Harriet

Yesterday we had David Buxton and Svitlana Kurochkina from Harriet in for the afternoon. We work closely together but normally across several time zones and an ocean. A few hours of in-person conversation covered more ground than I expected (though that often seems to be the case!). The texture of the conversation is simply more nuanced and faster-paced when we in the same room. Things that would have taken three Slack threads to sort out got resolved in five minutes. In-person collaboration still matters in a remote-first world. Yesterday was a useful reminder that the connective tissue of a working relationship can form faster when you're in the same room. Ryan Irwin Chaz Woodstock

In which phase is your business right now? ⬇️

Here's a breakdown by David of the practical challenges IT leaders are actually running into and the solutions that are finally getting entire organizations up and running on AI.

It's absolutely crazy the pace at which AI software has to be unbuilt at the moment because of changes in the underlying model capabilities. Every few weeks, it feels like whole files can be deleted where we previously built harnesses to improve outcomes. Now, the raw model capabilities are simply better than what we can achieve with those harnesses. I worry that medium to large-sized companies are shooting themselves in the foot by doing massive amounts of internal development around AI tooling. They are going to end up with white elephant projects that don't connect well to the changing state of the art. They will be left stuck with stuff that performs at 2024 levels when the rest of the industry has moved on, or they are just going to have to pay a huge maintenance tax on an ongoing basis.

A very valuable workshop today with one of our newest additions to the client list. The question high on agenda today is HOW do you ensure you enable your users with skills and data access without jeopardising privacy and data security. It is an important one! I wonder if anyone in my network thought about this from AI skills, tech stack/system connectivity and personnel's data and has solutions in mind on a org wide level. We solved this partially with Harriet AI Endpoint, but I am sure there are still areas to consider!

A great question is when companies decide these second-tier models are good enough and stop paying top dollar for the frontier. Anecdotally, cost-sensitive teams are already leaning hard into things like Cursor's Composer models. One wrinkle: for interactive use, cheaper usually means faster, which means better UX. That matters less as models get capable enough to run autonomously over many cycles without a human in the loop. The prevailing mindset is to push spend up people have drunk the Jensen Kool-Aid that developers should be burning $250k a year. Reining costs in is hard right now. The mechanisms are poor, and most teams don't have the telemetry to tell whether the marginal capability of the top model is actually worth it over something slightly less capable. It's a tough spot for a CEO: tell everyone to get AI-pilled, and in the same breath ask them to keep costs under control. Luckily, they have Harriet.

Out of the back to back meetings with clients and prospects conversations this week, I learnt that the gap is getting larger between the AI enabled and AI feared group of employees across the workforce. How can you help bridge the gap for your people, without forcing them, without making them feel uncomfortable? Your boss says AI is the future but no one shows you how to use it at work, you talk about agents all day but you have no clue how to connect Claude, ChatGPT... you name your AI tool... to your existing systems, to your legacy systems and even if you can, you need an engineer's time to set it up. One of my clients landed a very good point: "The pressure to use AI is high, the motivation to provision and understand employee use cases, overlooked!" What we have done is try to understand, by team or persona, what skills they will utilise the most, and with least friction, and enable those. So here's my question for you: if you don't have a tool that bridges that gap automatically, what's your plan to bring your people with you? That's what we're doing at Harriet! DM me to understand how we're tackling this problem.

Two AI adoption patterns I've seen this week: Company A is running Claude Enterprise, ChatGPT Enterprise, building MCP infrastructure, provisioning via Skills, hedging against the model wars. Expensive, but probably rational if you think the landscape stays volatile. Company B has an IT person who has quietly become the internal AI team. MCPs on Cloudflare, a single playbook pushed to 50 staff. Nobody fully understands how to maintain the MCPs, but it works. For now. Both are letting a hundred flowers bloom 🌼. The question nobody's asked yet is who waters them when the person who planted them leaves. 🥀 At Harriet we are building the infrastructure that keeps the flowers alive. DM me if this resonates.