Portfolio recommender

The problem

This came first, in 2022: one kind of decision, share expansion and the moves around it, run on business rules, as the first AI feature in Dr. Reddy’s portfolio tools. It shipped inside the early Panorama product shell, which is why the screens carry its name. Panorama 360 later rebuilt the whole landing page on what it taught.

From the docs

60 percent of a product manager’s time went on analysis instead of strategy.

• Two weeks per analysis. Business rules applied by hand across large spreadsheets.
• Errors and no history. No version control, manual rules, no record of past recommendations.
• The cost. Delayed product decisions and inconsistent recommendations across the portfolio.

What the research found

How. Interviews with 15 PMs, shadowing quarterly planning sessions, and a workflow analysis of the tools in use, structured with jobs to be done.

• The functional jobs. Analyse portfolio performance, generate strategic recommendations, track market opportunities, manage rule-based decisions.
• The emotional jobs. Feel confident in data decisions, spend less time on manual analysis, show strategic value.
• What they did today. Large spreadsheets with no broad view, constant number crunching before any step, and no way to filter out a type of recommendation.
• What they wanted. A small, almost one-click flow, a history of past recommendations, a zoomed-out view, and rules applied for them.

What PMs did and said	Design decision
Constant number crunching before any step	Rules applied automatically; the PM checks the result
Large spreadsheets with no broad view	A portfolio view first, then one recommendation at a time
No record of past recommendations	Every recommendation closes with an action, and a no records why
No way to filter out a type of recommendation	Filters by impact and opportunity on the list
Wanted a small, almost one-click flow	Review, decide and act on one screen, reasoning beside the data

One decision, three ways out

Every recommendation ends in one of three actions, so nothing is left half-decided in a spreadsheet:

1. Convert it to a task, with the inputs and outputs attached.
2. Edit it, changing an input and seeing the output recalculate.
3. Disqualify it with a reason, so the next PM sees why it was turned down.

Version A against version B

Version A · modal detail

Version B · side panel, shipped

1. Version B was 40 percent faster to complete.
2. PMs preferred the reasoning beside the data, not on top of it.
3. A bookmark feature request came out of the test.
4. The insight: PMs want the why without losing context.

Patterns built to outlast it

• Recommendation cards for AI insights.
• Drill-down tables for sortable data grids.
• Rule visualisation to show the business logic.
• Progress indicators for multi-step guidance.

Together they kept five areas consistent and made each new area faster to build.

After launch

Version one · the recommendations list

Added in the analytics round · disqualify with a reason

• Adoption. 45 percent tried the new workflow in week one, 73 percent used it as their primary tool by month one, 89 percent were still using it at month three.
• Efficiency. Analysis from two weeks to ten minutes. Most of that is the rules running automatically; the design’s share is that checking each result takes one view instead of a spreadsheet hunt. 35 percent more recommendations implemented, satisfaction 8.4 out of 10.
• Requested next. Bulk export, collaborative comments and a better mobile experience.

The recommender later grew into Panorama 360.

With the team

• Lead designer and researcher, with two product owners, four engineers and 25 users giving weekly feedback.
• Data scientists on explainability, engineering on real-time data, product owners on the balance between users and technology.
• Design workshops with a fifteen person team kept the stakeholders aligned.