Case 01 · Agents
Ramp Creative
Capacity wasn’t the problem. Ramp’s creative team was drowning in requests. The fix wasn’t more designers. It was giving everyone else the brand.
- Role
- Senior Staff Design Engineer
- Year
- 2026
- Built with
- Slack, Inspect, HTML templates, DTCG tokens, Remotion
- Status
- Launched August 2026
01The ask
A deck that’s ready before you ask.
Early on, I sat down with someone in sales. I was there to understand his problems.
He described one I hadn’t heard before. An agent that reads his calendar, sees a pitch coming up, and has a custom deck ready before he asks. On brand, every time.
I was stoked. It was genuinely difficult. It wasn’t impossible.
That meeting reframed what I’d been seeing. People at Ramp weren’t sitting around waiting for design. They were ambitious, and they were already making their own decks and graphics in Google Docs and Claude. It all just came out off-brand.
So the real question wasn’t how to push more work through the creative team. It was this: how do you give ambitious people the brand without making them wait for a designer?
02Before
Every request went through a human filter.
Requests landed in #help-creative-studio. Slide decks, whitepapers, executive business reviews, launch graphics. The studio got more than 375 of them in six months.
Each one followed the same path. A producer gathered details and wrote a brief. A designer dug up old files and built the piece by hand. The requester waited.
One deck. Three people. Days.
Sales didn’t have the licensed typeface or a clean logo, so when they couldn’t wait, they made their own. The designers weren’t much better off. Production design is tedious, and their time went to execution instead of ideas.
Three people, one deck.
Every step of a routine request waited on a person.
BeforeOne deck, three people, days of waiting.
- a person does the work
- waits or reviews
- hands off
03The insight
Capacity vs. capability.
Everyone thought the problem was capacity. It was actually capability.
The obvious fix for too many requests is more designers. But hiring wouldn’t change who could make on-brand work. Two capabilities were missing:
- Requesters didn’t have the brand. No typeface, no templates, no way to get it right on their own.
- The brand lived in designers’ heads. Which layout to use, how much text fits, what to leave out. None of it was written down, so nobody else could use it.
As Paul Jun puts it, about half of brand work is concept. The other half is execution. So I split requests the same way: judgment vs. execution. Judgment calls go to people. Routine work goes to an agent that can only make on-brand output.
We called it Q, after the character in the Bond films who hands the agent the tools to do their best work. It later became the Ramp Creative Agent. The original name stuck with me because it describes the job. Q doesn’t go on the mission. He makes sure whoever does is ready for it.
Judgment goes to people. Routine work goes to the agent.
- Request
I need a deck for the Northwind renewal.
- The agent asks
Who is it for?
Their CFO and finance team.
What should it do?
Show what they saved this year and make the case to renew.
When is it due?
Thursday.
- It looks it up
- NotionAccount notes and this year’s savings
- #sales-northwindLast quarter’s review deck
- TemplatesBusiness review deck: a match
- Audience
- Northwind’s CFO and finance team
- Goal
- Show this year’s savings and make the case to renew
- Must include
- Savings to date, renewal terms
- Due
- Thursday
- Format
- Editable deck and PDF
It fits the business review template, so the agent builds it inside the brand rules.
04Slack
Meet people in Slack, not in a new tool.
The standard answer here is a template library: a web interface where people type their own content into templates.
I didn’t build that. It’s the old way of working. It asks people to learn a new tool and change how they work.
People already asked for creative help in Slack, so that’s where the agent lives. Requesters type in the same channel they always have. The agent asks the producer’s questions, pulls context from Notion, writes an internal brief and builds the deck from approved layouts.
What comes back is an editable HTML deck or a PDF. People can comment slide by slide and change any of the words. They can’t change type, colour or layout. When the agent isn’t sure, it makes an honest first attempt and tags the right producer.
New kinds of requests don’t break the system. They become new templates.
Same steps as Fig. 1
One message in Slack. People only where judgment matters.
In SlackOne message, and people only where they’re needed.
- a person does the work
- the agent does it
- waits or reviews
- brand rules applied
- hands off
Try it. Break a rule.
Change any word. The type, colour and layout hold.
05Designers
Write down what designers know.
The hardest part wasn’t code. It was getting designers’ judgment out of their heads and into the system.
Designers annotated every Figma layout: when to use it, what it holds, when not to use it. Brand rules became values the system enforces: design tokens, a brand file, skill files, and character and line limits for every slot. Motion timing came straight from a motion designer’s Lottie file. I built the renderer that reads all of it.
I half expected pushback. Designers have every right to be nervous about AI. Instead, they were excited, and I watched them go through the whole process with real willingness and conviction.
I think they felt connected to the output because the decisions were theirs.
The agent doesn’t have taste. It runs on theirs.
Everything still requires them.
This is where the taste lives.
Every step hides a decision a designer used to make in their head.
- 01AskWho is it for? What should it do? When is it due?The agent’s first questions, taken from the producers
- 02BriefWhat’s the real request behind “I need a deck”?A creative brief the agent writes and keeps
- 03StructureWhich layout tells this story? Does it need three images?Designers’ usage notes on every layout
- 04FitWill this headline fit? Is it too many lines?Character and line limits on every slot
- 05StyleRight weight, right logo, right colour?Design tokens and the brand file
- 06MotionHow should these names come in?Timing from a motion designer’s Lottie file
- 07CheckIs this on brand?Fit checks and an image-model review
- 08EscalateShould a person look at this?A tag for the right producer
A layout is a box of parts until someone writes down when to use it.
- Use when
- Three things belong together: three customers, three steps, three screens.
- Don’t use when
- You have two images. Take the two-up; don’t go looking for a third.
- Holds
- A headline up to six words, and one-line captions.
- Use when
- One number carries the whole point.
- Don’t use when
- The number needs a footnote to make sense.
- Holds
- Up to six characters and a four-word label.
06Onboarding
Onboard it like a human.
My first build in March was a Node and Python Slack app running ImageMagick. I also tried a markdown-to-slides app. Most of my time went to plumbing: the chat service, the database, memory, image manipulation. It was a lot for one person.
Then Inspect, Ramp’s internal tool for running agents, started hosting Slack agents with custom skills. I dropped the plumbing and put my time into the brand.
That changed how I think about architecture. It became less about coding best practices and more about taste and communication. You step back and trust the AI to do the work.
But trust doesn’t mean no rules. I gave the agent what I’d give a new production designer: tools, references, guidelines and a delivery pipeline. I just did what I’d do for a human.
It wasn’t perfect. Early ads came out ugly because we hadn’t shown it any good ads yet. The fix wasn’t code. Designers made top-quality examples, and the output followed.
The agent got what a new production designer gets.
- 01ToolsWhat to make the work with.Skills, and a renderer that reads the brand
- 02ReferencesWhat good looks like.Top-quality examples, made by the designers
- 03GuidelinesHow the brand works.Design tokens, the brand file, limits on every slot
- 04DeliveryHow finished work gets handed over.An editable HTML deck or a PDF, back in Slack
07OnRamp
Proof at conference scale.
The first big test was OnRamp, Ramp’s 2026 conference. Every speaker got an editable deck, a PDF, a presenter package and an eight-second title video. The videos were built in Remotion from one motion designer’s example.
Talks had one to five speakers. Of 38 talk titles, most were too long for two lines. The easy move was to make speakers cut their titles. Instead, titles wrap and the animation staggers, so every speaker kept the title they wrote.
One motion example, any title, eight seconds.
08Impact
Nobody else got involved.
The UK launch ran itself. A product marketing manager needed launch images with dollar signs swapped for pound signs. They asked the agent, and the images came back with layout and numbers intact. They solved it alone, then told others to use it. The launch team made and localized more than 80% of UK launch assets themselves.
A deck costs about a dollar. Paul Jun estimates an hour of a designer’s time at $100. A routine deck from the agent costs about $1.
It grew past decks. It launched in August 2026 with decks. It now makes whitepapers, one-pagers, print, product graphics and image edits.
Other teams picked it up. Requests ran in a public channel. Ramp’s applied AI team saw it working and adopted the system for their own work.
Designers got their time back. They now make the top-quality examples the agent learns from, and turn new kinds of requests into new templates.
One message, a launch’s worth of localized images.
What changed
About a dollar a deck.
Against $100 an hour.Fig 10.A · Cost of one deck
Fig 10.B · Who made the UK launch assets
09Lessons
What I’d take with me.
Map, find, fit. Map the process from request to delivery. Find the friction. Fit the simplest solution into the process people already have.
Meet people where they are. Nobody should have to change how they work to get help.
Write down expert knowledge early. The designers’ notes are the most valuable part of this system. More valuable than the code.
What I’d do differently: ship the smallest version first, watch how people use it, then add guardrails. I spent too much of my early time on plumbing before anyone had touched the thing.
Five months from first build to launch.
- Jan 2026Joined Ramp
- Mar 2026First build
- Mar 2026Building it aloneNode, Python, ImageMagick
- Jun 2026Burak and Mary join
- Jun 2026Templates, tokens, rendererwith Burak and Mary
- Aug 2026Launched
- Aug 2026Running and growingtemplates, fixes, OnRamp
10Why it matters
Production can be automated. Creativity can’t.
That sales rep wanted a deck that was ready before he asked. On brand, every time.
That’s what Q was for. Not to replace the people who make the brand, but to hand everyone else the tools to use it well.
The pound signs were the easy part. The designers defined what on brand means, and the system makes sure every request starts there.
Production can be automated, to everyone’s benefit. Creativity is still a human skill.
- Brand and creative lead
- Paul Jun
- Brand operations
- Emerline Ji, Staff Brand Operations
- Web engineering
- Burak Aslan
- Mary Jiang
- Design
- Mint Zhai, Kyla Arsadjaja, Claire Chen, Coco Yuan
- Quality bar
- Mark Weaver, Aodan Reddy
- Support
- Diego Zaks, VP of Design