---
title: "Embedded agent fleet case study · Derra"
description: "How a 100+ person BI and performance marketing agency connected creative production, performance optimisation, analytics dashboards and project management."
url: https://derra.io/case-studies/embedded-agent-fleet/
---

CASE STUDY 001

# Rolled out to all 8 clients, the automations freed 71% of the team's delivery time.

A 100+ person BI and performance marketing agency. We built the automations with its largest client, across creative production, performance optimisation, analytics dashboards and project management. Then we rolled them out to every client.

71%

Delivery time freed, all clients

97,000

Hours a year, 57 full-time people

$35M

Ad spend saved a year

THE AGENTS AT WORK

## From brief to platform-ready assets, with two human decisions.

Slack · #campaign-creative

Channels

\# campaign-creative

\# pm-updates

\# performance

\# dashboards

**Creative pipeline** AGENT 9:12 AM

Brief received with 12 product images. Copy drafted for 4 markets and 36 base assets produced. Ready for review.

DL

**Design lead** 9:40 AM

Approved. Two edits made on the hero banner.

**Creative pipeline** AGENT 10:05 AM

QA passed on all 36 assets: brand, copy, sizing and text legibility checked. Uploaded to Meta as paused drafts.

QA passed · 36/36

CM

**Campaign manager** 10:21 AM

Released.

Agent pauses. Creative is never released without a person.

Illustrative reconstruction, sample data.

THE CLIENT

## A 100+ person agency running repeat promotion cycles.

The agency works across strategy, performance, creative, data, project management and client operations, and runs its own analytics product with bespoke analytics dashboards. Its main client runs a repeating promotion cycle, so briefs, creative and performance follow a consistent pattern.

### ClickUp

Job boards, tasks and handoffs the PM agents create and track.

### Slack

Where people work with the agents, review output and approve.

### Notion

Briefs, processes and company knowledge the agents read.

### Figma

Creative templates the pipeline builds assets from.

### Frame.io

Creative review and feedback.

### BigQuery

Performance data for the optimisation loop and dashboards.

### Meta

Approved ads uploaded as paused drafts through the Meta Marketing API, ready for the campaign manager to release.

THE ENGAGEMENT

## Four workflows in scope.

### PM coordination

8 campaigns a year, each with a job board, handoffs and follow-ups.

### Creative production

5,200 assets a year from 8 briefs.

### Performance optimisation

8 campaigns and $25M in ad spend a year.

### Analytics dashboards

200 bespoke dashboards a year for the agency's analytics product.

Figures from here to the results are for the case-study client, the agency's largest. The rollout to all 8 clients follows.

THE AUDIT · WEEKS 1–4

## What the audit found.

We spent four weeks inside the operation, following real work across departments and interviewing the people doing it.

01

### Work crossed too many inboxes

One campaign request passed through 5 people and 5 handoffs before it shipped.

02

### Specialists became queues

Campaign managers waited 72 hours on average for analysts to answer repeatable questions.

03

### Approval did not mean delivery

Approved creative still needed an hour of manual preparation, QA and upload per asset.

04

### Dashboards needed developers

Each dashboard went from analyst brief to developer, through QA and back to the developer for revisions, over several rounds. It took 2 weeks and 8 developer hours.

THE REDESIGN

## Four connected parts. People decide where it counts.

We did not replace the tools or the team. PM agents coordinate the work. An event-driven creative pipeline produces the assets. A performance loop feeds what worked back into the pipeline. Analysts ship analytics dashboards without waiting on developers.

Most of the system is event-driven automation and headless agents that run on triggers. Chat only comes in when a person gives feedback.

PM agents

ClickUp · Slack · Notion

Creative pipeline

Figma · Frame.io

Brief and assets in

Copy from the brief

Base assets produced

Review and edits

QA: agent + image models

Uploaded to Meta

Review and release

Performance loop

BigQuery · A/B tests

Analytics dashboards

Analyst → agent → approval → customer

Person decides

Coordination

Performance feedback

Event-driven: runs on triggers, nobody prompts it

Designer reviews and adjusts assets

Campaign manager reviews ads in Meta and releases

Chat only here: designer feedback goes to the agent

FIG. 1 · ONE CREATIVE ASSET, BEFORE AND AFTER

Before

15 steps, 7 people · 2 weeks

After

Brief in

Copy

Assets

Review

QA

Upload

Release

7 steps, 2 human decisions · half a day

Manual step

Event-driven automation

Headless agent

Person decides

Chat feedback

5 steps run on their own. People review twice.

FIG. 2 · ONE ANALYTICS DASHBOARD, BEFORE AND AFTER

## One analyst and one agent replace a relay of handoffs.

Before

2 weeks · 8 developer hours · several rounds

Analyst

writes the brief

Developer: backend

queries, data API

Developer: frontend

QA

Platform

revisions

Now

1 day · 0 developer hours · about 1 analyst hour

Analyst

Agent

Analysis and revisions happen here · Slack

BigQuery

pulls and samples data

Backend

views, queries, API

Frontend

dashboard components from templates

Data definitions

metrics and rules analysts maintain

GitHub

reviewed and merged

Analytics platform

Analyst approves

Customer's account

Person in a handoff

Handoff

Agent

Person decides

BEFORE AND AFTER

| Metric                                | Before         | After        | Change      |
| ------------------------------------- | -------------- | ------------ | ----------- |
| Turnaround per asset                  | 2 weeks        | Half a day   | 96% shorter |
| Person-hours per asset                | 3 hours        | 6 minutes    | 97% fewer   |
| PM hours a week on coordination       | 32             | 6            | 81% fewer   |
| Dashboard turnaround                  | 2 weeks        | 1 day        | 93% shorter |
| Developer hours per dashboard         | 8              | 0            | None needed |
| Analyst hands-on time per dashboard   | Several rounds | About 1 hour | One loop    |
| Creative variants tested per campaign | 3              | 10           | 3.3x more   |
| Customer acquisition cost             | $75            | $53          | 29% lower   |

TIMELINE

## Four weeks of audit, then build in stages.

| WHEN        | FOCUS                | MILESTONE                                                                             |
| ----------- | -------------------- | ------------------------------------------------------------------------------------- |
| Weeks 1–4   | Audit                | Map the work, interview the teams and set the baseline for every figure on this page. |
| Weeks 5–6   | PM agents            | PM agents and shared company knowledge go live in Slack and ClickUp.                  |
| Weeks 5–8   | Analytics dashboards | Built in parallel. Analysts ship dashboards through the agent.                        |
| Weeks 7–10  | Creative pipeline    | Built, run in shadow next to the creative team, then live.                            |
| Weeks 11–14 | Performance loop     | Connected to performance data in BigQuery and feeding the pipeline.                   |

Dashboards ran in parallel and shipped first.

WHAT THE AGENTS DO

### PM agents

- Turn a new campaign into a ClickUp job board with owners and dates.
- Hand work between teams with the context attached.
- Monitor progress and chase follow-ups in Slack.
- Flag anything at risk to the PM.

### Creative pipeline

- Write copy from the brief.
- Produce base assets from the copy and provided images.
- Send them for human review and edits.
- Run QA with a headless agent and image analysis models.
- Upload passed assets for the campaign manager to release.

### Performance loop

- Monitor creative performance in BigQuery.
- Identify what performed best.
- Produce new variants from the winners.
- A/B test them in the campaign and carry the learnings into the next one.

### Analytics dashboards

- The analyst works through the analysis with the agent in Slack.
- The agent pulls and samples data from BigQuery and writes the frontend and backend code.
- Code goes through GitHub and ships to the platform.
- The analyst checks it and one approval ships it. No developer handoff.

RESULTS

## The same team, more capacity.

Asset turnaround fell from 2 weeks to half a day, and the hours to produce one asset from 3 hours to 6 minutes. PMs got back 26 hours a week each. Dashboards ship in 1 day instead of 2 weeks, with no developer time. Optimised creative cut customer acquisition cost by 29%, worth about $7.3M a year in ad spend, or around 138,000 extra customers for the same budget. That one client released 20,200 hours a year, 15% of the team's delivery time. Figures combine audit measurements and team estimates.

ROLLED OUT TO EVERY CLIENT

## Built on the largest client. Running on all 8.

The case-study client is the agency's largest. The smallest is about a fifth of its size. Taking the other clients as spread evenly between the two, all 8 together equal about 4.8 clients the size of the case study.

Case-study client

Smallest, about 20%

8 clients ≈ 4.8 case-study clients

| Step                              | Value   | Note                                            |
| --------------------------------- | ------- | ----------------------------------------------- |
| Hours released, case-study client | 20,200  | a year                                          |
| × Client-equivalents              | 4.8     | 8 clients, average 60% of the largest           |
| = Hours released, all clients     | 97,000  | a year, the work of 57 full-time people         |
| ÷ Delivery hours a year           | 136,000 | 80 people × 1,700 hours                         |
| = Delivery time freed             | 71%     | the same team can carry about 3.5x the workload |

Agency-wide figures are estimates scaled from the case-study client.

WHERE THE VALUE COMES FROM

| Source                                   | Annual value        | How                                                                                                                    |
| ---------------------------------------- | ------------------- | ---------------------------------------------------------------------------------------------------------------------- |
| Creative production time                 | $679k               | 5,200 assets × 2.9 hours saved each at $45 an hour                                                                     |
| PM coordination time                     | $187k               | 3 PMs × 26 hours a week × 48 weeks at $50 an hour                                                                      |
| Developer time on dashboards             | $102k               | 200 dashboards × 8 developer hours removed at $70, less 1 analyst hour each at $50                                     |
| Ad spend saved by optimisation           | $7.3M               | Customer acquisition cost down from $75 to $53 on $25M of ad spend, or about 138,000 extra customers at the same spend |
| Hours released                           | 20,200 hours a year | Creative 15,100 + PM 3,700 + developer 1,400 (rounded)                                                                 |
| Share of delivery time freed, one client | 15%                 | 20,200 hours released ÷ 136,000 delivery hours a year (80 people × 1,700 hours)                                        |
| Ad spend saved, all clients              | $35M                | $7.3M × 4.8 client-equivalents                                                                                         |
| Delivery time freed, all clients         | 71%                 | 97,000 hours released ÷ 136,000 delivery hours a year                                                                  |

WHAT STAYS WITH PEOPLE

## Agents prepare. People decide.

Agents never approve creative, release a campaign or ship a dashboard on their own. Designers review and edit every asset. Campaign managers release every campaign. Analysts own the data definitions and check every dashboard before one approval delivers it.

WHAT THIS DEMONSTRATES

## Forward-deployed engineering compounds.

The first value comes from solving recurring work. The larger value comes from instrumenting that work and creating a durable picture of how the operation behaves.

As the system sees more decisions, exceptions, handovers and outcomes, it becomes possible to improve the automation, redesign the workflow and identify changes to the business itself.

[BOOK A CALL](https://cal.com/derra/intro-call) [EMAIL US](mailto:hello@derra.io "hello@derra.io")

---

Book a call: https://cal.com/derra/intro-call
Email: hello@derra.io
Site: https://derra.io
