Confidential. Grow is a codename for an unreleased Constant Contact product. Reading this case study constitutes agreement to keep its contents confidential. Participant names and identifying details have been removed or changed.

Grow

Took a 0→1 agentic marketing product from an open-ended charter to a funded beta by finding who it was for, proving what they'd trust, and building research systems the team could build from, through two resets.

My Northstar prototype on mobile: onboarding, reviewing and editing the week's ideas, and checking results.

00 — Overview

Problem

Constant Contact set out to build an agentic marketing product for 1–5 person businesses: owners who squeeze marketing into 5–10 minute "stolen moments" and have been burned by tools and agencies. The charter was broad and the deadline close. There was no PM for six months, leadership changed mid-project, and research was spread across teams, vendors and outside researchers. Every reset risked starting from zero.

Solution

I chose the first customer with a scoring model instead of a hunch, and turned the research into tools the team builds from: a data-driven customer journey, a markdown knowledge base that AI tools read directly, an AI copy style guide and prototypes in code. Along the way I checked the AI's work, testing synthetic research against real people and catching AI summaries that dropped participants.

Result: green-lit and funded to production. Closed beta in October 2026, open beta in January 2027.

My Role

Principal Product Designer, 0→1. Partnered directly with the VP and Principal Engineer through June 2026 with no PM, then led a Senior Product Designer within a larger team.

  • Led the research strategy, and ran concept, functionality and usability testing on working product data
  • Built the ICP scoring model, persona, market sizing and customer journey that set the first target customer
  • Directed a UX researcher, and supported former IDEO researchers (contractors) on future-of-marketing and AI sentiment studies
  • Designed and built prototypes in Figma Make, Claude Code and v0, including the chat patterns
  • Sole designer on Send, a developer email product, from April to July while keeping Grow moving

Targeted impact

🎯 The right first customerA scored, evidence-backed ICP instead of "build for everyone"
🧭 Findings that get usedResearch that drives briefs, specs and prototypes, not a Figma file nobody opens
🗣️ Trustworthy AICopy that sounds like the owner's business, and research the team can trust
⚡ Less overheadTeammates get answers from the research themselves, without waiting on readouts

Discovery

"I don't want another tool. I want an expert partner I can actually rely on." — composite SMB owner voice

I joined 15 days before the company kickoff. Every team kept hearing "I don't have enough clients," so more broad discovery wouldn't move anyone. I reshaped it into a hybrid sprint: 20 minutes on each owner's problems, then 30 minutes pressure-testing concepts with a prototype.

What the first wave told us

  • Seeing it work was convincing. Net intent rose from 0.64 to 0.74 once owners saw the prototype.
  • Owners feared looking fake, not AI errors. "Selling feels fake" was the top pain point (63%). "Nervous about AI errors" ranked last (22%).
  • Control won. Easy edits scored a net intent of 0.92. Influencer outreach scored −0.37 and was cut.
  • Proof of money beat reach. A results tracker scored 0.85.

Choosing who to serve first. The first 27 conversations were mostly service providers. I pushed to narrow to product sellers, whose sales are trackable so we could prove ROI, with an exit if the need wasn't there. We grew the study to 63 owners across four segments and scored them on pain, activation, willingness to pay, time to value, reach and competition, down-weighting small samples and screening out "polite" yeses.

Confidence-adjusted ICP scores from 63 owners. Makers and merchants led on pain, interest and willingness to pay.

Makers (0.733) and merchants (0.725) led service providers (0.610). Makers spend about $109 a month on marketing tools and called $145 fair. Service providers spend about $45. The recommendation was a creative merchant who already sells on Shopify, captured as the persona "Marcus," with about 135K obtainable stores. Leadership aligned on it in February.

The creative merchant persona that set the first target customer.

Reframe: The product wasn't a marketing tool. It was a teammate owners could supervise.

Over ten months the research grew to 200 different small business owners across six studies. With a senior designer partner, I coded the 69 deepest sessions into the corpus behind the customer journey.

Research timeline · 200 SMB owners · 6 studies
  1. Dec '25 – Feb '26Concept validationSMB owners unmoderated
  2. Mar – Apr '26Usability + functionalityE-commerce working product data
  3. Apr '26Leadership resetResearch carried forward
  4. Spring – Summer '26Future of marketing + AI sentimentWith former IDEO researchers
  5. Aug '26Research PreviewSalon + med spa owners
  6. Sep '26Alpha + brand sentimentLocal retail owners
  7. Oct '26Northstar testingNext

200 different small business owners across six studies in ten months, with Northstar testing next.

Testing synthetic research against real people

Recruiting was the bottleneck: UserTesting qualified 2 of 1,600 product sellers. To see if AI could help, I ran a synthetic study of 40 AI-generated merchant personas in Gemini and compared it with real merchants.

What the synthetic study predictedWhat real merchants said
Struggling beginners who need a coach→Skilled professionals who lacked certainty, not knowledge
"Guidance" was the feature worth paying for→High-value owners put $0 on guidance and up to $75 of $100 on performance metrics
A $149/month price, anchored against hiring an agency→Most anchored around $20/month, the price of ChatGPT
The main fear was AI errors→The main fear was generic content: "AI has no vibe"

The verdict: useful for brainstorming, dangerous for pricing and priorities. On its own, it would have had us launch at roughly 7× what owners would pay, around the wrong core feature. We kept recruiting real people and used AI to speed up the work around them.

Making research impossible to ignore

With no PM and a principal engineer moving fast, research that sat in a deck was useless. So I treated findings like product infrastructure and used AI tools to put them where decisions happen.

Where findings usually liveWhere they lived on Grow
A readout deck nobody reopens→An interactive customer journey with a research receipt behind every claim
Annotations in a Figma file→Markdown files that Claude and v0 read directly to draft briefs, specs and prototypes
"Can you send me the research?" threads→A NotebookLM anyone on the team can ask, with cited answers
Copy notes in a handoff doc→An AI copy style guide loaded into every generation in the product
Static mocks→Working prototypes tested on real product data

Engineers and leadership got answers without waiting on a readout, one source drove briefs, specs and prototypes, and when the April reset came, nothing had to be rebuilt from memory.

The interactive journey: friction and ease across six phases. Each phase opens to its context, pain points, UX mechanics and research receipts.

Insight in action: the attribution split. The journey showed where churn would hit. Owners on Shopify or Square expect automated ROI, but custom and offline operators won't adopt a point-of-sale system at all. We designed two tracks, automated proof and guided logging, so neither group sees vanity metrics and leaves.

Build & test

To check that agent output could earn owners' trust before we scaled it, I ran 8 moderated sessions with creative merchants on a Claude Code prototype connected to working product data. Each owner first saw a marketing plan, images and a blog post generated from their own business. I synthesized 288 coded quotes in three passes, two with AI and one by hand. The AI summary alone left out 4 of the 8 participants, so the human pass carried the findings.

What we learned

  • Proactive, but in their control. All 8 wanted the AI to lead, and all 8 required approval before anything went live. Trust builds over months: babysit, then review, then semi-automatic.
  • Only sales matter. Revenue and new customers lit everyone up. Time saved was acknowledged, then forgotten.
  • A contractor, not software. 7 of 8 compared it to hiring help: "easier to hire an AI instead of waiting 3–4 days from a person."
  • The first action has to feel novel. A cart-abandonment email fell flat ("where's the romance?").
  • The AI copy guide worked. I shipped a style guide in a day that turned raw model output into plain language. Participants said it "sounded like their business."
  • Everyone missed the same step. All 8 missed the choice between entering a URL and connecting Shopify. That drove the next fixes: URL first, plan first, clear results navigation.

What changed after those sessions: from the original prototype to Grow's research preview, with what testers told us at each step.

A reset, and what carried forward

In April a leadership change reset the product direction and roughly four months of in-flight work. The research didn't reset. Because the findings lived in shared, AI-readable artifacts, the new team got up to speed quickly and rebuilt on the same evidence.

The new direction broadened the target to non-technical local businesses: a bigger market, but many run on closed booking platforms or no point-of-sale system at all, which made proving ROI much harder.

Testing kept turning into redesigns. The onboarding audit originally borrowed a competitor's loss framing, "your business is losing $X a month." Owners hated it, and 44% rejected claims that weren't backed by evidence. I reframed it around "where you shine" and "you could be making $X more a month," with every claim tied to something we found about their business.

Research-driven redesigns across the flow: lead with value, show owners their business instead of a checklist, less reading, generate then hand over the keys, and prove the work action by action.

The Northstar

I built the Northstar in v0 after months of iterating in Figma Make and Claude Code, testing along the way. It's mobile-first, because owners market in stolen moments, and every screen traces back to a journey phase.

  • Onboarding that learns the business: story, goals, links, brand voice and photo look
  • A weekly plan to review, not a blank page: recommended ideas, each with estimated impact, a "Why" and a "Not for me"
  • Review before anything goes live: every idea opens as a real preview, ready to edit and schedule
  • Results owners can read: revenue driven, new customers and clicks, with a chat for follow-up questions

Onboarding → goals → home → the week's ideas → results.

Chat usability: making an AI teammate easy to direct

Chat wasn't owners' first choice, but leadership wanted a future-facing interface. I designed patterns that keep it fast, guided and in the owner's control:

  • Tappable answers. The agent asks single or multiple choice questions, with room to answer in your own words.
  • Object-oriented editing. Tap part of a draft, like the header or hero image, and the prompt applies to just that. It became the app-wide pattern.
  • Photo pickers in chat. Real photos from the owner's library or Canva, over AI imagery, as the research demanded.
  • "Why" and preset pills. Every recommendation explains itself, and every screen suggests the next question.

Multi-select voice chips, tap-to-edit the header, an in-chat photo picker, and the reasoning behind each idea.

When leadership moved to a more chat-centric second Northstar, I designed its usability study around seven jobs-to-be-done, testing how much owners will let the agent do alone and whether the ROI report is credible enough to renew.

Designing for proof without the plumbing

Owners stay when they can see what marketing earned them, and a profit tracker was the most requested feature in the corpus. When the target shifted to businesses we couldn't connect to, I treated that as the biggest risk to retention.

  • Made the tradeoff visible. I used the journey to show where churn would hit without proof of ROI, and recommended starting with customers we could connect to, beginning with Square.
  • Disagreed, then committed. Leadership chose the larger market. I documented the risk and kept designing the strongest version of that bet.
  • Designed directional proof. One-tap sale logging, tallying where new clients came from, quick check-ins after campaigns, and estimates that always show their math.

Results designed around what marketing earned: revenue driven, new customers and every idea tracked.

Brand study

I also rebalanced and audited Grow's 30-person brand study, and caught AI summaries that left out the small business round entirely. It has its own case study: Grow brand study →

Constraints worth naming

  • What ships in closed beta is narrower than the Northstar because of technical constraints. The Northstar remains the target the team builds toward.
  • The April reset set back roughly four months of in-flight work.
  • Proving ROI for local businesses without integrations is the biggest open risk to retention. I continue to recommend an integration path such as Square.

Impact

Grow is funded to production, with a closed beta in October 2026 and an open beta in January 2027. Concept validation held up across every study. The journey, knowledge base and AI copy guide remain the team's shared foundation for briefs, specs and every generated line of copy.

With the foundation in place, I handed Grow to the team and moved on to bring the same research-to-product methods to Constant Contact's core email experience.