Account-based marketing · Agentic AI · HubSpot · 2026
From 200+ manual steps to one chat: agentic, personalized pages for every target account
The business was losing account-based marketing to competitors’ tools because we couldn’t personalize content at scale. I’d already stopped one AI effort that missed the quality bar, so I quickly tested AI output on real campaigns before building anything. Content passed; design didn’t. That split shaped the product.
My role
UX lead, strategy to end-to-end. Wrote the requirements that shaped the engineering approach.
Team
Triad (PM/UX/FE/BE), plus adjacent teams.
Timeframe
Apr–Sep 2026
Scope
0→1: one personalized page per account, created in chat; agentic templates as the governance layer.
The call I made
Test AI quality with customers before building the UX. Design stays deterministic; content non-deterministic, within guardrails.
Result
Content rated 4–4.5/5; funded the private beta, launched Sep 2026
customers named tool fragmentation as their main pain
70%
of HubSpot customers attempting ABM, taking it to competitors’ tools to execute
GTM teams need hyper-personalized content to win those accounts. In our product, 50 personalized pages took 200+ manual steps, so customers were leaving for competitors. In discovery, 8 of 9 named tool fragmentation as their main pain, 6 of 9 already used AI elsewhere, and sales reps were sending off-brand, untracked AI pages built with Claude to their buyers.
The risk wasn’t the UX. Two earlier AI efforts hadn’t cleared the bar: the AI-gen output wasn’t good enough to trust, publish or even start editing.
“I would love to hit a button and have 65 landing pages.”Head of Growth Marketing, B2B SaaS customer
My decisions
01
Tested the AI-gen output quality before building the full experience
I ran the alpha as a human evaluation: a quick POC where customers created real pages for real campaigns, live, and rated visual and content quality separately on a 0–5 scale where 5 meant “I’d publish this tomorrow”.
What I didn’t do
Design the full creation flow first
Trade-off I accepted
Weeks before any shippable UI
02
Split the page into two layers: deterministic design; non-deterministic content, within guardrails
The ratings split cleanly: off-brand visuals broke trust even when the copy was strong. I set the new model: pages now start from an existing on-brand page, so the design stays deterministic, and AI fills the non-deterministic content without touching the visual rules. Requirements I got from user research shaped a new engineering approach.
Human eval of AI-gen content
AI-gen content(up to 4.5/5)
AI-gen design(as low as 1/5)
012345
Customer ratings from the alpha, on a 0–5 scale where 5 meant “I’d publish this tomorrow”.
What I didn’t do
Let AI generate the whole page
Trade-off I accepted
Less generative freedom, so brand safety stops blocking adoption
03
Designed for two personas: Ops writes the templates with rules, everyone else creates
Research surfaced two mental models: Marketing Ops wanted to set visual and personalization rules up front; marketers and sales reps wanted pages in minutes, with no prompt engineering.
With the triad, we split the product: Ops authors AI templates (deterministic design, plus non-deterministic personalization instructions), and everyone else uses them. 3 of 5 participants backed it: “Sales and marketing should not be setting up templates.”
For people who never build a template, I added a curated library: select, generate, publish.
What I didn’t do
Build one flow for both personas
Trade-off I accepted
Governance first, to earn trust; start-from-scratch creation for marketers waits until after MVP
04
Replaced the template wizard with prompt-first setup and live preview
The alpha wizard hid the prompt at step two, with no preview. At ~10 minutes per generation, four small edits could cost 40 minutes; users called it “blind iteration” and “a black box”.
I moved template setup to prompt-first with live preview, grounded in a human-AI interaction audit run through a Claude Code skill I built.
Before: the alpha wizard. The prompt sat at step 2 of 2, with no preview of the page.
Default
After: prompt-first setup. The agent keeps to the template’s fixed visuals and follows the prompt’s guardrails to personalize the content.
What I didn’t do
Patch the wizard with more steps
Deferred, deliberately
Section targeting: research was inconclusive, so private beta tests it on real usage
05
Moved creation from a wizard to chat
A fixed wizard was too rigid for real campaigns: alpha users asked for required fields on templates, and sales reps wanted to add context, upload materials and ask questions mid-flow.
I moved creation into an AI chat assistant that builds one page per account from what the CRM already knows (discovery-call pain points, objections, pages visited), linked as variants to review and publish together.
What I didn’t do
Add more wizard steps for each new need
Trade-off I accepted
Less predictable flows to design and test, in exchange for one that fits how campaigns actually get built
06
Shipped a self-contained beta, and prepared the long-term vision across hubs
The long-term home for AI templates depended on open strategy questions across several teams. I chose a self-contained home for MVP (no dependencies, ships fast) over placing it inside landing pages, and kept the long-term placement as a named open question for private beta.
In parallel, research with the PM mapped three personas (marketers, sales reps, Marketing Ops) to four jobs across the funnel. I mapped each job to its entry points across Marketing Hub, Sales Hub, the CRM and Data Hub, and sequenced them: alpha, private beta, public beta, then cross-product alignment.
The direction: a proactive system that spots the signal in CRM data, recommends the action and generates the asset, fully tracked and attributed inside the platform.
The vision: one personalized motion across Marketing, Sales, CRM and Data, run by an agentic layer.
The vision prototype: from an insight on the Marketing home to a fully personalized campaign, through chat. One flow across 5 different teams, closing the loop across Marketing, Sales and CRM. Based on customer research, it proposes new interaction patterns for agentic, chat-based experiences.
What I didn’t do
Guess the long-term architecture to look decisive, or pitch the full vision as the beta scope
Trade-off I accepted
A likely move later and a reduced beta scope, in exchange for an MVP that shipped on time; the full motion moved to the teams that own those surfaces
Beyond the triad
UX leadership
I set the agenda for the review: a coherence check at 50–70% fidelity, with three questions for the room (does the end-to-end flow hold, do the AI interactions make sense, are we on the right surfaces). Feedback went to structure, not pixels.
Receiving teams
I wrote the handover document the receiving teams now work from: the decisions, the rationale and the open questions, so the work carried on without starting over.
My triad
I used the alpha ratings and the two-persona split to sequence the roadmap: governance for Ops first, speed for marketers next.
What happened
4–4.5 / 5
Content quality on real campaigns: the riskiest assumption, retired
Private beta
Funded on that evidence, live Sep 2026 as targeted pages
[Beta result]
[Private beta metric, to add when available]
Shipped. Targeted pages: “individual landing pages with unique URL links generated from a reusable template” (public docs).
Customer pull. Customers said they’d use it the next day; one wanted to show it to their manager immediately.
The gap I scoped for beta. Pages were ready for broad campaigns, not yet for the highest-value accounts. I scoped private beta to test the three things that decide whether this scales: quality at volume, section targeting, and whether non-technical users can write templates.
The vision carries on. That vision is what the receiving teams now build on.
What I’d do differently
Test the interaction model alongside quality. A clickable prompt-first prototype next to the quality test would have caught the 40-minute edit loop a cycle sooner.
Put a revenue number on the bet earlier. I framed the stakes as accounts lost to competitors, but never sized them in pipeline. With that number, the beta’s success bar would have been set in revenue, not ratings.