1
Plan
Define audience, goals, and channels for upcoming campaigns.
Strategy lives in slides; tactics live in spreadsheets.
Steady
Case Study
Designing AI marketing tools at HubSpot
I led a nine-person design team across two product groups and set the AI design strategy for the four products we shipped inside Marketing Hub. Over four months of a rapid design overhaul and AI adoption, Marketing Hub's CSAT rose from 70 to 84 across all users, and setting up a campaign took half as long.
Role
Senior Design Manager
Team
9 designers across 2 product groups
Timeline
4 months: design overhaul and AI adoption
Shipped
AI Campaigns Assistant · Content Remix · AI Copilot · Social AI Composer
Outcomes
+14 pts CSAT (70 → 84) · 2× faster campaign setup
A platform doubling down on AI
I joined HubSpot's Marketing Hub at the end of 2022, as the platform was doubling down on AI. Marketing Hub helps businesses attract, engage, and grow their audience through automation, CRM, and personalized content, and my team owned its design across every tier, from Starter to Enterprise.
The work I cared most about was making AI pay off in the work marketers did every day. The AI Campaigns Assistant cut the time it took to produce targeted, on-brand content, and multi-touch revenue attribution showed which of that content actually drove revenue, so teams could see what the time savings bought them.
Two opposing pressures, one workflow
Marketing teams were being asked to ship more campaigns, faster, while making each one feel personal and on-brand. AI could help with both, but only if it fit the way marketers already worked and left the judgment calls with them.
1
Disconnected orchestration
Campaigns required coordination across email, social, ads, landing pages, and blog. The workflow lived across disconnected tools.
2
Parallel AI surfaces
AI features were arriving from multiple teams in parallel, each with its own UX patterns and trust model.
3
Cautious customers
Marketers were curious but skeptical. We needed to earn trust before asking them to delegate work.
Building the team, then the work
As Senior Design Manager, I led nine product and content designers across two product groups. My first priority was the team itself: establishing clear ownership, calibrating critique, and aligning on what "good" looked like for AI features specifically.
Product Group 1
Reach & Reputation
AdsSocialAIInsights
Product Group 2
Marketing Orchestration
PlatformCampaignsAI AssistantMeasurement
9
Designers managed
I set the AI design strategy across both groups and worked with PM and engineering leadership on the order the releases went out in. I stayed hands-on in critique on every AI surface, especially anywhere two agents showed up in the same flow, because that's where products built by different teams were most likely to contradict each other.
Three principles for AI in marketing tools
The team converged on a small set of principles that shaped every AI surface we shipped.
Principle 1
Visible by default, never silent
Marketers see what the agent did and why, and they can edit it. Every AI action leaves a clear trace.
Principle 2
Editable in the canvas
Outputs land in the same fields and states marketers already know, so they edit AI's work the same way they edit their own.
Principle 3
Trust earned, not assumed
Each AI surface had a clear handoff: preview, edit, regenerate, approve. Marketers stayed in the driver's seat.
Where marketers got stuck
Those principles came from real research. Across five campaign phases, friction concentrated in the middle, where building and executing fragmented across tools and teams.
2
Build
Brief, draft, review, and approve assets across teams.
Work fragments across email, sheets, and project tools.
Strained
3
Execute
Schedule, publish, and monitor campaigns across channels.
UTM tracking is manual; nothing lives in one place.
Frustrated
4
Measure
Pull performance data and assess what moved the needle.
Attribution means stitching together five reports.
Mixed
5
Inform
Share results with stakeholders and shape the next plan.
Lessons rarely loop back to the next campaign's brief.
Lifted
Mapping AI to where marketers felt friction
From those friction points, each AI surface mapped to a specific moment in the marketer's lifecycle. The five phases collapsed into four where AI could move the most work.
01
Strategy
Starting from a blank page each year. Aligning goals, audiences, and bets across product and brand teams.
02
Calendar
Coordinating channels, dates, and owners. The plan lived across disconnected tools and inboxes.
03
Campaign
Producing high-volume, on-brand assets across email, social, ads, landing pages, and blog, and doing it fast.
04
Performance
Knowing which assets actually moved revenue. Closing the loop back to next year's strategy.
Where AI helped
Templates
Where AI helped
AI Campaigns Assistant
Where AI helped
AI CopilotSocial AI ComposerContent Remix
Where AI helped
Multi-touch attribution
AI agent shipped under my team
Four agents, four jobs
Each AI product targeted a different point of friction. Together they covered the marketer's content lifecycle, and the seams between them are what we'd refine next. My team also designed the agent catalog where marketers picked AI Specialists and AI Teams to plug into a campaign.
AI Campaigns Assistant
Generates campaign briefs, target audiences, and asset drafts from a marketer's prompt, taking a plan to first drafts in one step.
Content Remix
Reformats one piece of content into channel-native versions, like a long blog into a thread or an email into social posts, so one idea can run everywhere.
AI Copilot
In-context help inside the editor to draft, edit, summarize, and translate without leaving the page.
Social AI Composer
Channel-aware social posts with each platform's voice, length, and CTAs built in, so marketers don't have to rewrite for every network.
From brief to live campaign
A look at the AI Campaigns Assistant in flow: plan, build, and distribute, with measurement closing the loop.
Designing for an AI that made things up
This was HubSpot's first generation of generative AI, and the models hallucinated. We knew they would, so we built guardrails into every surface to catch a bad answer before it reached a customer.
How we tested it
Tested in widening rings
Each feature went to HubSpot's own marketers first, then to beta customers, then to our Solutions Partners and the product council before general release. Every ring was another chance to see where the output broke while the stakes were still low.
Failure by design
Small jobs, so mistakes stay small
Agentic AI was experimental, so we kept each agent's task simple and chained the steps: brief, then audiences, then drafts. A hallucination stayed inside one step the marketer could check. The price was a workflow that felt fragmented, the first seam I'd fix today.
Human oversight
A person approves every draft
Nothing an agent drafted went out until a marketer signed off. Brand voice settings kept drafts sounding like the company, and admins could switch AI features off for their whole account.
What could go wrong, and the way back
What went wrong
What the marketer saw
How they recovered
A made-up fact, number or claim
What the marketer saw
A draft in the editor, never a published asset
How they recovered
Edit it inline, or regenerate before approving
Copy that didn't sound like the brand
What the marketer saw
A draft shaped by the account's brand voice settings
How they recovered
Tune the voice settings, or edit the draft
A task too big for the model
What the marketer saw
Small chained steps, each with its own preview
How they recovered
Fix the step that went wrong and keep going
A team not ready for AI
What the marketer saw
AI features an admin can switch off
How they recovered
Turn them on when the team is ready
How customers took to it
HubSpot reported these on its earnings calls. They cover all of HubSpot's AI, my team's products included.
54%
Content Hub attach
Up from 13% in 2024, which HubSpot credited to Content Remix, its most popular AI feature
75K+
Weekly Copilot users
By the end of 2024
43%
Copilot repeat use
Of users came back during the public beta
2/3
Enterprise on AI
Of enterprise and half of Pro customers using AI features, Q3 2024
What changed in four months
CSAT comes from HubSpot's Marketing Hub-wide survey of all users, taken before the overhaul and again four months later. Over the same four months, setting up a campaign took marketers half as long, with the Campaigns Assistant now drafting the brief, audiences, and first assets.
+14pts
CSAT lift
70 → 84 on the Marketing Hub-wide survey, all users
2×
Faster setup
Setting up a campaign took half as long
4products
Shipped in window
Campaigns Assistant, Remix, Copilot, Social Composer
9designers
Team led
Across 2 product groups, 4 surfaces
The decisions worth naming
Every shipping window forces a choice. Two stand out from this one.
Trade-off 1
Speed vs. cohesion
Chose
Ship four AI products in parallel
Over
Sequencing them behind one unified system
I pushed to ship all four in parallel, because steady releases showed customers real momentum and earned their trust sooner than one big launch would have. The cost was that each product's UX grew on its own track, so we standardized the shared patterns after launch instead of before.
Trade-off 2
Generation vs. control
Chose
Scaffolding around AI output
Over
Letting the model “just do it”
Early prototypes leaned on raw generation. I asked the team to add preview, edit, regenerate, and approve to every AI surface, even where the model could have skipped them, because marketers who were skeptical of AI would only hand over work they could check first.
What I'd give more time to
The thing I'd give more time to, in retrospect, is the seam between agents. We shipped Campaigns Assistant, Content Remix, AI Copilot, and Social AI Composer in parallel. Each was strong on its own, but the handoffs deserved more attention than the schedule allowed.
Some of this was the moment in AI: this was before LLMs could do what they can now. We were designing for narrower, more deterministic models, and “agent collaboration” meant something closer to coordinating four well-defined products than orchestrating one fluid assistant.
With current models, I'd revisit how those four surfaces compress into fewer, more capable agents, and how the marketer moves between them feeling like one continuous conversation.
