With everyone and their mother jumping the AI bandwagon, let’s do a sanity check.
Is generative AI in 2025 the internet of 1998, one year before the dot.com bubble?
Or are companies seeing some real returns on investments (ROI) from their generative AI and agentic AI initiatives? What is the value-add of generative AI exactly?
I asked Gemini Deep Research to do a deep dive, and it was able to pull ROI figures on Generative AI initiatives for Q1-Q2 2025 from sources like EY, Deloitte, Gartner, Accenture, McKinsey, BCG & co—companies that make it their business to know.
Here is the full Gemini-generated report in case you want to read it:
Some things that stood out to me:
Financial services—mainly banking and insurance—are leading the pack when it comes to generating ROI, with a 4.2x return in investments on any dollar spent on AI.
This is in no doubt down in some part due to an industry bias on tracking ROI—I’ve also been involved in quite a few successful financial services AI implementations myself, so my professional experience backs the internet data in this case.
The average ROI across industries on generative AI projects is none too shabby either:
A 3.7x ROI on projects that tend to run 2-6 months is pretty good!
That’s a lot better than most things listed on Wall Street these days..
The fact that most generative & agentic AI projects either meet (43%) or exceed (31%) expectations was surprising to me—I was expecting a lower success rate based on what I’ve heard from peers, somewhere around 50%.
This relatively high success rate could of course be down to a bias towards the kind of marketing materials shared online that Gemini has access to.
People somehow prefer to share their wins rather than their losses with the world.
On the other hand, we’re also 3 years into this thing—it’s no longer 2022, and most large organisations already had established AI & data practices that are able to steer these kind of initiatives based on years of experience.
The success and impact reported are primarily for LSE, with SMB lagging behind somewhat both in terms of AI adoption rates and successful implementations.
Even with those established AI & data practices, two-thirds (66%) of companies still reported difficulty of establishing a positive return on investment on their generative AI initiatives.
Implementing generative AI projects in your org is definitely not a walk in the park.
You need to know what you’re doing, but then the rewards are also there:
The stat I expect to rise most both this year and in years to come is the direct impact of (generative, agentic and regular) AI initiatives on company revenue—19% from the sources cited by Gemini, and I’ve also read 17% in other reports.
Those percentages seem roughly precise (which happens a lot when you’re working with AI agents) based on my experience and given how early into this thing we are.
That’s all fine and dandy of course, but what are some steps you can take to end up on the right side of the balance sheet with your AI initiative(s)?
Four success factors found across all ROI-positive generative AI projects
Make sure your AI project checks the following four boxes:
The project has measurable success criteria: establish SMART objectives (Specific, Measurable, Achievable, Relevant, Time-bound) for your generative AI and agentic AI implementations. Define expected success and failure rates for the AI solution, such as "reduce customer service response time for Tier 1 queries by 35% while maintaining 95% accuracy", rather than vague goals like "improve customer service efficiency." Translate those metrics to money to ensure you are able to do a proper cost-benefit analysis of the AI project.
Ensure the right data foundations are in place: run a comprehensive evaluation of your existing data, data quality, data accessibility, and relevance, identifying any gaps that would prevent your AI agent from functioning effectively in the real world. Gen AI and AI agents run on data just like any other AI app—so make sure the data is present and available in the right formats before starting off.
Focus on mapped-out, functioning processes: a key mistake I’ve seen made too often is that people jump headfirst into an AI project to innovate their way out of a broken process. While reducing (human) error rates can be a key success metric for an AI project, a major process overhaul will put your AI initiative on thin ice because now you’re not just rolling out an AI solution—you’re doing change management. And what is worse, you’re relying on technology to make something work that people couldn’t. Which is a dangerous bet to make any day of the week.
Your project is staffed with people that know what they are doing: an often overlooked but evidently crucial element in any successful AI project is that the people on the project have a clear roadmap and relevant experience. If you don’t have the experience in-house, consider hiring outside help or up-skilling your team. Even with outside help, up-skilling your existing team should be part and parcel of your project roadmap.
The AI implementation gap is real, but it's also entirely avoidable. By focusing on clear success metrics, solid data foundations, optimized processes, and building the right skills in your team you'll position your generative and agentic AI projects for the win. Rather than another case of shiny new object syndrome, you'll be methodically building systems that deliver measurable business value.
This Thursday I’ll be speaking in Amsterdam at Digital Marketing Live (in Dutch) about how to get your marketing team ready for generative & agentic AI initiatives. Can’t make it or don’t speak Dutch? I also offer Team AI Readiness workshops through Lodestone Digital:
Last week in AI
Mistral AI launched its new Agents API, enabling developers to build AI agents capable of interacting with external tools and databases via function calling. This allows for the creation of more autonomous and sophisticated AI applications that can execute complex tasks.
DeepSeek released DeepSeek-R1-0528, a new open-source model now available on Hugging Face, offering another powerful option for developing custom AI solutions. This model provides an accessible alternative for businesses and freelancers looking to build AI applications without proprietary restrictions.
Google officially launched Stitch following its acquisition of Gallileo AI. Stitch is an AI-powered tool that converts Figma designs into production-ready React code. This platform is set to significantly accelerate web development workflows for businesses and freelancers by bridging the gap between design and code. Check out this old Youtube video of mine to see what you can do with Galilleo AI.
OpenAI's Operator automation tool now supports their o3 model, enhancing its capabilities for users. This upgrade provides improved performance, efficiency, and automation possibilities within the Operator environment. Businesses using Operator can now leverage this more advanced model for AI-assisted tasks.
Manus AI launched a new feature that generates structured and tailored slide decks from the research the AI agents perform. Users can easily edit the generated decks and then export or share them, streamlining presentation creation for various settings.