Scaling Product Management for Agentic Delivery

Your developers are six weeks ahead of your roadmap. The PR is open, the IDP cut a working microservice in an afternoon, and the spec it was built against is two iterations stale. The agent did its job. Product didn't get the chance to.

‍

That's the new shape of the bottleneck in software delivery, and it lands harder in regulated financial businesses than anywhere else. A consumer SaaS team that ships the wrong thing rolls it back and moves on. A payments processor, a custody bank, a digital broker, or an embedded-finance startup doing the same thing runs the rollback into change control, crosses a compliance boundary, and gives the audit log a story to tell. The shape of the friction shifts between an incumbent bank and a Series C fintech, but both feel it.

The math behind the bottleneck

Marty Cagan's 1:6-10 PM-to-engineer ratio was an accurate read of the world in 2007. Coding was the constraint. One PM could keep eight engineers fed with well-defined work and still have time to talk to customers. Almost twenty years of that assumption shaped how product orgs were sized, what PMs were measured on, and which rituals filled their week.

‍

The inputs have changed. GitHub's controlled study of Copilot users showed a 55% improvement in completion time on a typical task. Agentic systems push that gain further by parallelizing requirement analysis, scaffolding, test generation, and refactor. PM throughput has improved too, but mostly on document production, not on the parts of the job that take real time: customer research, hypothesis design, scope decisions, stakeholder alignment.

‍

When the ratio gets pulled toward 1:3 or 1:4, the work doesn't just intensify. It changes character. A PM is no longer mostly translating intent into tickets. They're feeding decisions into a system that will act on them faster than they can correct.

What breaks first

The 2024 DORA report flagged something the industry hasn't fully reckoned with. AI adoption raised individual productivity, but throughput went down by an estimated 1.5% and delivery stability by 7.2%. The likely culprits are bigger batch sizes, weaker review discipline, and unstable upstream priorities. AI exposes whichever part of your operating model was already brittle.

‍

In platform engineering, the brittle part is usually scope. Engineers shipping through golden paths produce more, faster, against whatever spec they were given. If the spec was a one-line Jira title and a Slack screenshot, the platform ships a fully governed, audit-ready, compliant version of the wrong thing.

‍

For a regulated financial business, that's a much bigger problem than a sprint of rework. SOX evidence gets generated. PCI scope shifts. Customer data may cross a boundary that requires a control owner's sign-off. A fintech operating under a BaaS sponsor bank may trigger a partner notification it didn't intend. The platform protects you from the technical mistakes. It can't protect you from a mis-specified requirement, and the agent can't tell the difference.

The new operating model

The product orgs getting ahead of this aren't bigger. Their PMs spend their hours differently. Three things move.

‍

Discovery stops being a phase. The classic discovery-then-delivery cadence assumed delivery was the slow part and you could batch discovery up in front of it. With agents in the loop, delivery isn't slow anymore. Discovery has to run continuously, alongside delivery, feeding the next decision into the queue at the rate the platform consumes it. SVPG's writing on the product operating model covers some of this in general. The financial-sector implication, whether you're a bank, a payments processor, or a fintech, is that customer research, regulatory interpretation, and risk framing all need to become standing capabilities, not project phases.

‍

Specs become executable artifacts. A Jira ticket and a wireframe were enough when a person was reading them and asking clarifying questions. An agent doesn't ask clarifying questions; it makes plausible-sounding assumptions. The move is toward spec-driven development, where the input to the agent is a structured artifact with outcomes, scope boundaries, constraints, prior decisions, and verification criteria. Tools like GitHub Spec Kit make the artifact concrete. Writing one of those is real product work, and PMs need the room to do it.

‍

Guardrails move into the platform. Compliance controls, data classification, change traceability, segregation-of-duties checks, and blast-radius limits all live in the IDP, enforced as policy-as-code, not in a reviewer's calendar. The bottom layer of correctness becomes platform engineering's job, which frees PMs to spend attention on the top layer: are we building the right thing, for the right user, under the right constraints? Banking Exchange's piece on agent governance is a fair summary of what regulated firms are starting to require of AI agents, and most of it is platform work.

Where the bottleneck moves next

Adding ten more product managers won't close the gap. Headcount isn't the constraint; the distribution of time is. More standups won't fix it either, because the bottleneck isn't coordination; it's decision quality at speed.

‍

The harder truth is that even a well-resourced human PM function can't define scope fast enough to keep a fully agentic delivery team at full capacity. Discovery, research, regulatory interpretation, and spec writing all happen at human speed, and human speed is now the slow part of the system. Hiring through it is a 12-month answer to a 12-week problem.

‍

The agentic model has to expand upstream. Discovery agents that synthesize customer signal across support tickets, sales calls, and product analytics. Regulatory agents that flag jurisdictional constraints before they hit a spec. Draft-spec agents that produce the first version of an executable artifact for a PM to edit, challenge, and approve. The product team's job stops being to produce specs and starts being to govern the pipeline that produces them. The PM becomes an editor and a decider, not a typist.

‍

That shift is where product judgment compounds. A feedback loop that used to run in weeks now runs in hours, and the PMs who learn to drive the pipeline well end up making more high-quality decisions in a quarter than the previous operating model produced in a year.

‍

The cost of standing still is a regulated system that ships, accurately and at speed, the wrong thing — and an auditor with very specific questions about how that decision was made.

‍