33) AI Command Center

Behind-the-scenes building Vambrace AI, a company on a mission to figure out its mission. Please pardon the stream-of-consciousness style. Subscribe to follow along or visit the site here:

(typos are to make sure you’re paying attention)

Introductory Remarks

Dear Vambracers —

In last week’s post, Launch, I reflected on the recent launch of the production website and environment associated with our first engagement. So far everything has been mostly operational, and there haven’t been any major issues. There’s still a lot to work out, but now it feels like we can attack things from a position of strength. And that brings us to today’s topic: the AI Command Center.

AI Command Center

Background

Following the successful launch of a revamped operating platform for my first client—which included taking on hosting responsibilities for the entire backed infrastructure—I’ve developed more intuition around how legacy software operates. This might be obvious to many people, but software really is just massive spreadsheets stitched together that talk and interrelate in explicitly-defined ways that users can interact with and manipulate in explicitly-defined ways via the frontend. I know there are new database paradigms popping up, but for now let’s just move forward with the basic assumption that relational databases are pretty good.

Intuition

Now, as I was building new customer-facing components of the software platform, I had to understand the backend architecture to make sure that the frontend was speaking to the backend appropriately. But I also ended up building custom commands with Claude Code that directly manipulated the backend—pursuant to highly strict criteria—to help with testing. This was very efficient and fun in the moment. But now that we launched and I have the luxury of time for some more reflection, I’m realizing that we can apply similar techniques to the production software to help management get insight into the current state of the business and, eventually, directly manipulate the backend.

Phase 1: Read-only

So that gets us to the main idea here, which is to build an AI-native command center that (initially) has read-only access to the company’s entire database (with some restrictions set for sensitive information), and can provide real-time responses and recommendations to management. In the near-term, the value would be immediate insight into how things are going and direct responses to highly specific questions. This (hopefully) would help save time and improve operations, etc.

Phase 2: Recommendation-based

In the long-run, you could collect intent information and keep track of behavior such that you can start to model intuition around why certain questions are being asked or why management is interested in certain things. From there, you could actually move towards a recommendation-based AI Command Center where, each morning, management receives an Executive Digest complete with specific one-click recommended actions that they can take based on the current state of the business as told through the data.

These could be, for example: “schedule mow route for [xxxxx] zip code”; “increase prices for [yyyyy] customers”; “initiate billing and payment collection for yesterday’s services,” etc. Management could theoretically click through these recommendations each morning and either accept, modify, or decline each recommendation—and prompt the AI to behave differently where relevant. The value here would be time saved and more indicators of improved economics (cash conversion cycle compression, more revenue through better pricing, etc.).

Phase 3: Autonomous action

In the long-long-run, you could have the AI actually proactively decide and execute certain sorts of tasks around route scheduling, payment collection, price increases, crew composition, etc. Assuming you permit the system access to communications technologies as well (i.e., SendGrid) then it could even proactively send out marketing material to recover abandoned customers, offer discounts to certain disaffected customers, and directly respond to customer concerns or issues that arise. If it has access to all the data in the system, then really it should be able to provide more accurate and effective answers to customers more quickly than management could.

So this I think is where we start to get to the autonomous back-office, where the physical action in the real world (i.e., mowing a lawn) still requires humans and physical machines, but everything else can eventually start to run on autopilot (again, pursuant to some boundaries).

What really excites me about this future is that the AI could theoretically come up with novel ways to more effectively generate and capture economic value from customers. When the DeepMind team created an AI system that won the first game of Go (the most statistically complicated board game that exists), the Go community was amazed to find that the AI was doing unorthodox and seemingly-stupid things that have since become orthodoxy to Go strategy. And so who knows what seemingly-stupid business strategies are out there just waiting to be discovered?

The “Insight”

I don’t know that this is actually all that insightful, but if these types of businesses—and most businesses at that—really are all built on these massive spreadsheets, then it’s all just data, and once you help AI understand the first derivative of the data-state, meaning the reason that a change in one variable occurs—then you can start to let it proactively intervene and “run” the company autonomously, by just changing the data as it sees fit. It would “understand” what it means to need to schedule routes for [x] customers, or to raise prices by [y]% in response to [z]% increase in average service times and [a]% increase in wages, etc.

It could eventually even read-in information on competitor pricing, competitor behaviors, etc., and then even somehow (I’m sure) ingest information on prospective customers, directly create custom ads for certain customer cohorts, use the historical data to cluster customers by purchase patterns and then find qualifying indicators for those clusters and then target them on socials, etc.

The possibilities really are endless when you realize that all the data we’ve put on the internet is effectively understandable and manipulable by AI. And so then if businesses run on data, then AI can run businesses. It’s just about bridging the intuition gap around the first derivative of data-state changes to get there—which I think will come with time.

Legacy vertical SAAS

The reason for the subtitle, really, is that I personally believe that legacy vertical SAAS platforms probably won’t move quickly enough to embed these types of autonomous capabilities into their systems. I also think there could be some value compression there if the analog features that were built in a software 1.0 and 2.0 world, and also there would conceivably be less time spent in the platform probably.

Also I think it’s just criminal (not actually) how restrictive these legacy software solutions are when it comes to data interoperability and “data freedom” (for lack of a better term, although I’m sure there is one). It is such a pain to get data out of the system in any sort of sophisticated and useful way. Usually the only option is just download a massive CSV file, and then what good is that? You’re only getting point-in-time snapshots of the state of the business.

So that’s also why I think there increasingly is value for companies that have real-time access to their own data and are run on their own software systems. That data is invaluable in the modern AI-first world where each and every transaction, communication, route, etc., can help inform an autonomous future.

Looking Forward

I feel more excited and optimistic about the future than ever—and I’m also grateful to be doing what I’m doing and learning what I’m learning. There feels like so much opportunity right now and I hope to be part of the cool stuff that will be built in the coming years.

I might do a more traditional annual-planning type of thing next week—we shall see. But in the meantime, I hope you had a wonderful new year and that you have a great week!

Sincerely,

Luke