Why Your AI ROI Numbers Are Probably Wrong – And What to Measure Instead

Forethought’s Antoine Nasr discusses how outcome-based AI is reshaping the way enterprises measure the true value of customer service automation 

In this CX Today discussion, Associate Editor Rhys Fisher sits down with Antoine Nasr, Head of AI at Forethought, recently acquired by Zendesk. The pair discuss Forethought’s agentic AI push to tackle one of the most uncomfortable conversations in the CX space right now: is the ROI you're reporting on your AI investment actually telling you anything useful? 

Twelve months of AI euphoria in the enterprise is giving way to a harder question: was it worth it? Antoine Nasr makes the case that most organizations are measuring the wrong things, and explains what Forethought is doing differently: 

Deflection is a broken metric: A deflected ticket tells you nothing about whether the customer's issue was actually resolved. Forethought's outcome-based pricing model only charges when the AI agent genuinely resolves a conversation, and shows admins exactly why each one was classified the way it was. 

The $1B ROI figure? Already outdated: Antoine is candid that Forethought's headline customer ROI number has already been surpassed, and walks through the framework for calculating both the hard dollar savings and the softer but real gains in CSAT, brand value, and churn reduction. 

AI as a single entry point: User expectations have shifted. Customers no longer want to navigate your org structure, they want one interface that routes intelligently across support, commerce, sales, and beyond. Forethought's RevTech integrations are built around that reality. 

Self-improving agents are here: Forethought's Discover product automatically identifies support topic gaps, builds agents to handle them, and improves after every interaction, closing the loop between performance data and agent behavior.

Transcript

SPEAKER_01 0:08

Hello and welcome to CX Today. I'm Reese Fisher, Associate Editor, and today I'm delighted to be joined by Antoine Nassau, the head of AI at Forthought. Antoine, thanks for joining me. How are you doing today?

SPEAKER_00 0:21

Hi Reese, thank you for having me on. I'm doing very well. How are you?

SPEAKER_01 0:25

Yes, yes, I'm really good. Thanks. I'm looking forward to today's chat. You know, we're going to be talking about integrant integrating agentic AI with RevTech stacks and how that's impacting the CX space. So I guess before we maybe dive into the topic, I thought it'd be helpful for you to use your uh your expertise to give us a bit of an overview of this topic. So could you maybe just provide us a little bit of a background around this and speak to perhaps why this matters so much right now?

SPEAKER_00 0:52

Of course. Uh as just as an introduction to forethought, what we do is AI agents for customer support specifically. Uh we were recently acquired by Zendesk to help accelerate uh their adoption of Agentic AI. And what we've seen over the last 12 months is that there has been a massive amount of attention and money that has gone towards implementing AI in the enterprise. Uh the one of the first uh destinations of that attention and money has been customer support. This is for a few reasons, including the fact that, well, it's easier to measure uh the impact and the ROI. Now, 12 months in, uh, as expectations around what those tools can do uh and uh both from executives and from and from users have increased, we are now in a position where we're finding people wanting to understand the actual ROI of this investment, uh, wanting to see how it impacts revenue outside of simply well ticket deflection and cost savings.

SPEAKER_01 2:01

Yeah, you speak you spoke there quite a bit around ROI, which like you said, it's it's always a big topic whenever we speak about AI and customer service. I was just wondering, where do you think organizations perhaps are feeling the most pressure when they try to prove that financial return of customer service AI?

SPEAKER_00 2:18

Yeah. So the main the main uh area where they're they're feeling the pressure is that well, people's expectations have really increased in terms of what the AI should be able to do for them, uh, which means that you're no longer asking your customer to basically just go to customer support for customer support questions, go to sales for sales questions, go to go to um media inquiries for media inquiries questions. Customers now expect to have one interface uh that is an entry point to the rest of the organization. And so when you talk about, well, I want to prove the ROI, well, the obvious way is of course, how many tickets have I deflected? That even itself is an imperfect metric, and we can talk about that later. Uh but when there is impact across the entire organization, across the revenue and the cost uh of a of an enterprise, uh you start having to figure out, well, okay, uh maybe I uh deflected uh X amount of tickets, but then how do I how do I explain that uh through those conversations we've had an increase in customer satisfaction, which therefore means you know our brand has our brand value has increased, which also probably means that customers are less likely to churn and things like this. And that's even leaving alone all of the integrations with things like commerce and marketing and so on.

SPEAKER_01 3:41

Yeah, thanks for that, John. Some really interesting stuff in there. I guess pulling things back to the to the RevTech Stack situation, you know, when you guys at Full Thought start looking at solutions to connect CX with RevTech Stacks, what perhaps shapes your decision-making process?

SPEAKER_00 3:59

Thank you for the question. Uh so what we have been noticing uh across the market is that user expectations are have have increased in terms of what they expect out of AI tools, uh what they expect out of their interactions with a company. And so previously used to get routed to the sales uh interface whenever you need it to talk to sales, customer support whenever you need when you had a question, marketing, etc. etc. Uh what AI has enabled and the expectation that it's created is that there is now a single entry point for any use case that you have with respect to a given company. So it was almost a we had to do it basically in order to continue serving our customers best. Uh we had to expand into use cases that go beyond um typical traditional customer support use cases. So one example of that would be uh commerce. So you're on a website, let's say you're chatting with support and you are it's maybe a company that is uh that sells you know uh electronic devices and you're trying to troubleshoot a problem, but then it turns out you know the device that you currently have uh is very outdated. You should be able to, and we enable you to uh inside of the chat bot, inside of the that experience, uh, to figure out what device you should purchase instead. And it doesn't just stop there. Uh we also uh we also have so this was announced actually last week. Um we released something called AI Studio. AI Studio allows uh our users, so the the customer support admins, uh, to have very conversational experiences uh and analysis of the state of their of their customer support uh CX automations. Uh there is a headless mode of that, which then allows you to bring in kind of all the insights and the intelligence of Forethought into whatever tool they're using, uh things like Cloud Cowork, ChatGPT, etc., um, and then connect your other systems to it, and then have that kind of um analysis, deep analysis of correlating all customer support issues uh with you know other metrics that you care about. And then finally, we also uh, and this is extremely important in terms of observability, we allow our users to set up uh custom events essentially that let's say, again, an example of that uh e-commerce use case. If you well, you want to know if someone added something to cart, you want to know what that was, maybe the user's location, so on and so forth, you can easily track that uh with with your custom events and then build out all the dashboards that you want. And so, in doing all of those things, uh we are basically addressing the reality that users now expect an end-to-end experience and no longer a basically uh we're no longer asking users to try to figure out your org structure for you. Instead, this is the portal, this is how you interact with our enterprise, and it looks like uh an AI agent.

SPEAKER_01 7:10

Yeah, now some really interesting stuff in there. You mentioned observability, Anton. I find this interesting. It's maybe we've gone off on a slight tangent here, but it's obviously it's an area that a lot of vendors seem to be investing in almost all of a sudden. I was wondering why do you think there's kind of this uh emphasis on observability right now?

SPEAKER_00 7:29

Yeah, uh I think it's almost so before when you when it was humans doing all the work, the observability looked like uh one-on-ones with your direct reports, uh recordings, and so on. Now, with AI, many people think of it as a black box. Uh, and so they are worried that, well, I'm spending a lot of tokens to answer a given question. Is that money going to something that actually is going to uh impact my business outcomes uh positively? So when we're our focus on observability is basically to preempt this question, uh and we up we offer observability at every level of our product. And so, even you know, we talked about deflection uh earlier. So, what is deflection? A deflection is a customer try to get in contact with your company. Doesn't matter what happened, they didn't end up talking to a human. But that's a really important thing. It actually matters a lot what happened. And so measuring deflection in and of itself is not the most aligned way to think about to think about value generated by AI. And this is why at Forethought we are going towards and we have gone towards uh outcome-based pricing. And the way that we determine the outcome beyond just a deflection is we actually evaluate the conversation, we look at how it went, and if the conversation actually uh was, or if the AI agent actually addressed um the customer's issue, if the if the information provided was correct, if the steps taken were relevant, that's what we consider a resolution, and that's when we get paid. And we, whenever we determine that something was a resolution, we actually, in terms of observability, we show that to the admin. So the admin can audit every single conversation and they can see, well, okay, uh the AI thinks that this was a resolution, let me see why. The AI thinks that this was not a resolution, let me see why.

SPEAKER_01 9:27

Yeah, no, I think that's really really interesting stuff. Like I said, I do apologize. We went slightly off topic there, but I think uh yeah, it's a really interesting, really interesting trend in the space right now. I guess jumping back onto the onto our main topic here. Um obviously at Forthought, you you guys have documented a really impressive number. It's one billion dollars in ROI for customers. What evidence or maybe early signals would maybe convince a cautious buyer that these returns are actually achievable and measurable?

SPEAKER_00 9:59

Yeah. Uh so we're very proud of that. Uh but if I'm being honest, this is an outdated number. Uh so but uh to your question, um I was I was mentioning prior about how for any given individual conversation you can easily see why we thought that this was a resolution or not. So that's already sort of in the way of evidence that, well, before we even talk about ROI, the R part of ROI, here's how you can see the return, is because yes, it did get, it was a real conversation that got handled by an AI agent. Now, in terms of the kind of the end-to-end RI, well, customer support NCX is one of the uh the domains where it's relatively easy and straightforward to see what kind of your worst case RRI is. So if you don't take into account any of the extra stuff, your ROI is essentially well, my agent would have spent, let's say, 20 minutes. Um, my agent would have spent, let's say, 20 minutes uh handling this ticket, one hour, whatever it is. Here's what the ticket costs in terms of dollars or in terms of hours. So that ticket didn't happen, but it was handled correctly. So again, that's uh that's a really important point that I'd like to make again, which is if I wanted to if I wanted to only maximize deflection, I could do it very simply by just refusing to answer questions, right? And not providing an escalation path. So that's obviously not what we do. Uh and so when we talk about um uh the base case RI, it is essentially uh dollars and cents, how many tickets that would have gone to a human uh don't end up making it to the human. But there's also all the extra stuff, which is great, which is well, now you have control over a unified voice of your company. Uh it is the AI agent that you configure, you put um effort and attention into configuring it correctly, the sources that it consults, the personality that it has, and so on. So then you get an increase in a CSAT, customer satisfaction. Uh, you have opportunities to kind of assist the customer beyond what they initially thought they wanted to. Back to my example about troubleshooting electronic equipment. So all of that slightly less tangible, but still it it's it's a nice extra uh compared to what we already can measure very easily.

SPEAKER_01 12:24

Yeah, yeah. I think it's it's just a nicer way to look at the technology, looking at its wider use rather than, like you said, just focusing on ticket deflection, which can be not the clearest metric as you've outlined. Yeah, some really interesting stuff. Uh I guess Anton, next I wanted to ask you, and this is what I like to do with kind of a lot of people like yourself and I speak to who are working at the co-face of this technology is what maybe shifts or trends are you noticing in the market right now, you know, and particularly with AI becoming this primary service layer. And obviously, as you touched on, your your recent news with Zendesk?

SPEAKER_00 13:00

Uh so the trend that I'm noticing in the market right now is that, well, for the last 12 months, I think there has been euphoria at how amazing those tools are. And that is absolutely true on an individual basis. So when you use ChatGPT, when I use Cloud or whatever it is, uh the capabilities of the tool is amazing. Uh, and so that has prompted a massive wave of investment into uh the enterprise. And now what we're seeing is the executives are asking, well, was that money justified? And that is where thinking of AI of your AI strategy at the organizational level requires a different set of thinking as it does compared to when you're just doing it, we're just using it as a personal user. Uh and that has been a fascinating trend for us to look at. Uh, and it aligns very well with our core capabilities of we're not just trying to sell you a tool, uh, we are trying to accelerate and improve your business. So we take this consultative approach, we try to figure out well, how what is the best way? And we try to set you up and give you the tools to set you up automatically with the best uh possible uh AI automations. This is, for instance, our Discover product. Our Discover product will automatically figure out what are the topics that you're receiving, uh, what are the topics that your customers are asking you about, and create AI agents to handle those topics. It'll find gaps in your knowledge source, uh knowledge base, and it will help you fill those gaps based on how your agents have handled uh similar questions in the past. And then it will constantly self-improve by learning from every interaction. So we announced the self-improving agents uh at Fore Thought and Zendesk. And that has been a massive effort uh from the team to basically figure out the best way to create AI agents that get better every time it they encounter a situation where maybe they underperformed. So that that that has been kind of the trend, the trend that uh I think is the most exciting.

SPEAKER_01 15:08

Uh, and yeah, yeah, absolutely. I like what you said there about the consultative approach because, like you said, you know, there's uh the AI space isn't uh isn't short of vendors right now. Like with the and I'm gonna like that phrase, euphoria around the tech. Obviously, there's uh there's plenty of options, but like you said, having uh the right option for the right problem you're trying to solve rather than just AI for AI's sake is is I think always one of the key key lessons around this. So I guess just on my final question here, and again, this is something I like to ask a lot of our a lot of our interviewees. You know, if you could offer one piece of advice to our audience this week when it comes to turning AI into you know a proper revenue driver as we've discussed, what would it be and why?

SPEAKER_00 15:55

The main piece of advice I think I would give is that it treat AI and AI agents as infinitely patient and infinitely um energetic uh employees that need a good amount of onboarding. And what I mean by that is there the we we've all seen it in our experiences with the AI models that they're extremely capable. But when it comes to precise domains and requirements of you know uh complete accuracy and truth and effectiveness, those AI agents need to be guided. So don't shy away from the investment of time and effort. That is, I think, how roles are transforming uh in many areas of the economy, which is instead of being the person answering the question, you are training and you are guiding and you're managing the automation that could now answer the question and lean into that effort uh because it is definitely worthwhile and it will transform your organization.

SPEAKER_01 17:02

Perfect. I think that's uh yeah, a great place to end. Thanks. Thanks, Anton. Really, really great chat there. Like you've said throughout, I think AI ROI has been a bit of a pain point in the customer service space for a long time. So I think it's really helpful speaking to people like yourself who have that expertise who can get under the bonnet of it a little bit and really kind of unpack the do's and the don'ts, kind of the lessons that people can take from this. So uh yeah, thank you very much for your time.

SPEAKER_00 17:27

Thank you so much, Reese.

SPEAKER_01 17:30

No problem at all. I would also like to just quickly thank our viewers for tuning in. If you enjoyed this, and I'm sure you did, please remember to like and subscribe to the channel and head on over to cxthoday.com for more stories like this. Until next time, thanks for watching.