REAL AI WORKER DEMO · ELLĀTU
Watch one agency job go end to end.
Ollie moves from brief and Context to human review, real PAUSED campaign builds and verification.
English audio + captions · Use CC in the player
Real builds. Human approvals. Configured PAUSED.
Demo assumptions and reconstructed review emails are labeled in the film. Meta uses an authorized demo account. This is workflow proof, not campaign performance.
What this workflow demonstrates
- Inputs
- A demo launch brief, approved client Context and real product assets. Budgets are demo assumptions.
- Work performed
- Research, campaign planning, creative preparation and campaign objects in approved Google Ads and Meta demo accounts.
- Human decisions
- Commercial corrections and client approval before platform writes. Review emails are labeled reconstructions; client approval is a chat transcription.
- Verified output
- Campaign builds configured PAUSED, checked in the UI and saved API read-backs. No evidence of delivery, spend or performance improvement.
Full workflow transcript
Client brief
Most AI demos show you a model having ideas. This one shows a worker finishing a job. A US launch for a hair serum, from a client brief to paused campaigns in Google and Meta, with an agency in control the whole way. The brief and budgets are demo assumptions. The workspace, the assets and the campaign objects are real. Watch what a workplace lets an AI worker actually complete. One agency job: a new US launch for Ellātu’s Silky Touch Serum. The brief, audiences and budgets are demo assumptions. The references, workspace and campaign objects are real. This Gmail brief is staged; automatic email intake is not demonstrated. Our question is what work can be completed while the agency keeps control.
Workspace + Context
Here is DEMO 018, the workspace used for this job. The workspace matters because the worker needs the right client knowledge and access before it begins. This screen establishes the workspace identity. Saved Context describes the product, the US demo market, the premium visual direction and supported claims. It also establishes the review boundary: internal review before client communication, and approval before meaningful changes. This is the current Context used before the new-account build. Saved knowledge does not expand API permissions.
Meet Ollie
The worker you're watching is Ollie, an AI worker plugin running inside this Outloop workspace. Outloop gives Ollie the client Context, approved access, connected systems and approval rules it needs to actually finish the work.
Research + campaign plan
Research informs decisions rather than filling a report. Official product information sets the claims boundary. A dated competitor sample helps with positioning. Retrieved keyword estimates help group relevant searches, but they are not Ellātu performance. Hair-growth intent is deliberately excluded because search demand does not make a claim valid. Overlapping search estimates are not added together as unique demand. The resulting Search structure has two ad groups, ten keywords, fifty negatives and one responsive search ad. Meta uses one campaign, one ad set and one ad with the approved dark creative. The US audience and budgets are demo assumptions. USD and ILS are shown separately. Every delivery object is configured paused; a configured budget is not permission to spend. So far the worker has understood the client: the claims it may make, the market, and what to leave out. Now it moves from understanding to making, using the client's own files, not stock inspiration.
Assets + creative
The worker draws on approved Drive files: the actual wordmark, product photography and supporting references. These establish brand and product identity. Ingredient references support the composition. They do not replace the real product or turn inspiration into an authoritative asset. The accepted creative package keeps the product central, with restrained copy and a clear call to action. We can compare dark and ingredient-led variants and useful formats. These are generated campaign assets using the approved references. Showing a landscape asset here does not mean it was installed as a Google Display ad; the actual Google build is Search.
Human review
These next emails are a staged demo reconstruction of real approved decisions, not the original historical email exchange. Ollie presents the package for internal review before the client sees it. The Account Manager is asked for commercial judgment: how to treat past purchasers, and what planning information belongs in the client summary. The Account Manager supplies two rules: no discount-led messaging to past purchasers, and no forecast figures in client communication. The package and both rules are approved for client review. Ollie confirms the corrected package follows them. This internal approval does not authorize campaign writes. The current Context already contains these rules; these staged emails did not save them again. The preserved original Outloop receipt records revision 23 and both exact rules. This is historical receipt evidence, not a recording of the old settings screen. The current Context is revision 34. The lesson is persistent approved client knowledge, with the same permission boundary. Two decisions in that package were commercial, not compliance. How to treat past purchasers. Whether a forecast belongs in front of the client. Those are human calls. The Account Manager made them once, the worker corrected the package, and the rules stayed in the client's Context. That is what a workplace does that a chat window cannot.
Client approval
We now return to a real received email. It describes the new-account build and includes the scope document and approved creative attachments. The client receives a concise account of what is proposed, before the platform write. This email predates the staged review reconstruction; the film groups the steps to explain the workflow. Acting as the demo client, Adam approved this exact new-account scope, paused only. That approval occurred in chat. The on-screen presentation is a labeled transcription, not a Gmail reply. Internal approval and client approval are separate gates. Once the client approves, Ollie doesn't stop at a media plan or a recommendation. Through Outloop's approved access, it builds the campaign and pushes it directly into the ad accounts: campaign structure, ad groups, ad sets, keywords, negatives, targeting, budgets and creative, while keeping everything PAUSED for review.
Google Ads build
In Google Ads, the worker built the campaign, ad groups, keywords, negative list and responsive search ad, pushed the build into the account, and kept it PAUSED. The negative list is attached as an exclusion, not as permission to deliver. Then it read every object back through the API to verify the setup. This is a fresh build in the new account; no older ad was recreated to show retention.
Meta Ads build
On Meta, it created the campaign, ad set and ad with the approved dark creative inside an authorized demo ad account, not an Ellātu-owned account. Again configured PAUSED, and verified through the API. The read-back reported the ad's effective status as in process. That is not platform approval or delivery.
Verification + reporting
For the specified September 19 reporting period, Google returned numeric zero cost, impressions and clicks. Meta returned no reporting rows. No rows is not a measured zero, so we do not describe it that way. These are dated observations, not a promise about future account activity.
Completion + retained Context
The real completion email tells the client what was built and confirms it remains paused. Its disclosure about UI inspection was accurate when sent; the interface footage was captured later. The client gets a clear conclusion grounded in the actual objects, without an invented performance forecast. Finally, both commercial rules remain in Saved Context. The next job can use that approved knowledge. Human judgment persists, while approval and runtime permissions remain separate. That is the agency workflow: approved knowledge, useful work, human review, client approval, real paused execution and verification. AI workers need more than a model. They need a workplace: approved knowledge, the right access, human review, and real systems they can work in safely. That is Outloop. Start your free trial, or book an AI Workflow Review and we will walk through one of your own workflows.
Bring this to your own client work.
Start 14-day trialStart with one workflow.
Bring one real client workflow. We’ll map the systems, access, files, approvals and execution path needed to turn it into an AI-worker workflow.
Set up your first workspaceAI Workflow Review