ProductPlan strategic portfolio visualizations
ProductPlan
Closing the Idea-to-Strategy Gap: Designing ProductPlan's AI-Suggested Initiatives
ProductPlan is a roadmapping and strategic portfolio tool used by enterprise product teams. In late 2024 the company bet on AI to close its biggest workflow gap: ideas flowed in, but almost none became strategy. I led the design discovery, competitive analysis, and prototyping that became AI Suggested Initiatives, the feature that reads a team's idea backlog and proposes initiatives a PM can accept, reject, or edit.
27%
Increase in engagement in the ideas portal feeding the pipeline
7 months
Innovation Week prototype to GA
My Role:
I was the senior product designer on ProductPlan's AI initiative, working with a product manager and an engineering team. I owned the quantitative discovery across our production database and Pendo, the competitive analysis of six AI-enabled competitors, the segment analysis that defined our four target personas, the design direction for the AI feature set, and the design of the idea-clustering prototype built during the company's AI Innovation Week.
The shipped feature was my design end to end: the suggestion review experience, the account-level AI opt-in in settings, and the engineering handoff for the GA release. I also designed the Initiatives framework that the AI suggestions create into, which meant I controlled both ends of the idea-to-strategy pipeline.
Constraints:
The hardest constraint surfaced before design started: everything had to run on ProductPlan's existing stack, with Amazon Bedrock as the only new dependency. That ruled out model fine-tuning, external data pipelines, and anything requiring new infrastructure. Enterprise trust added a second constraint surfaced by the segment analysis: customers navigating privacy and compliance requirements would not accept AI touching their data without explicit, auditable consent, which shaped the account-level opt-in that gated the launch.
THE PROBLEM
Ideas flowed in; strategy never came out
ProductPlan had an ideas inbox and a roadmap, but the path between them was manual synthesis that rarely happened. Since 2022, customers had attached only 76 ideas to opportunities across the entire platform. The objects existed; the workflow between them did not.
The cost showed up downstream. In our PDLC research, 60% of product managers reported struggling to align ideas with strategy, which contributed to an estimated 30% longer time-to-market for affected teams. Enterprise customers were exporting ideas to Excel, Miro, and Confluence to do the clustering work ProductPlan should have supported.
My earlier work on customizable idea intake forms had made the problem more urgent, not less. That redesign lifted ideas portal engagement 27% in its first month, which meant more raw input piling up against the same synthesis bottleneck. The business trigger was clear: competitors were shipping AI features fast, and the idea-to-strategy gap was the place where AI could do work no competitor's bolt-on could.
Who it was failing
The primary user was the enterprise product manager: data-driven, cross-functional, and managing hundreds of overlapping idea submissions from sales, support, and customers. I profiled four segments (enterprise PMs, B2C PMs, Heads of Product, and project managers) through market and competitive research in order to understand what each would accept from AI. The same tension ran through every profile: demand for automation, paired with refusal to let AI make roadmap decisions.
That tension became the key insight. The PM's job is judgment, but most of their time went to the taxonomy work that precedes judgment: reading, deduplicating, and clustering ideas into themes. The right AI feature would do the clustering and hand the judgment back.
RESEARCH AND DISCOVERY
Validating that the gap was synthesis, not collection
The obvious diagnosis was that idea collection was broken, and my custom fields work had already improved it. The discovery here was quantitative: I queried our production database and Pendo usage data in order to find where the workflow actually stalled, and the answer was consistently downstream of collection.
The numbers told a clean story. Idea volume was growing, but only 76 ideas had been attached to opportunities since 2022 across the entire customer base. Teams were not failing to have ideas; they were failing to convert them, and the conversion step was unassisted manual reading.
COMPETITIVE LANDSCAPE
Finding the opening: every competitor bolted AI on
I analyzed Zeda.io, Aha!, Productboard, Chisel, Mixpanel, and Pendo in order to find where ProductPlan could differentiate rather than chase. The pattern across all six: AI added to existing surfaces, summarizing feedback or drafting documents, with no understanding of the customer's strategic context.
ProductPlan's data model already linked ideas, opportunities, objectives, and roadmap bars. AI that reasoned across those relationships, suggesting initiatives and flagging which strategic objectives they aligned with, was something no competitor's architecture supported. I made context-awareness the first design principle: any AI feature that could work in a generic document editor was not worth building.
INNOVATION WEEK
Proving clustering worked against real data
During the company's AI Innovation Week I partnered with engineers to build a working prototype on Amazon Bedrock in order to test whether clustering was viable with prompt engineering alone, since fine-tuning was off the table. The prototype took a full idea backlog and returned five suggested opportunities per run.
Two design decisions in the prompt mattered more than the model. First, I capped the first draft at five suggestions per run in order to keep the output reviewable rather than another backlog to triage; the shipped feature raised that ceiling to ten after engineering and I weighed generation time against user value. Second, I fed the model the list of opportunities that already existed on the board so suggestions were deduplicated against current strategy. Without that step the output looked impressive in a demo and created cleanup work in real use.
Five suggestions per run, deduplicated against the existing board, so output is reviewable rather than another backlog.
Trust was a top pain point in every segment profile, so nothing touches the roadmap without explicit acceptance.
DESIGN SOLUTION
The hardest call: suggestions PMs own, not automation
The tempting direction was a persistent AI assistant, the pattern Chisel used, available as a context menu across every surface. I rejected it for two reasons: every segment profile flagged trust and transparency in AI-driven decisions as a top pain point, and a floating assistant would be bolted on rather than embedded in the workflow where the synthesis problem lived.
Instead I designed the feature as suggestions inside the existing ideas workflow. The AI proposes initiatives as problem statements, flags potential alignment with strategic objectives, and stops. The PM accepts, rejects, or edits every suggestion before anything touches the roadmap. This was slower to demo and better to ship, and it is the interaction model that shipped at GA.
Designing for the model being wrong
A clustering model is sometimes wrong, and the design had to assume it. Rejection is a first-class action with the same weight as acceptance, so a bad suggestion costs the PM one click instead of eroding trust in the feature. Deduplication against existing opportunities removed the most predictable failure, the model proposing work already on the board.
How many suggestions to surface was a decision I worked through with engineering rather than a pure design call. Generating more suggestions meant longer load times, but too few meant a PM waited on the model and got little back. We settled on a ceiling of ten as the point where the value returned justified the wait.
The floor mattered as much as the ceiling. The feature requires a minimum of six ideas before suggestions can run, so the model never clusters a backlog too thin to produce meaningful themes.
Enterprise trust research demanded auditable consent, so the governance surface was designed alongside the feature, not after it.
Designing consent before designing features
Trust also needed an account-level surface, not just per-suggestion controls. I designed the AI opt-in experience in account settings in order to give organizations explicit, auditable consent before any AI feature touched their data: an administrator-only toggle that records who enabled it and when, covering current and future AI capabilities in one decision.
Designing the governance surface alongside the feature meant legal and enterprise-trust concerns never blocked the launch. The opt-in shipped in the same June 2025 release and is documented in ProductPlan's release notes and support docs.
Rebuilding the destination so suggestions had somewhere to land
AI-suggested initiatives are only as useful as the object they create. The legacy Opportunities object held a title and a plain-text description, which is why customers were writing business cases in Confluence instead. I had redesigned that object into the Initiatives framework, with rich text, customizable templates, stakeholders, status, and custom fields, in order to make an initiative something a team could actually prioritize and defend.
Unlike the AI discovery, this redesign was grounded in qualitative research. User interviews and usability testing on the idea forms and Initiatives work showed teams managing business cases across scattered tools and named rich text as a must-have, and those findings defined what the new object had to hold.
That sequencing was deliberate. Shipping AI suggestions into the old Opportunities object would have automated the creation of artifacts nobody used. The AI feature and the Initiatives framework were one pipeline designed end to end: intake forms feed ideas, AI clusters them, and Initiatives carry them to the roadmap.
Rebuilt the destination first; AI suggestions into the old object would have automated artifacts nobody used.
RETROSPECTIVE
Outcome
The clustering prototype shipped as AI Suggested Initiatives on June 4, 2025, seven months after the Innovation Week prototype. I designed the shipped UI and supported the engineering handoff through the GA release, along with the AI opt-in that gated it.
The shipped feature carried all three design principles from my direction document. Suggestions are generated from the customer's own idea backlog and checked for strategic objective alignment (context-aware). Every suggestion requires explicit PM acceptance, rejection, or modification (assistance, not automation). The feature lives inside the existing ideas workflow rather than a separate AI panel (embedded, not bolted on).
The GA release surfaced up to ten suggested initiatives per run, linked each suggestion back to the specific ideas that generated it, and supported bulk accept and reject. All of that was in the original release and still describes the feature in ProductPlan's current support documentation.
The intake side of the pipeline had already proven out: customizable idea forms drove a 27% increase in ideas portal engagement in their first month, measured against the prior month's active usage. That growing input volume is what the AI synthesis step was built to absorb.
What I'd do differently
I left ProductPlan in September 2025, three months after GA, before adoption data had time to mature. I would have pushed to instrument suggestion acceptance and edit rates from the first beta, because acceptance rate is the real measure of whether AI suggestions earn trust, and I left without that answer.
My discovery for the AI feature itself was entirely quantitative: production data, Pendo, and competitive analysis. It was enough to find the bottleneck, but I would have validated the trust findings with customer interviews the way the Initiatives redesign had been, because segment profiles built from market research are a hypothesis about people, not a conversation with them. I also underestimated how much of the design work would be prompt design; treating the prompt as a designed artifact with its own iterations and reviews is something I now do from the start.
Featured Case Studies
Bridging the Gap in Strategic Planning with ProductPlan
Driving strategic alignment with 54% month to month retention boost with portfolio visualizations
Curriculum Management for School and District Admins
Increasing Daily Active Usage by 15% through optimization of curriculum management
