A few years ago, ChatGPT had the intelligence equivalent of a high school student. Today, it surpasses the knowledge of someone with a Ph.D. As these tools swiftly evolve, so do their use cases.
Former Engagement Manager James da Costa (London, 2015, 2018-21) sees this firsthand as a Partner at a16z. James develops and invests in startups with a focus on AI Software as a Service. Previously, he was a founder, having also started one of Africa’s first digital banks, Fingo.
But now, James believes, it’s never been easier to build a company. “The best startups that are using AI to solve business problems are building extremely fast,” he explains. “You can meet someone with an idea today, and tomorrow they have a prototype, and the day after that they’ve shipped to a customer. It’s truly energizing.”
Former Business Analyst Micah Hill-Smith (Sydney, 2020-22) has witnessed this firsthand as well. As co-founder of Artificial Analysis, which specializes in AI benchmarking, he’s been tracking the development and efficacy of these tools and the Large Language models at their foundation.
In a conversation led by Senior Partner Lareina Yee, James and Micah sat down with current BA Natasha Maniar—who, prior to her time at our Firm, developed multimodal AI applications—to share insights on exciting AI trends across industries and recollect on firm frameworks.
As the resident benchmarker, Micah, can you give a high-level explainer of what you’re talking about when you refer to AI models and how to differentiate them?
Micah: So, imagine a massive offline CSV file that you run all the matrix multiplication through to predict the next token—a token being a unit of information that AI understands alongside other units. That’s the model.
When you use ChatGPT, you’re implicitly picking the model it uses—like GPT-3. That model is predicting next tokens, but it’s being given context by the application. This context includes your chat history as well as some additional tools to search the web, generate images, and execute code.
As an AI benchmarking company, we conduct a lot of testing on these models and look at many different factors to determine how smart they are. We are in this incredible moment of competition between all these companies, such that we have a robust global top ten.
With so many developments, what gets you the most excited?
Micah: I remember the feeling that I had when I first tried OpenAI’s GPT-4 not that long ago, and how shocking that was—the speed at which everything is continued over the last couple of years has been pretty wild. And now again with GPT-5.
These models are getting smarter, cheaper, and faster. Software and hardware gains continue to improve year to year. These developments enable a huge range of new AI uses, and I strongly believe these applications will profoundly change what work looks like over the next (terrifyingly small) handful of years.
Natasha: One tool that could really change work is MCP—the Model Context Protocol. It connects different agents, data sources, and apps so they can talk to each other seamlessly. Instead of juggling siloed tools, you get an ecosystem where AI and software work together across your workflow. That interoperability is what makes it a game-changer. MCP protocols can reduce the barrier to integrations.
Another exciting advancement has been the increase in modalities—the media forms we can use to engage AI. Everyone is likely familiar with text to text; that’s where you would submit a written prompt and get a text response. But now we’re seeing abilities of inputting text, voice, image, and video, which then can produce simultaneous outputs of text, voice, image, and/or video leading to more real time multimodal context-aware and useful assistants
James: On that front, I’m also really excited about voice AI that can be used within a conversational flow. This would involve speech to text, text to speech, voice to voice. I’m eager to see how it’s creatively deployed to solve existing business problems.
Voice AI is at an exciting inflection point. Voice agents allow you to scale to peak demand, staff phones 24/7, communicate in your customer’s preferred language, and operate at a fraction of the cost.
I’m eager to see how it’s creatively deployed to solve existing business problems.
What do these developments look like in practice?
James: Take customer support in banking. Financial services make up 25 percent of total spend on all global contact centers, whether it be verifying consumers with voice-based biometric authentication or managing complex customer issues via call centers. Even before the launch of ChatGPT, many banks were experimenting with voice automation already. So, deploying voice AI in customer support is a natural extension of this.
The other interesting development here has been with pricing. Historically software for customer service has been priced per seat. But when AI can handle ticket resolution, the natural pricing metric becomes successful outcomes. If AI can handle a sizable proportion of customer support, companies will need far fewer human support agents, and therefore fewer software seats. This forces software companies to fundamentally rethink their pricing models to align with the outcome they deliver rather than the number of humans that access their software.
Natasha: That pricing piece will have an impact across industries such as in healthcare with value-based pricing models.
I’m really excited about how AI is changing consumer health. Wearable devices don’t just track your health data anymore—they’re starting to provide personalized interventions. With LLM-powered research agents, you can ask questions by voice or text and get real-time medical suggestions tailored to you. You can even take a photo of your food to get nutritional information or build a personalized meal plan. Over time, that helps people understand and manage their health in a much more proactive way. AI is really enabling people to take ownership of their health.
Recently an MIT Media Lab report, “Your Brain on ChatGPT,” went viral, sparking concerns about how human thinking holds up in an AI era. This paper explored overreliance on technology and raised some interesting questions about how much agency we should have over the tools we use. I’m keen to see more extended user research on this so we can figure out how this tech will work within our world, rather than us working in the technology’s world. A lot of it will come down to using AI as a tool of role augmentation, not just automation.
Micah: I love that notion of augmentation. When I think back to my time at McKinsey, I had the privilege working on a couple of blue-sky strategy pieces for big banks. That big-picture strategy thinking was far more enjoyable than the grind of daily deliverables.
AI makes me excited for BAs of the future because they get to spend more time on those questions, while AI tools handle a lot of the manual, time-consuming work. The value of BAs and consultants in general was never the manual work of putting things into a slide; it’s the big ideas. AI can help reallocate time there.
Any other research we should keep an eye on to stay fluent on AI topics?
Natasha: One fruitful area of research that may advance intelligence of AI is simulations, as in: how can you simulate human behavior? How can you simulate agent-to-agent interactions? How can we push agent intelligence so that it can replicate human intelligence within reasoning models?
It’s also valuable to look into some of the model cards of the tools you use to understand some of the limitations of the current models.
James: The most important way to stay competitive when it comes to AI is to test the tools and use them regularly. Try them for software engineering or writing. Apply them in your daily workflows.
There’s a moment where you’ll go from just using these tools for projects sporadically to being someone who uses these tools day to day. Within a year, you’ll be so far ahead of the curve.
So, you three actually came together through Natasha, who was already riffing on AI topics knew James and Micah through separate ways. It’s a real testament to the power of network. How did you all initially connect?
Natasha: Micah and I met at an AI engineer conference in San Francisco. Then James and I met through friends of friends. When we first met, he asked me where I worked. I said, “McKinsey,” and he asked me “What’s that?” I spent five minutes trying to explain what we do, before he finally spared me and told me that he used to work in the London office. [Laughs].
James: We met through a friend, but the story goes further. Through your network, Natasha, I got reconnected with old colleagues from back in my London Office days. You pulled me back into those Firm communities, and that’s compounded into all sorts of wonderful conversations.
Natasha: I think the lesson is that one’s network really starts with friendship, I’d like to not think of it as a formal construct; I’ve learned the most just through conversations.
What’s your favorite way to play with AI?
James: I’m a big fan of AI-based coding tools, like Cursor, that I can use to generate code to build custom and personalized software. I encourage everyone to test out some of these coding tools.
Micah: I have to agree. Coding agents. As of this moment, here in 2025, they’re by far the single most magical thing I use in my day.
Natasha: Since coding agents have been claimed, I’ll say smart glasses. There’s a very cool one which just came out with a direct speech-to-speech model allowing people in real time to receive context aware answers to their questions based on the environment around them.
Finally, as you know, at the Firm, we love a good framework. What’s your favorite?
James: “Release the agenda!” You always need an agenda for a meeting, and then you need the flexibility to adjust as necessary.
Micah: Framework feedback. Honestly, in what I do now, I’m thinking a lot about how to offer the most helpful feedback. Culture is incredibly important, and the way it was framed at the Firm was massively informative in my career.
Natasha: The two-by-two matrix for use case prioritization, especially useful in product strategy
