Led 4 parallel research tracks—experience audit, 30+ app competitive study, user feedback synthesis, and telemetry analysis—to build a prioritized roadmap. Reframed chart drop-off from UX issue to adoption crisis, forming the strategic foundation for every charting feature shipped.
Liked this project?
Let's talk about what we can build together.
Introduction
How four parallel research tracks reframed Excel's charting problem, from a UX fix into a product adoption crisis.
Excel has ~400M users. Only ~8M create charts - a 98% drop-off. I led four parallel research tracks (experience audit, 30+ app compete study, user feedback synthesis, telemetry analysis) and converged them into a framework that reframed the problem: we weren't fixing a UX issue. We were solving a product adoption crisis hiding inside a legacy feature.
That reframe, and the prioritised roadmap it produced, became the foundation for every feature we built. The individual case studies (Modern Colors, Copilot Insights, AI Recommendations) each tell their own story - but every one answers a question this work asked first.
Results Overview
A preview of where four research tracks eventually landed - the full account is later, in What This Enabled.
Four parallel research tracks took the better part of a year to become a funded roadmap. Here's a preview of where that work eventually landed. The fuller account, including what still can't be proven, follows later in What This Enabled. Modern defaults reached general availability. 96% of users kept them. Chart Insights shipped its Crawl phase to an Insider ring - roughly two-thirds positive feedback, no fabricated figures, across more than 500 reviewed insight strings. And the blank-chart pane that preceded this work took recovery from 0% to 34%.
Starting Point:
No Brief, New Team, Loud Pain
September 2024. A new team, no roadmap, charting was our team's earlier tally at ~30% of Excel Web frowns - the OCV data itself puts it at 22%, the single largest dissatisfaction category. The context that shaped the structure.
September 2024. Newly formed team. No roadmap. Charting was our team's earlier tally at ~30% of Excel Web feedback frowns; the OCV data itself puts it at 22%, the single largest dissatisfaction category. The instinct was to jump in and fix the loudest complaints.
Instead, I structured four parallel research tracks — each answering a different question:
| Track | Question |
|---|---|
| Experience Audit | Where exactly does it hurt — in our own hands? |
| Telemetry & Funnel | How bad is it at scale? |
| User Feedback | Is our assessment real, or just our opinion? |
| Compete Study (30+ apps) | What does "good" look like — and where's the whitespace? |
The Conventional Wisdom (and the Paradox)
What the team assumed before the reframe - and the paradox that made picking a side impossible.
Before the reframe, the working assumption inside the team was narrower: Excel already had an Ivy/Legacy split in its charting stack, six paused feature bets, and a colour-modernization effort that had been deprioritized more than once. None of that had been assembled into one place. Buried in an archived deck was a predecessor result that should have changed the starting conversation: the HVS Blank Chart Pane had already taken blank-chart recovery from 0% to 25% to 34%, and healthy sessions from 68% to 78%. Nobody had connected it to what we were about to research.
The real tension wasn't a data gap. It was a paradox: by one measure, only 2% of sessions showed any charting struggle. By another, 98% of Excel's users never created a chart at all. Both numbers were true. They just described different populations - the people who tried, and the much larger group who never got that far. Resolving which one the team should build against is what the four research tracks were for.
My Role
What I actually owned on this work.
I was Lead Product Designer on Excel Charting, joining a newly formed team with no existing roadmap in September 2024. I structured and led the four parallel research tracks - the experience audit, the compete study, the user feedback synthesis, and the telemetry analysis - and converged their findings into the framework and roadmap that funded the three execution streams that followed.
Track 1:
Experience Audit
First-person walkthrough of the full chart creation journey — screen-recorded, timestamped, catalogued across 60+ friction points.
We used Excel to make charts from scratch — screen-recorded everything, timestamped every friction point across three scenarios.
| Scenario | Desktop | Web | Key Friction |
|---|---|---|---|
| Insert a chart | ~50s, 12+ clicks | ~30s, 7+ clicks | 3 decisions before any visual feedback. Hit-and-trial loop. |
| Basic formatting | ~80s, 20+ clicks | ~60s, 20+ clicks | Controls scattered across 4 surfaces. 20 clicks for basics. |
| Add elements | ~10s, 5+ clicks | ~24s, 8+ clicks | Web 2.4× slower. Inconsistent undo killed experimentation. |
We also catalogued 60+ craft issues — no snapping guides, overlapping labels, jargon-heavy UI ("bounds," "intervals"), and a Format Task Pane described internally as "extremely long and difficult to navigate."
Track 2:
The Funnel
About half of first charts never survived the session, and the data that forced a strategic pivot.
| Stage | Users | Drop-off |
|---|---|---|
| Excel MAU | ~400M | — |
| Aware of charts | ~368M | -8% |
| Creating charts | ~8M | -98% |
Within that ~8M: about half (45-58%, with real week-to-week volatility) deleted in the same session. 22% produced blank charts.
The pivot this forced: We'd been hired to fix customization for ~8M existing users. The funnel proved the real opportunity was the 360M who never tried. I led the hardest conversation of the project — telling partners "the thing we were hired to do is the wrong priority" — backed by three data points: the 98% drop, the ~45% discard rate, and the compete landscape. The reframe: not "abandon customization" but resequence — fix defaults first.
Leadership aligned. The roadmap shifted.
Track 3:
User Feedback & the 0% That Settled Every Debate
Thousands of OCV signals and one usability study — the 0% axis configuration rate that ended every internal debate.
NPS detractor pain-point breakdown, from ~800 verbatims (May-Oct 2024):
“Making a graph in the web version is painful."
"UI is cluttered... hard to find things."
"Web has less functionality than offline."
"Default chart looks outdated."
"Can't edit charts easily. Can't add a trend line."
Past research confirmed: trial-and-error was the dominant pattern, users searched externally (ChatGPT, social media) for guidance, and when presentations mattered, they exported to Canva. Even experts were frustrated.
The Ease-of-Use Study gave us the number that settled every internal debate:
| Task | Success | Ease (/5) |
|---|---|---|
| Configure axis | 0% | — |
| Set bounds/intervals | 43% | 1.6 |
| Change line colour | 86% | 3.7 |
| Add data labels | 57% | 3.2 |
| Overall | 47% | 1.6 |
Zero percent could configure a chart axis on Web. And right-click had 86% success — the highest path — validating that context-sensitive controls beat menu navigation.
Track 4:
Compete Study — 30+ Apps
Across AI-native, design-focused, and BI platforms — mapping where Excel sat on the ease-complexity matrix and where the gap was.
I pushed for breadth because Excel doesn't just compete with Google Sheets — it competes with every tool users reach for when they think "I need to visualize this."
| Competitor | Key Learning | Applied As |
|---|---|---|
| Napkin.ai | One-click, zero decisions | Sample data, AI recommendations |
| Canva | Template-first, beautiful defaults | Modern defaults that look polished instantly |
| Pitch | On-chart context toolbar | Floating toolbars for direct manipulation |
| Google Sheets | Explore suggests charts proactively | Contextual nudges on data selection |
| Power BI / Tableau | Auto-insights alongside charts | Copilot chart insights |
| Flourish | Vibrant modern aesthetics | Updated palette, typography, contrast |
The strategic insight: The upper-right quadrant — high ease AND high data complexity — was nearly empty. Every tool traded one for the other. Excel had the data power (400M MAU, formulas, enterprise trust). It just needed to move rightward on ease. Copilot was the bridge.
The compete study also decoded OCV: "I use Canva because it looks better" = aesthetic gap. "Charts suck" = first-impression problem. "I wish it told me what's interesting" = intelligence gap.
Convergence:
Four Streams → One Framework
The moment four contradicting research signals collapsed into one framework: Pre-Insert → Insert → Post-Insert.
By this point we had four research tracks each pulling in slightly different directions:
The experience audit said: "Fix the 20-click formatting journey, the fragmented surfaces, the 60+ craft bugs"
The funnel said: "Forget formatting — nobody's even making charts. Fix adoption."
OCV and usability said: "Everything is painful. Charts look bad. Web is behind Desktop. Customization is impossible."
The compete study said: "Everyone's ahead of you on ease, defaults, and intelligence. But nobody owns ease + complexity."
| Phase | User Question | What Research Said | Direction | Metric |
|---|---|---|---|---|
| Pre-Insert | How do I start? | 98% drop before insertion. No triggers. | Nudges, previews, sample data, AI recs | Insertion rate |
| Insert 🔴 P0 | Does this look good enough to keep? | 40% deleted. 0% could customize. Default IS the product. | Modern colours, typography, smart defaults | Chart Kept Rate |
| Post-Insert | What does this mean? | 20+ clicks for config. Zero insight generation. | Copilot insights, AI recs, floating toolbars | Retention, frown reduction |
The Roadmap
Every initiative prioritised, sequenced, and tied back to specific evidence — the framework that drove FY25-26 investment.
| Initiative | Phase | Priority | Timeline | Why |
|---|---|---|---|---|
| Modern Default Colours | Insert | P0 | FY25 H1 | 100% reach. Fixes 40% deletion. Low-med effort. |
| Copilot Chart Insights | Post-Insert | P0 | FY25 H2 | Fills compete gap. Drives dual flywheel (charts + Copilot). |
| Contextual Nudges | Pre-Insert | P0 | FY25 H1 | Attacks 98% discovery drop-off. |
| AI Design Recommendations | Post-Insert | P1 | FY26 H1 | Builds on modern defaults foundation. |
| Sample Data / Cold Start | Pre-Insert | P1 | FY26 H1 | Solves 22% blank chart sessions. |
| AI Chart Recommendations | Insert | P1 | FY26 H1 | Addresses 43% generic Copilot requests. |
| Craft Bug Fixes (30+) | All | Shield | FY25→ | Can't build AI on a broken foundation. |
Modern Colours ← Audit (dated output) + Compete (Flourish/Canva) + OCV ("looks outdated") + Funnel (~45% discard)
Copilot Insights ← Compete (Power BI generates insights; Excel had zero) + OCV ("tell me what's interesting") + Usability (0% axis config)
AI Recommendations ← Audit (hit-and-trial) + Compete (Napkin one-click) + Copilot data (43% generic requests)
Sample Data ← Funnel (22% blank sessions) + Audit (blank chart dead-end) + Compete (every modern tool shows something immediately)
What This Enabled
The research produced more than a roadmap. It produced organisational momentum.
Modern defaults reached general availability. 96% of users kept them. Chart Insights shipped its Crawl phase to an Insider ring - roughly two-thirds positive feedback, no fabricated figures, across more than 500 reviewed insight strings. And the blank-chart pane that preceded this work took recovery from 0% to 34%.
The number I would most like to give you is the change in the share of Excel Web users who insert a chart. Even the baseline is contested. One internal deck puts it at 0.89% in January 2025. A later one puts it at 0.5% in March 2025, repeated again in a September 2025 case study. The dates run backwards from what a growth story would need, and neither source has been retracted. A third count leans toward 0.89%. It's 0.89M Web chart users against 7.6M on Win32, and the arithmetic only lines up cleanly against one particular estimate of Web's total user base. I'll state both figures and both dates rather than pick one. What I won't give you is the movement. A number for it exists in our internal summary write-ups from six months later, but it traces only to those summaries - I cannot find the underlying telemetry that would let me stand behind it. So the baseline stays contested, the movement stays unverified, and I would not close on a number I cannot stand behind.
My Reflections
The hardest part of this project wasn't any single design decision. It was operating in an environment with no clear direction, multiple competing priorities, a legacy codebase with deep technical debt, and a team that was new to each other and to the problem space.
My job in that context wasn't to have answers. It was to build the framework that makes answers discoverable. The Pre-Insert → Insert → Post-Insert model emerged from weeks of systematically absorbing data from four different tracks until the structure became obvious.
There's a temptation at the senior level to focus on big strategic moves and delegate the small stuff. This project proved that craft is strategy. Each of those 60+ craft bugs was individually trivial. But their cumulative effect across a 200+ click journey is the experience users describe as "painful." You can't separate the forest from the trees when users are tripping on every root.
Every prioritisation argument in this project was backed by specific evidence:
Data didn't eliminate disagreement. But it moved debates from "I think" to "the evidence shows." That's a fundamentally different — and more productive — conversation.
Internal research had identified charting pain points for years. What changed the conversation was showing how competitors solved the same problems — and how far ahead they'd gotten. The compete study didn't introduce new information so much as it made existing information undeniable. When you can show leadership a side-by-side of Excel's default chart vs. Canva's, the argument makes itself.
Missed question: what had already been tried, and why it died. No graveyard was assembled up front, so we spent four weeks re-evaluating Legacy-to-Ivy trade-offs that archived decks had already settled, and the HVS Blank Chart Pane result surfaced late enough that we nearly rebuilt its lesson from scratch.
Missed question: what I was explicitly not looking at. There was no written boundary, and early AI Insights scope drifted toward conversational BI before it got pulled back to on-canvas callouts.
Missed question: what decision this changes, and when to stop. The audit had no stopping rule - it logged 60+ friction points when the first 15 had already justified the pivot.
Conclusion
A strategy doesn't need every question answered before it ships. It needs the right ones asked first, in the right order, before anyone commits budget or headcount to an answer.
Four research tracks ran in parallel, each one redirecting the others along the way. That gave this team something more durable than a single number - a shared, evidence-backed reframe everyone could act on. Modern Colors, Copilot Insights, and AI Recommendations all trace back to it. A fourth stream is still in flight.
That's the model. Not a UX fix chasing complaints. A product adoption crisis, argued with evidence, and prioritised accordingly.
This is the foundational research behind everything that followed. It combined the experience audit, the telemetry deep-dive, the user feedback synthesis, and the 30+ app competitive study. The individual feature case studies (Modern Default Colours, Copilot Chart Insights, AI Recommendations, and Sample Data / Cold Start, still in planning and pre-GA) each tell their own story. Every one of them answers a question this work asked first.
