The AI Diaries: week 3
It’s clear that AI is changing the world, but just how much of a difference can it really make to enterprises? Is now the right time to sign on – and how can it be best employed?
In our AI diaries series, we’ve been exploring the opportunities AI affords – and the unique challenges it presents. We’ve asked two of the companies we back to track their usage of different AI tools and report back.
Oliver Crowe is the technical product manager at Octopus Ventures-backed Flock, an insurtech specialising in whole-fleet vehicle insurance for commercial operators. With thousands of vehicles on the books, Flock has long struggled to leverage the vast quantities of data generated across an extensive tech stack. Oliver hoped that Anthropic AI’s Claude, with a model context protocol (MCP) which allows the integration of other tools, would allow the team to break down the silos and centralise a historically fragmented pool of information.
Patrick Van Deven is CEO at another Octopus Ventures-backed business: VaultSpeed. While his team found a customised ChatGPT assistant useful, he reported some issues with hallucination. The speed and effectiveness of other tools, used for vibe coding (writing software through prompts, without code) raised some more philosophical questions. How does the frontier between product management and engineering look when an (ostensibly) functional prototype can be built in a shadow of the time it would have taken before?
If you haven’t read them yet, catch up with Oliver and Patrick’s AI journeys in Week One and Week Two – or read on to learn how they did in Week Three.
Oliver Crowe, Technical Product Manager, Flock
Week three has been excellent. The different tools we’ve integrated into Claude using the MCP are pulling in data from across the business and massively supplementing how we work. When it comes to the biggest positives, efficiency is right at the top.
In our operations, many of our tasks used to be particularly manual, with lots of different steps and systems involved. We’ve tried some of the workflow and agent tools like n8n and Zapier with the MCP, with a view to setting up agent-enabled workflows.
Now, with these AI-driven workflows, we can do what was taking about 20-30 minutes, down to just minutes on a daily basis. For example, we’ve leveraged our telematics data and our policy-holders’ connectivity to create a new system that automatically alerts our internal team, opens a ticket in our customer relationship management system (CRM) and notifies our hardware supplier to reach out, in just a few steps. It’s saved us at least half-an-hour, every day.
The other benefit is it’s a huge enabler. Of course, the tech teams have enjoyed some of the benefits, but different teams across the business are seeing gains as well, and confidence in the tools is growing. You can really see this on a regular basis, whether it’s daily interactions in the office or our AI breakfast where team members with no technical background showcase workflows they’ve set up themselves.
Instead of the default ‘Ask the tech team how to do this,’ we’re finding less technical colleagues setting up their own workflows; the cultural shift is really powerful.
As to whether it’s worth converting a trial into a paid subscription – if the trial adds value I can say these tools pay for themselves. The costs are tiny in comparison to the efficiency gains and the insights we can get, from different data sources to customers.
But the key lesson, again, much like last week’s, is focus. The number of opportunities on offer can be very overwhelming which makes clarity on the problem you’re setting out to solve super important.
Moving forward, we’ll be expanding our AI solutions into additional workflows. One of the things that we’ve understood and validated on some calls with customers is that it’s important not to sell it as AI to customers.
As a marketing tool, it has limited effect. Instead, we focus on the problem being solved. With telemetry insights, for example, the real value is what the chart tells a fleet manager about risk or safety. Not the fact that it was AI-generated or the data points that contributed to it.
The other thing is that the MCP and a lot of these AI tools in general work best when the problem is really well defined. For workflows you need clear steps, for analytics you need precise questions and definitions of data points so that results are consistently strong.
Overall, the experience has been very positive. It’s only been a few weeks working with some of these tools, but there’s a huge opportunity in the future and I think it’s only going to get better over time. I’m excited to see how it goes.
Patrick Van Deven, CEO, VaultSpeed
Last week, I highlighted some of the problems I’d been seeing. Features being hallucinated, sales being lazy, product management vibe coding new releases that cannot be put on the market.
The answer to all of this, I’ve realised, is that it demands the development of a new type of AI-specific management processes from leadership, so that we don’t run ahead of ourselves.
It’s good to ride the wave of AI, but the way we do it is we work as a team and we keep one another in check. If somebody makes a huge step forward, a leap in innovation, with some AI features, we pause, reflect and challenge ourselves: how do we bring this to the market? How do we message it? How do we equip it with the non-product items that go along with it?
This new class of problems introduced by the speed of AI is answered by more teamwork at the leadership table. It’s faster and easier than it has ever been to produce good software. The question is, what do we want to do with it and who runs with it?
One consequence is that responsibilities are changing in a very fluid way. I’m not sure I have all the answers – in fact, for sure I don’t. But we’re exploring how this whole thing changes our processes.
A concrete example can be found in product marketing. With a custom GPT you can produce a new piece of product marketing stuff every day. But do we want that? And if we do, what do we do with it, and who checks it? It’s the same story with product management.
In the old world, product managers would write user stories, before turning to Figma to design what the user experience would be. Now, they just vibe code the next iteration of the product and hand over the vibe-coded app to engineering to build it for real. It’s blurring the limits of everyone’s functions, and because the limits are blurring it calls for even more teamwork and situational leadership at the top.
We’re also finding that not everyone is moving at the same speed. This means working to keep functional experts who may take a little more time to adopt AI – or bring a greater degree of scepticism – inside the company. The opportunities are huge, but the interplay of these unanticipated consequences means it’s us as leaders who have to ensure the solutions are properly leveraged – and everyone gets brought along for the ride.
