Agentic Quote Creation

‍

Helping HubSpot sales reps close deals with fast and accurate quotes.

‍

HubSpot sellers lose deals when quotes take too long. I led design for an AI agent and Breeze skill that drafts the quote from the deal, guides reps through their company's rules and approvals, and leaves publishing and sending in the rep's hands.

Industry recognition

This work was a key feature of HubSpot’s CPQ offering, awarded Leader in The Forrester Wave™ for Configure, Price and Quote Solutions, Q4 2026 (opens in a new tab).

company
HubSpot
timeline
90 day sprint
Role
Senior Product Designer, design lead
Tools
Claude Code, Figma, Storybook

Outcomes

  • Time-to-quote improved by more than 50% — compared with the current quote editor — in testing with beta customers
  • Beta customers reported increased confidence navigating complex business pricing rules when creating their quote, compared with the current quote editor
  • AI chat + canvas artifact pattern contributed to HubSpot's design system

Framing design objectives

01Customer | Task success
Objective

Increase customer task success

Outcome

Reduce the time required to create a quote

02Customer | Satisfaction
Objective

Increase customer satisfaction

Outcome

Reduce negative customer feedback

03Business | LTV
Objective

Increase product LTV through an agentic platform

Outcome

Increase the % of deals converted to quotes

The problem

4 days
from deal-ready to quote sent

The quote is the time-critical last step to close a deal. Reps build it from the deal’s products, pricing and terms, while admins set the templates, rules and approvals every quote must pass. When that takes too long, buyer interest fades.

Deal to signature
Deal to signature: Deal, Quote, Publish, Share, Sign
↑ Deals stall between building the quote and publishing it
What reps told us
01Reps re-enter details already confirmed with the buyer.
02Pricing and discount errors are caught late, in a maze of compliance requirements.
03Quoting feels disconnected from the rest of the CRM.

Reframing the problem

From
Deals die when quotes get delayed

Two people shape every quote, and they want different things from it.

Admin | sets up quoting

When I set up quoting, I want every quote to follow our templates, rules and approvals, so pricing stays accurate and compliant and scales.

Problem

Compliance lives in settings reps can’t see, so errors surface late in approvals.

Seller | closes the deal

When my deal is ready, I want a complete, compliant quote fast, so I can send it while the buyer is still interested.

Problem

Reps re-enter details already in the deal and navigate a maze of compliance requirements.

To
Quotes are ready for reps to finalize and share

We're all builders — design's superpower is taste

The old linear handoff

Customers reacted to unwieldy, static Figma prototypes. Gaps surfaced late, during QA, and engineering had to translate designs into design system front-end code.

Research and requirements→Static mocks in Figma→Customer and team review→Spec and engineering handoff→QA after it’s built→Submit QA fixes…
Design-led front end
  1. 01I prototype in code and review with customers and the team
  2. 02The front-end engineer completes the back-end wiring
  3. 03Ship

Production-fidelity prototypes settle debates with customers mid-project, not after launch.

Engineer-led front end
  1. 01The front-end engineer prototypes, customers and the team review it
  2. 02I polish and refine the UI in the repo and submit a PR
  3. 03Ship

Engineering can spike and prototype innovative front end where back-end connections are critical.

Current toolsClaude CodeConductorSidekick agentFigJamStorybookNotebookLMFigmaLoom

Designing for agents and AI

A tighter loop with customers and stakeholders: fuller context going in and production-fidelity prototypes coming out for faster, more productive feedback.

01

Define and align

Claude CodeFigJamNotebookLM
  • Synthesize research, call notes and data in NotebookLM for fuller context
  • Map the end-to-end flow with stakeholders in FigJam
  • Align on the problem, vision, scope and measurable goals
02

Explore and build

Claude CodeFigmaStorybookFigJamNotebookLM
  • Explore divergent directions in Figma and with AI
  • Prototype in Claude Code with HubSpot design system components from Storybook, so it looks, moves and behaves like the product
  • Craft and test UI for AI and agentic skills, manually and with AI
03

Share and learn

LoomClaude Code
  • Put working prototypes in front of customers and stakeholders
  • Share async walkthroughs in Loom
  • Bring what we learn back into exploration
↺ Feedback loop · design crits, stakeholder reviews, customer sessions

Shaping the agent skill

Engineering first owned tuning the quote skill. Their tuning was technically sound, but it missed UX patterns like progressive disclosure and the end-to-end flow, and the Breeze conversation felt dense and disjointed.

Debate real moments, not opinions.

I audited the skill's conversations in the agent platform, then brought working prototypes and real agent conversations to early design reviews. That let us point at the exact moments that broke a rep’s trust, and those reviews became a shared backlog of skill issues.

Design now owns how the skill talks, paces the work and asks for input. Engineering owns the tuning, and the shared backlog keeps us in sync.

HubSpot quote agent skill
Trigger“Create quote” from a deal card or in Breeze chat
ContextDeal products, contract length, account tier, conversations, quote template
ToolsTemplate verification, build line items, run Quote Rules, run Quote Approvals, publish, draft email
PermissionsDrafts on its own; can’t publish past a failed Quote Rule or Quote Approval
CheckpointsRep confirms template, resolves rules, approves publish and send
HubSpot agent builder showing the Quote Agent

Agent design

01Transparency
Principle

Proactively do the work, show the reasoning.

In practice

The agent says what it did and why, so reps can check it.

02Guardrails
Principle

Compliance applies to agents, too.

In practice

Every draft runs against the same Quote Rules and Quote Approvals as a person.

03Human control
Principle

People commit.

In practice

Publishing and sharing are explicit, rep-owned steps.

Leading a unified experience across many teams

Agentic quote creation sat at the center of five teams. Here's what we worked through with each.

Product
Our product triad set the vision for a fully automatic experience with as little friction as possible, then managed scope for a private beta and a public beta so we could learn with customers fast.
Engineering
Engineering first owned tuning the quote-building skill. Early prototype reviews surfaced missing UX patterns like progressive disclosure, and we built a shared backlog of skill issues.
Platform AI
We didn’t own the agent. I partnered with HubSpot’s platform agent team on how quoting fits together: the deal agent orchestrates, our quote-creation skill holds the workflow, and tools do the work. Together we set what the agent can read, do on its own, and must ask the rep.
Design system
I created a seamless end-to-end quote sharing interaction pattern that was contributed back to design system.
Beta customers
I brought both multiple template selection prototypes to customer sessions. Customers use only a few templates and could verify one from a thumbnail, settling a techinical constraint for AI quote template inference: Breeze recommends, the rep confirms.

Creating a quote, start to finish

Impact

>50%
faster time-to-quote vs. the current quote editor
8
beta customers
This (Breeze) is so much easier than trying to figure out all the menus. I don’t even know where to start in the new one (quote editor)…
Sales leader · HubSpot beta customer

Keys to success & Learnings

  • Tuning the AI skill is a design problem. How the agent reads a deal, which steps it proposes and where it stops to ask are as much a part of the experience as the screens. Design should own that work, not hand it off after the UI is done.
  • The details matter most when AI and agents act for the user. Some steps carry real consequences, like changing a discount or sending a quote to a customer. An agent shouldn’t take those steps without the seller’s say. I mapped where sellers needed to review and confirm, and made those confirmations a required part of the AI skill so sellers stay in control of their quote.
  • Test concepts early, especially ones generated with AI. Early prototypes missed how customers wanted to choose products from their HubSpot product library. Because we tested several concepts with customers early, the fix was a small pivot to a chat-based product picker that matched how they already thought about it.