Amit IsteinProduct Designer

WorkBecky

AI backlog governance for product teams

Becky is an AI-powered governance layer that helps product organisations keep their backlog complete, consistent and trustworthy — without replacing the tools they already use.

A Jira backlog for a demo project with the Becky panel open on the right. The panel sums up suggestions across the backlog and groups them into Auto Fix (Safe) with an Apply All Safe Fixes button, AI Suggestions listed by missing field with a Review button each, and Manual Review with a View list button.
Type
Independent product project
Year
2026
My role
Product vision, AI workflows, governance model, MVP strategy, UX

Overview

  1. BeckyGovernance and data-quality layer: detects gaps, drafts what’s missing, keeps it consistent.
  2. JiraStays the system of record. Planned next: Azure DevOps, Monday.com, Linear.
  3. Product teamsKeep working where they already work.
Becky is not another product-management or Agile tool. It sits on top of the platforms teams already use.

The problem

The problem isn’t simply “bad data”. Backlogs slowly accumulate missing and inconsistent information:

  • Missing mandatory fields, incomplete user stories, missing acceptance criteria
  • Inconsistent prioritisation, epics without business goals, missing ownership
  • Incorrect or inconsistent labels, and governance rules nobody follows
Product managers
Spend time maintaining backlogs instead of managing products.
Scrum Masters and Agile Coaches
Enforce governance by hand.
Leadership
Gets reports built on incomplete or inconsistent data.
Everyone
Plans and decides with product data that is harder to trust.

The deeper problem

Poor backlog quality ultimately leads to poor decision-making.

The design question

How do you design an AI product that improves the quality of the product data teams already depend on?

Most backlog analysis tools stop at detection: they flag missing acceptance criteria or stale epics and leave the clean-up to a person. Becky goes further — it fixes what it safely can, drafts what it can’t, and does it inside the tool teams already use.

Detect, Complete, Maintain

  1. Detect

    Automatically identify governance issues in an existing backlog.

  2. Complete

    Use AI to suggest or generate missing information: acceptance criteria, business value, labels, priority, organisation-specific metadata.

  3. Maintain

    Continuously monitor backlog quality, so governance doesn’t become another manual task.

Where AI belongs — and where it doesn’t

Not every fix deserves the same treatment. Every finding is sorted by how much judgment it needs, and each kind gets a different experience.

Deterministic

Auto Fix (Safe)

Rule-based changes with zero AI judgment.

  • One summary, one action: Apply All Safe Fixes

AI-assisted

AI Suggestions

Missing content Becky can confidently draft.

  • Reviewed one issue at a time
  • Current vs. suggested, side by side
  • Confidence and a “Why?” explanation
  • Accept writes to Jira; Reject writes nothing

Human

Manual Review

Structural problems, or where AI has no confident signal.

  • Grouped by owner
  • Notify owners, or email the unassigned items
The Becky panel. Auto Fix (Safe): 7 auto fixes, described as deterministic with no AI judgment involved, and an Apply All Safe Fixes button. AI Suggestions: issues missing acceptance criteria, description, labels, estimate and summary, each with a Review button. Manual Review: issues that need Product Owner attention, with a View list button.
The three kinds of finding, as the panel presents them. The counts come from a deliberately messy demo backlog, not a real team’s data.

Decision 1

Review AI like a pull request

An earlier version listed AI suggestions as checkboxes grouped by field, applied later in one step. The redesign groups them by issue, so an issue with two drafted fields is reviewed once, and moves through them like a code review: previous, next, and status for each one.

Why It matches what actually reduces a product manager’s effort.

Decision 2

Show the least confident number

When one issue has several drafted fields, the confidence shown is the lowest of them, with the reasoning one click away.

Decision 3

Write exactly what was reviewed

Accepting sends back the exact values the person saw — nothing is regenerated at write time.

Governance model

  1. The organisation’s rules, not ours

    Consistent with the organisation’s own conventions, not an arbitrary external standard.

  2. Policy-driven, traceable, reversible

    Any change Becky makes must be explainable, bound to policy, and undoable.

  3. A hard constraint, not a nice-to-have

    Becky writes to production project-management systems, so auditability and safety come first.

Inside Jira, not beside it

The prototype is a browser extension that adds one collapsible Becky panel to the real Jira page. It never recreates any part of Jira’s own interface, and its styles are isolated so neither side can affect the other. The screenshot at the top of this page shows it running on a demo backlog.

  1. Analyze Backlog

    Reading issues, running governance rules, running AI analysis

  2. Backlog Health

    A simple, transparent score, findings and recommended actions

  3. Fix My Backlog

    Findings sorted into safe fixes, AI suggestions and manual review

  4. Review

    Apply, accept or reject, notify owners

  5. Review Complete

    What was applied, rejected or failed, with retry

MVP strategy and current state

  1. Jira first, read-only

    Connect to a real Jira Cloud project and map every issue into Becky’s own platform-agnostic model.

    Current milestone

  2. Analysis and fixes

    Governance rules and AI drafting, reviewed in the Jira panel.

    Prototyped

  3. More platforms

    Azure DevOps, Monday.com, Linear — each a self-contained integration.

    Planned

A deliberately messy backlog
A demo generator seeds a fictional payments company, “BeckyPay”, with 50 issues of inconsistent quality, and records which problems were injected — ground truth to check Becky’s analysis against.
Build only what’s needed
Every milestone ships something working and demonstrable; abstractions wait for a second real use case.
AI-assisted development
Built with an AI coding assistant working to my product and engineering ground rules.
Validation
Customer validation is the next step. Nothing here has been tested with customers yet.