Is This Just Another Silicon Valley Promise?

It’s that time of the year in India. Over 49 days, a small religious town in northern India hosts the world’s largest gathering of people - about 450 million at last count - the Kumbh Mela. Estimates say that about 250 people, including several children, are separated from their families and lost in the crowd.

Why am I bringing that up now?

Because this year, AI (and facial recognition) is brought to help reunite these families. If that isn’t a use case for scaling AI effectively, I don’t know what is!

Despite the exciting adoption of AI, challenges are plenty - even just technological ones.

Amazon takes its time with GenAI on Alexa

Since the launch of ChatGPT, Amazon has been looking to make Alexa into an AI agent or a personalized concierge that can do a lot more than it currently does. Why don’t we have it yet?

According to Amazon, the technology isn’t really there yet. Hallucinations, latency, reliability, and scale are genuine problems that the tech industry is grappling with. This is especially true with general-purpose applications like Alexa, which have to do everything and then some!

[More on Alexa’s ‘brain’ on FT]

Others are more positive about the ‘future is now’ story

Google Cloud’s AI Business Trends report for 2025 takes a more positive outlook toward AI in the mainstream. What caught my attention was the prediction that multi-agent systems will take off in a big way this year.

If you haven’t heard, multi-agent systems are simply several independent AI agents that collaborate to complete a workflow. In a multi-agent sales support system, you might have one agent doing research, another simulating pricing models, yet another perusing RFPs, and a final one coordinating all this for writing a proposal.

What this does is create specialists within the org—data agents, creative agents, code agents, etc.—to do one task well, improving model efficiency and performance.

[The entire Google report can be found here]

CIOs are confused, naturally

70% of CIOs told IDC that their failure rate for custom-built AI app projects is 90%! That’s right. 9 out of every 10 projects fail, CIOs lament.

On the other hand, those who have successfully implemented GenAI see significant gains. BCG finds that organizations adopting AI claim 1.5x revenue growth over those who don’t. One of our clients reduced the cost of data indexing from $18 to a meager ¢60 with GenAI.

So, what gives?

In our experience, successful AI implementation requires a combination of strategic thinking and technological expertise. Here’s what we’ve learned.

Start small; build fast

Integrating an LLM into every process and expecting all teams to be productive is easy. It’s also easy to fail that way. Choose small use cases and smaller models to start with. Rapidly stack more use cases, like you would in a modular piece of software.

Fine-tune, rinse, repeat

No model out there is perfectly designed for your unique needs. In fact, only your own proprietary data can give you that X Factor. So, choose the best-fit model and fine-tune it continuously.

Personalize and evangelize

When you choose the agentic or the copilot approach, personalize it to the needs of each department/role, if not team member. Like you’d hire an assistant with prior knowledge/experience in the area, fine-tune the agent with the right data.

Then, create a sense of positive enthusiasm among your teams for using AI. Show them how it’s done and how it can make their lives easier.

If you need any help along the way, just hit reply, and we’ll bring the power of our AI team right to your meeting!

Best,
Anshuman Pandey

P.S. Llama 3.3 70B is now available on Tune Chat