A2

A four-person AI Enablement and Automation team at Lumen Technologies was absorbing AI solution demand from several thousand people. Requests arrived through email, Teams messages, scheduled meetings, and hallway conversations, with no two arriving in the same shape. There was no queue, no prioritization rule, and no way for a requester to learn the state of their own request except by asking the person building it. The team spent its capacity reconstructing context the requester already held, and requests that did not survive the reconstruction were backlogged.

The proposed solution sat inside enterprise governance. AI agents required registration with a platform Agent Registry and approval from IT Operations before production use. Structured record writes from an agent were blocked pending a separate approval. Anything touching direct seller engagement fell outside the marketing organization entirely.

Discovery

In a single week the team fielded access requests from five separate individuals in one consolidated effort. That established the scale of latent demand rather than the volume of tracked work, which located the bottleneck: the value of a request is created in the conversation that turns a vague ask into a structured brief, not in the build itself.

Framing

Intake then became the product. Everything downstream of that conversation was designed as the cheapest mechanism that would hold. A pipeline of simple parts that exists outperforms an elegant one that would have waited months for approval.

Tradeoffs

• A shippable pipeline over the most architecturally complex one
• Enable the requester over building for them: a team of four cannot scale by building for everyone
• Human routing over a scoring engine with no volume to score

Routing stays a human decision.

A scoring engine can rank requests it has volume to rank. With no volume, automated routing would have encoded a judgment nobody had made yet, and done so invisibly. I argued this position directly and it held across the intake, packaging, and status model.

The internal record and the requester-facing one are separate.

Two dashboards ran in parallel: delivery detail stays internal, status is visible to every requester and leader without a permissions conversation. Status was the most frequent interruption the team absorbed and thus needed remedy.

A five-stage pipeline carrying a request from a conversational intake agent, through a structured record, a routing decision, a generated PRD and proof of concept, and a status the requester can read without asking anyone.

AEOS won the Strategic Impact Award in the company’s inaugural AI in Action competition, and took honorable mention in Innovation, Scale & Reusability, Data & Decision Intelligence, Best Use of Microsoft AI, and Future of Work.

Core loop

Ask → structured brief → route: enable or in-house → build package → status the requester reads alone.

I designed the system before checking what the platform permitted.

The v1 architecture failed because it assumed an approval and a request volume that did not exist. The rewrite to a document-driven pipeline is why there was a working MVP to demonstrate. I would now open a build by establishing what the platform permits, then design to the permitted surface.