is what we're gonna kind of talk about quite a bit. So the firms are really deploying AI into highly fragmented systems. You know, inconsistent workflows. Greg talked about this as well in his intro. You know, deploying this into highly fragmented systems, inconsistent workflows, into data that was never designed to be read by a machine. And then we expect AI to compensate for these gaps. AI doesn't compensate for these gaps. It actually amplifies the gaps. So an AI tool that you drop into a broken workflow makes the broken workflow run faster, but it doesn't fix the broken workflow. And this really brings me to the question in the title. The real AI ROI question that no one is asking. Now most firms I talk to are asking some version of this question. How do we adopt more AI? Is it more licenses, more pilots, more use cases? And the firms that are actually pulling ahead are really not asking that question. They're asking a very different question, and the question is a two part question. The first half of the question is, where in your operating model does AI actually change the outcome? Not where it can save a few minutes. Where does it change what the firm actually delivers? Where it changes what it can charge or what it cost to them? In a professional services firm, there's only a handful of places where that is true. How do you win work? You know, how do you start your projects and deliver it? How do you bill it? How do you manage the risk? Things of that nature. Everywhere else, AI is a convenience. Right? It's... I mean, it kind of certainly helps, but it's in those places that it's truly a level for your PNL. The second half of the question is how. Right? How will you know if and how AI has actually changed the outcome? That sounds obvious. Right? So if you remember the 18% that we just talked about, right, the percentage of firms who actually track their return on AI, Only 18% of the firms actually track their return on AI. That means four out of five professional services firms are not tracking their returns on AI. And the ones who can, most of them are counting activity. Right? So the counting... Okay. I saved this many hours. I see, you know, this many documents were generated. So they're counting activity. They're not counting the money. And these two questions really depend on each other. Right? If you cannot define where in your operating model does AI actually change the outcome, you have nothing to measure. And if you cannot measure how, you will never find out if you pick the right place in your operating model where AI can change the outcome. So most firms that, you know, I have seen have picked a dozen places where the thing AI can change the outcome. And it's it's mostly by accident. Right? Because, you know, they bought a tool that they got to their vendors, the various systems that they use, and whatever the vendor provided, they kind of decided to turn it on and use it. And most of them have no way to really tell how AI has impacted them or not. So here's kind of the 18% that I talked about. Right? The 18% who do not know or do not track. And here's kind of the full picture. Right? So 42% said, no. They do not track returns on AI. 40% do not know whether they... They don't even know if they're tracking. So if your CFO does not know whether the firm tracks returns on AI, then it's very likely that your firm is not tracking, you know, editor. So so I wanna propose something that's kind of, you know, provocative in an AI conversation, but it's completely ordinary and kind of taken for granted in every other conversation in your firm. Assess each AI initiative by its net present value. So treat it like a capital investment because that's what it really is. And why net present value? Because net present value comes cash kept. Right? It's not minutes saved. It's not activity like we talked about before. So for example, if your associates say, you know, two hours a week, right, they say two hours every week because of using AI, it really tells you nothing until you know what happened to those two hours. So were the two hours built to the next client? Were they absorbed into a fixed fee contract? Were they just written off? Or were they spent kind of reviewing the output of the AI? Right? So depending on what the answer is, once you count the money that you're actually paying for the AI tools and the AI capabilities, the NPV could actually be negative. The other reason why NPV is a good way to look at this is it holds up under every pricing model in this room. Right? Whether it's hourly, whether it's fixed fee, whether it's outcome based or a hybrid of those, it kind of forces you to state the assumption about where that saved our goals. And the firms that do this discover, and they discover very quickly that half of their AI initiatives have no financial path at all, and then they stop them. And that's fine. It's not a failure. Right? That's actually the point. Right? You wanna kind of have the wherewithal to make sure that you can measure your AI initiatives and see if there's a financial path to success. And if not, you call it off. Now so far we've been talking about kind of what's happening inside of the firm. Right? So now let's kind of turn it around and look outward. Right? So Thomson Reuters surveyed over, you know, 1,800 corporate clients, and 78% said that AI enabled quality improvements from the providers are essential or very important. But only 6% said that most of the providers are actually delivering that. And 32% say that they will actually reconsider their relationship with their providers within twelve months if the provider falls behind on AI. And of the 32%, one third of them have more than a million dollars of work at stake, meaning these are sizable contracts. So what does this tell you? So this tells you that kind of the demand side has already priced AI. Right? I'm sure all of you have been hearing this in bits and pieces. Right? Corporate legal departments are explicitly asking for, you know, billing that kind of reflects the AI efficiencies. They expect whatever efficiencies you all gain to be passed on. They expect quality, they expect the speed, and they expect the price to move lower too. Now there's a second half to this that actually gets less attention. Right? If the AI output is poor, that's not free either. So Stanford and Bettor have found that, you know, forty percent of US test workers received low substance, you know, AI generated work from a colleague in the past month, and they each spent around two hours dealing with it. So I'm sure all of you have heard of the term AI slop. Right? That's that's kind of being used quite a bit right now. So if AI slop is what your firm is sending to your clients, then you're just exporting rebook and you're calling it productivity. Right? So the quality and the fidelity of the AI generated work and output is extremely important. So now how do you kind of tell if AI is actually working? Right? So you don't tell by just counting how many people are using it, But by looking for four things. Right? How much of the AI's work is good to send to the client as is? Number one. Number two, how many mistakes it made? How many hours did your people spend checking and fixing it? And finally, whether your clients say the work got better or worse. And these four tell you what AI did to your profit and whether your clients sustained. And this, I would kind of argue, is almost industry agnostic. This applies equally to Deltek as a company, as a technology provider as well. Now AI really rewards, you know, fixed and outcome pricing. Right? Now faster delivery is a financial gain only when the fee is fixed or tied to an outcome. And it's a financial gain only when the backlog is full enough to sell that free capacity. Because under hourly billing, faster delivery obviously cuts your revenue, right, unless you kind of reprice the world first. So if you look at where the industry is sitting right now, right, over 90% of US legal fees still bill hourly. Around 72% of architecture engagements have an hourly component. In these models, right, in these billing models, every hour that AI saves is an hour that you're doing to invoice. Now look at who's already moved. Right? About 60% of Accenture's, you know, 2025 revenue was fixed price up from roughly half three years ago. Engineering firms are carrying a twelve month median backlog of work, which means that any capacity that AI frees up has a place to go and make money. Now even on fixed fees, the gain gets shared. Right? Like we talked about earlier, clients will ask for efficiency to show up in the price, and some of it should. But but a firm on fixed fees negotiates how much of it they can share from a position of strength. But a firm that's on hourly billing just gives it away by default. Right? They don't even have an option to do anything with it. And that is, I think, the cleanest case of, you know, AI amplifying the operating model. The same two in two firms with two different pricing models produces a margin in one and a revenue hole in the other. Right? Very, very simple to kind of understand. Now let's look at kind of the other side of this. Right? So SPI's 2026 benchmark looked at 509 firms by, you know, how far AI had gone into the service execution. Right? So firms with kind of no AI in the service execution showed around 6% EBITDA. Those with experimental use 9%. Kind of selected use cases 10%. And widely used with measurable benefits, 18%. So twice the margin of experimenters, right, between experimenters and widely used, nine to 18%. And on time delivery, the graph on the right, follows the same pattern. Also, going from around 74% for experimental to 82% for, you know, widely used. Now there's a little bit of correlation and causality built into this. Right? Now you could make the point rightly that well run firms, you know, adopt AI more widely and measure it better. So, you know, they were already well run to begin with, so some of the gap comes from that rather than truly just coming from AI. But in a way, that kind of supports the point that I've been trying to make, which is AI just amplifies the operating model that it lands in. A firm that has clean project data, consistent workflows, fixed fee pricing gets 18% EBITDA out of it. A firm without these gets 9%, at least on an average. And here's why I think, you know, these firms will be difficult to catch. Right? The gap is not static. The firm that's at 18% EBITDA now has more capacity to sell. Right? They have better data to feed the next model. They have a pricing model that keeps the game. So every cycle kind of keeps widening the lead. Right? The firm at 9% is spending the same amount of money on AI, but it's falling further and further behind. So what do the firms at 18% actually do differently? We kind of spoke about it a little bit, but that's gonna be kind of the next few sections that we talk about. So what do these firms do differently? Four things. Redesign the workflow rather than retrofitting AI into it, budget for the new bottleneck, contract to keep the game, and start small enough that you can actually measure it. And we'll kind of talk about these, in order. So first, redesign that on the air. Right? So McKinsey asked over 1,700 organizations whether, you know, they had redesigned their workflows around AI. Among the high performers, right, the ones who attributed 5% or more of the EBIT to AI, 75% said yes, and among everybody else, about 25%. So this is kind of the largest single gap in the survey. No model choice differences, no spend differences, no headcount differences, just whether the workflow had changed or not. So like I've been saying, right, when you drop an AI tool into an unchanged workflow, it it can save minutes. Right? No doubt. It's... You get some efficiencies. But nobody builds those minutes. Nobody removes those minutes from the plan, and they just kind of evaporate into the day. And when you redesign the workflow, you decide upfront what happens to the same time. And you need to redesign your billing model as well to fix price so that the saving becomes a line on your p and l because you actually designed it for... To find a place for it to land. Now the clarity study on any firm reaches the same conclusion from a different angle. The firm seeing results picked a very specific business problem. They ran it on connected project data with a very clear success metrics, and they treated AI as a capacity group, not as a seat license. And I would really stress kind of the connected data point. Right? If your if your project financials, your resourcing, your delivery data live in separate systems with separate definitions, then AI really has nothing coherent for it to read. And and that's kind of the fragmentation that I talked about, and no model can actually fix that for you. Second thing, review is the new bottleneck. Right? And I think this is the part of AI conversation that I think is kind of most underweighted. And, you know, I say this as somebody who runs a very large kind of product and engineering organization that's really gone deep on agentic tools and have moved to kind of an agentic software development life cycle. So, you know, I'm painfully aware of this. So two independent studies that's being quoted here. One is the University of Chicago looked at 25,000 Danish workers across, you know, 11 different occupations. Right? And they found that AI saved around 3% of the working hours on average, but it created new checking tasks for 8.4 percentage of the workers. So more people picked up review work than they saved in term term... In terms of meaningful time. Similarly, BCG surveyed around 12,000 employees across 12 markets, and 67% said that they're actually left with more complex work. 60% said the bar for good enough is now higher, and 41% said that they spend more time making decisions. So Microsoft and Carnegie Mellon kind of put it in, you know, one sentence, right, which is effort is shifting from task execution to task stewardship. The work does not go away. Right? It's it's just changing shape from doing to reviewing. And reviewing actually requires the most experienced people that you have. So for all your firms, there's two consequences. Right? One is review is a cost line item on every AI initiative. Right? So the hours that you save in in production, you know, because of AI should partly be reinvested in the review process. Second, the people who are good at judging AI output are actually the senior, the seasoned people whose time was already yours, you know, the most precious resource that you have. And think about this, and this is something that, you know, we all think about quite a bit. Over time, if AI kind of builds the junior pipeline of people that actually grow into the next generation of senior seasoned resources. Then we have a problem five years, ten years from now, even though you're getting, like, a one, two year productivity gains because of AI. Because that pipeline of junior people who learned on the job became seasoned and kind of rose up to become the senior resources. That pipeline is diminishing because of AI. So back to pricing. Right? Like, everything that I kind of said about pricing a few minutes ago is is really theory until a firm changes its contract. Right? So like I mentioned, right, Accenture moved from, like, 50% fixed price work to 60% in three years. Right? So roughly 6% a year. And they talk about it on their earnings call as a deliberate response to AI. Now BCG CEO says that, you know, three out of their four largest AI engagements actually carry outcome based fees. And about a quarter of McKinsey's global fees are now outcome based. And in the broader market, around 40% of professional services firms already sell AI related services in the line of business. So these are firms, you know, with kind of the deepest, hourly billing traditions in the world, and now they are going off of it as fast as, you know, their clients are letting them. Not because, you know, fixed fee is philosophically better or anything like that, but just because under hourly billing, every efficiency gain that you make goes straight to the client. And under fixed or outcome pricing, they get to negotiate how much of that this year. Of course, you're still passing on some of it, but you're in a much better position to negotiate how much of those you pass on. And I would argue that for most firms in this room who are watching this, that's probably the hardest change on the list. Right? It touches kind of partner compensation. It touches client relationships. It changes how you forecast, but it's probably the one chain that determines whether everything else on this list shows up in your margin or in your client's margin. So I don't wanna kind of leave you with a framework. I want to leave you with something that you can actually start next week. Right? So I would say first is pick. Right? Pick one consequential workflow. Not a dozen, just one. Take one accountable owner, a person, not a committee, and one financial outcome that it should move. Right? Margin or cash. And then baseline. Right? Before you switch anything on or you kind of AI enable it, record the net present value assumptions. Right? The quality level, the throughput, the cost, including the review time. You cannot measure how something is being impacted without having this baseline. And then test. Right? You run it, bomb it, run it against a control group, you know, where you can kind of have human review in the loop and count its hours as, you know, part of the cost. And then you decide within one quarter. Right? Don't kind of take months to do it. Right? Within one quarter, you know, decide and expand where the NPV is positive and quality has held. Right? And revise when, you know, the pricing or the review load or the data needs to change and just stop when it doesn't work. Right? Stopping is a legitimate outcome. It's okay to kind of start something, try it out, stop, and move on to the next use case. The firms that are good at kind of stopping more initiatives that they expand, they are the ones who actually realize value faster. So let me give you a quick example, right, to kinda make this process real. So let's take a mid sized engineering firm that responds to around 40 proposals a year. Right? So that's kind of the business case. Right? So 40 proposals a year. Now in the pick part, let's say the workflow that you pick is drafting the proposal. So the single owner is kind of the pursue direct. Right? And the one outcome is cash. Right? The cost of pursuit. Because every hour that you're spending on a proposal is kind of unbilled overhead. Right? So the outcome is cash. Now measure the baseline. Let's say the baseline is around hundred and twenty hours of kind of senior engineer and marketing time per proposal. Right? So hundred and twenty hours for each proposal with a 30% win rate. So that's kind of the before baseline. Now test for one quarter, let's say that you you know half the proposals kind of use AI for the first draft, and they do this based on your firm's past proposals, your history, your terms and conditions, and all of that, and your project data. And the other half kind of run like they did before. Right? So now every hour that a senior engineer spends kind of reviewing and correcting the AI generated draft should be counted against the savings. Right? And at the end of the quarter, now you decide. Right? So let's say that AI came in at eighty hours per proposal. Right? So that's, you know, forty hours saved per proposal times 40 proposals, which is one thousand six hundred hours of a senior resource, right, that you can actually put back into billable work. And now, obviously, that is a positive NPV even against the AI license spend of the first year. And then you kind of expand this. Right? But let's say that win rate had dropped. Right? If the win rate had dropped below the 30% that you started with because the quality of the proposals was bad, then the NPV is negative. Right? So either way, you're gonna know this within a quarter. Right? So this is kind of a good example of, you know, a real case of how you can take a very specific workflow, transform it, use NPV to realize the value, and kind of operationalize it. So kind of the the top of the line in the slide is pretty much summarizes the whole thing. Right? AI amplifies whatever operating model it lands in. So if your data is fragmented, it's gonna give you fragmented answers faster. If pricing gives away the efficiency gains, it's gonna give it away faster. If your workflows are inconsistent, AI is just gonna make them inconsistently faster. So you gotta fix the operating model first. So the firms that are actually pulling ahead are not not pulling ahead because they're kind of adopting more AI than you. Right? They really ask the core question, which is where in their operating model is... You know, where they can actually change the outcome. And they actually built a way to know if the outcome had changed. And then they changed the contracts so that they could keep the benefits and the results of those changes. So thank you, everybody.