Problem: AI accelerated actions, not the workflow

A model can draft text, notes, or analysis faster. Approval queues, RFC returns, and context loss between Jira, meetings, and documentation remain unchanged.

The business needs changes in decision time, rework, Lead Time, and cost of error—not a larger volume of generated material.

A practical starting point

Pulsar already collected delivery data through ETL, PostgreSQL, and Metabase. That created the prerequisites for AI: accessible data, metric definitions, and a repeatable management rhythm.

Four verifiable agent tasks were selected: pre-development RFC review, delivery-risk digests based on Aging and blockers, decision-context assembly, and detection of divergence between requirements, design, and implementation.

Agent-centric loop

The agent collects signals and prepares options; a person defines intent, boundaries, and quality criteria; the agent performs a verifiable part; and a person decides where accountability, trust, and trade-offs matter.

This is not a “manager bot”. It reduces the mechanical work around a decision without giving the model authority to choose priority or accept business risk.

What had to become explicit

The workflow required current context sources, shared definitions, RFC quality criteria, a final decision owner, and a before/after outcome measure. “Employees use AI” was not accepted as success; each scenario uses decision time, return rate, manual review effort, Lead Time, or cost of error.

Result and limits

Pulsar brought data from 20+ teams into one management system. A leader can obtain the aggregate view in 1–2 hours instead of collecting it through many messages. LLM analysis remains an experimental layer whose value is tested through process impact.

This case does not attribute the RFC System Lead Time improvement to AI alone. It demonstrates the condition for organisational impact: data, criteria, and the human decision point must already be designed.

Business impact

Weak RFCs can return before costly implementation; current-work risk is assembled without manually visiting every team; decisions receive connected context; AI quality is measured through time, returns, and errors; and analysis can scale without proportional coordination headcount.

Before an AI pilot

  • Which repeatable workflow contains waiting or context loss?
  • Where is the human decision point?
  • Which criteria are verifiable?
  • Are current sources accessible?
  • Which business or process measure should change?

A delivery diagnostic selects one AI scenario that can be tested in a real flow before a broad AI strategy begins.