Will Hunter · GTM Engineering

Rippling Paid Social Agent

1

Paid Social Agent

Creative was capping campaign velocity.

Why we started here

Sat with the domain experts on Meta and LinkedIn. Creative creation kept coming up as the constraint on campaign velocity.


A Figma canvas zoomed out, showing dozens of near-identical duplicated ad variants scattered across frames.

Brand built a template. Paid social hand-edited every variant: new copy, new image, new incentive, repeat.

8–12 hrs / week

An experiment won? They rebuilt it by hand for every other Rippling product.

2

Paid Social Agent

Before the agent: taking every ad apart.

The team already looked at performance. The work was manual, one experiment at a time, so the non-obvious patterns stayed hidden.

A real Rippling IT ad annotated with its six component elements: headline, incentive, body copy, background, format and concept.
Scoring

Every element gets scored on its own, across every ad we've run.

Lesson

Built the data layer before any agent work.

An agent is only as good as how easily it can find out what's true. The agent queries something simple. It never has to go exploring.

3

Paid Social Agent

Stop asking “did this ad win?”

Ask what's true across everything we've run.


A Rippling HR platform comparison ad with a free AirPods incentive badge in the lower left.
The same Rippling ad with the AirPods badge removed and the free gift line dropped from the body copy.

A lot of concepts we rated highly were being carried by the incentive.

Strip the incentive and the concept wasn't doing much on its own. A few held up unaided, and those are the ones we should keep experimenting with.

Not a criticism of anyone's judgement. It's a pattern that only shows up at that zoom level, and no single experiment would ever have surfaced it.

4

Paid Social Agent

Separate the plan from the pixels.

“This ad format did very well for our IT product. Can you create a version of it for core HR, spend, and our global product?”

That plus the ad itself is the entire input. A strategy problem and a rendering problem aren't the same problem.

Stage 1 · Plan

What's changing, what's held constant

Checked for reasoning, approved claims, and real evidence.

Human
approves

Stage 2 · Generate

Execute the approved plan

Checked for fidelity to that plan, not whether the art is good.

The proposed plan: a direction reading product-line reposition, IT ad template adapted to Core HR, Spend and Global, followed by three numbered variants each with its exact headline and three bullet lines.
Marketers align on the strategic direction before generation begins.
5

Paid Social Agent

Guardrails on the plan.

Before a pixel is rendered.

The claims field of a plan, listing eight approved claim IDs such as EXT-GLOBAL-F9FAE070 and EXT-FINANCEACCOUNTANT-EDE0A536.
Every line of copy traces back to an approved claim ID.
Claims

The check corrects rather than blocks.

The model drafts “save 95% of your workload” when the real number is 90. It doesn't kill the run. It points the planner at the right number and carries on.

Other guardrails

One experiment at a time

Only one thing varies.

Product-line completeness

Every leftover from the source product is replaced, neutralized, or kept on purpose.

Source grounding

Base assets exist. Doesn't invent copy, badges, or audience signals.

6

Paid Social Agent

Three methods that didn't work.

The failures are what pointed at the answer.

Attempt 1

Pure generation

The model reinvents the ad every time. Same prompt, wildly different results, nothing holding still between variants. You can't hand that to a brand team.

Attempt 2

Deterministic image creation and editing

Total control, and consistently sloppy output. It looked assembled rather than designed.

Attempt 3

Driving Figma through its API

Only part of what we needed was exposed, and what was exposed hit rate limits and ran too slowly for the volume we wanted.

7

Paid Social Agent

Editing beat generating.

The edit endpoint

Hand it the existing ad and a clearly defined prompt about what we're switching. Everything else stays identical.

Source IT ad: Meet IT that runs itself.
Original
Edited HR variant: Meet HR that runs itself, same layout and photography.
Edited Global variant: Meet Global that runs itself, same layout and photography.

Photoshop at scale.

Editing means you can trust that only the experiment changes. Layout and the legal footer stay constant.

8

Paid Social Agent

Is this asset good? Is it brand safe?

Did we execute the approved plan?

The old question

Subjective. No fixed reference. Fires constantly on things that are fine.

An HR variant of the ad beside its quality checks: font, layout, wardrobe and background unchanged all passed, product line coherence passed, ready to use.
Layout, type, and photography held. The product-line copy is what changed.
The new question

The plan is a contract. The source ad is a reference. Did the swap happen, and nothing else?

9

Paid Social Agent

Expanding outside the paid social team.

~30 people in marketing use it. ABM and the vertical teams are using it too.

~50

ads live, still ramping

Three account-based ads generated in one run, each naming its target company: Security Compass merging two teams, AlayaCare scaling fast after a fifty million dollar raise, and Absorb Software ready for global payroll.

Hyper-personalized ads for named accounts. Doing this by hand was never feasible.

10

Paid Social Agent

What's next.

Image creation is built. Building and trafficking is in progress.

Image creationBuilt 1 · Building & TraffickingIn progress 2 · Top of Funnel Health 3 · Bidding & Budget
1 · Building & trafficking · in progress

Campaign creation is still manual, and it caps throughput no matter how fast creative gets.

2 · Top of funnel health

Are the leads real and in-ICP, and do platform numbers match the warehouse?

3 · Bidding & budget

Move spend on pipeline signal, not a weekly review.

11