An exploration of what product management becomes when AI participates at every step.

Even small teams end up with a stack of disconnected tools. A project tracker, a docs platform, a help desk, a CRM. Each carries a per-seat fee. Each is racing to bolt AI features onto an interface that was not designed for them. The result is five chatbots that do not know each other, sitting on top of five data models that cannot talk.

ProductIntel collapses the stack into a single self-hosted instance on the cloud provider you already use, plus an AI token budget you control directly. A starter configuration runs around $100 a month all in. An active-active redundant setup runs closer to $200. Teams that want the platform to build new features for them can add a flat-rate Claude Max or ChatGPT Pro subscription on top, since bursty agent execution runs cheaper on a subscription than on per-token billing. The cost shape works for a five-person team in a way enterprise SaaS does not.

Six months in, the more interesting question got bigger. The design problem was not “replace SaaS.” It was: what does the daily rhythm of building a product look like when an agent team is real?

Anti-platform architecture. Each company forks the codebase and owns their instance, instead of buying access to a multi-tenant SaaS. The cost economics of multi-tenant break down when AI inference is the largest line item in the system.

A unified knowledge graph. Stories, docs, cases, requests, APIs, learned heuristics, all the same artifact type, all addressable by typed edges, all queryable as a single graph. Federated data does not let AI reason across the surfaces a PM actually works on.

Agent workforce as a first-class primitive. Not a chat interface that calls tools. A roster of named agents with specialties, assigned to teams, supervised by other agents, with their cost, performance, and reasoning trails tracked the way you would track an employee.

The Today module is what came out of the bigger question. It replaces sprint cycles with daily cycles. Six steps: capture an idea in plain language, refine it into a structured spec, plan the work units, resolve open questions, execute as a PR, deliver. AI runs the steps in between. You stay the one making the decisions that matter.

A morning prompt becomes an afternoon PR, not a story sized for a sprint two weeks out. The vocabulary is deliberately different from agile. No story points, no velocity, no retros. Those words were built for human-only delivery cycles. The work has changed.

The Today workflow assumes someone inside the team captures the idea. In real product life, most work originates outside it. A customer reports a bug, asks for a small feature, points out a flow that does not behave the way they expected. Cases is where that signal lands. From there, the same cycle takes over: refine, build, deploy, deliver. The difference is that the customer stays in the loop the whole way, not waiting on a release-notes email a month later.

The honest scope is small, well-bounded fixes. The compression is not from doing more in less time. It is from removing the queues, ticket triage wait, sprint planning wait, release approval wait, customer notification wait, that today eat 80% of the elapsed time. Strip the queues and a customer-reported bug becomes a customer-verified production fix in an afternoon.

The mistake most AI products are making is forcing the customer to switch. Today’s workflow is real, but it is not where most teams are. ProductIntel ships in three modes: agent-native (the Today workflow), traditional (the classic Work module with sprints and a kanban board), and hybrid (both at once, picked per project or per story).

A team can adopt the platform as a lift-and-shift of how they already work, then push toward AI-native at their own pace. The dual-mode design is not a hedge. It is the actual shape of how companies adopt AI.

ProductIntel’s intelligence stack is built around the second pattern. The Discovery Agent watches your backlog and surfaces patterns the team has not named. The Strategy team subscribes to your industry (Reddit, Hacker News, RSS), filters for relevance against your product context, and writes daily briefs on what is shifting and what to do about it. The Behaviors module observes how your team operates and learns the heuristics no one wrote down.

Most AI features wait to be asked. These do not.

ProductIntel runs at productintel.io. Twenty-one agents organized into six teams. Eighty-five-plus database tables. The autonomous execution pipeline is wired: agents pick stories from the backlog, run on a Mac Mini worker via the Claude Code CLI, and open PRs. Most of the day-to-day shipping happens through the platform itself. I started on a $20 Claude Pro subscription and graduated to $100 Max once agent execution picked up. Flat-rate beats per-token API billing for bursty dev work, and the math holds even at $200 if you push it harder.

Some pieces are still rough. The roadmap below is honest about what is done, what is in progress, and what is next.

Building this has been the most direct way to develop fluency in the tools that are reshaping the PM role. Context engineering, agent orchestration, token economics, the trust UI for non-technical stakeholders. These are not theoretical questions when you are shipping a product that depends on getting them right.

I write about what I am learning at mikeholloway.dev.