Agentic Workflows in IT Project Management: Lessons from the Field

Field report: 6 months of agentic experimentation

In 2024, I integrated agentic workflows into the daily management of the EQUANS Site Factory project at Talan/Guidepoint. Here is what I learned — successes, failures, and what I would have done differently.

Workflow 1: Automatic sprint report generation

Every Friday, an agent extracts data from Jira via the REST API, calculates team velocity, flags tickets blocked for more than 48 hours, generates a structured Markdown report, and automatically posts it to the team Slack channel.

Result: 3h30 saved per week. Team adoption rate: 100% by week 2. Report quality rated "better than the manual version" by 4 out of 5 team members.

Workflow 2: Incoming ticket triage and prioritisation

An agent analyses each new Jira ticket at creation: it evaluates criticality against business criteria, suggests a priority, identifies the impacted component, and adds the appropriate labels. A human validates the suggestion with a single click.

Result: triage time reduced from 45 minutes to 8 minutes per day. Prioritisation consistency significantly improved.

Workflow 3: Automatic technical documentation

Upon closing a "feature" ticket, an agent automatically generates a documentation sheet: feature description, user impact, technical dependencies, rollback procedure. This sheet is submitted for validation before being archived in Confluence.

"Documentation is no longer an end-of-project chore. It's generated continuously, validated by the team, with no extra effort."

What didn't work

The automatic estimation agent

I attempted to deploy an agent that automatically estimated ticket development time based on historical data. Error rate of 40% — too high to be useful. The project's contextual complexity was too specific to capture without fine-tuning.

Lessons learned for your deployments

• Start with the most repetitive processes: reporting, triage, notification
• Keep a human in the validation loop at the start — even if the agent is 95% reliable
• Document system prompts the same way you'd document code
• Measure systematically: time saved, error rate, team satisfaction