An AI ad workflow for agencies — workspaces, seats and exports

An agency has constraints a solo creator does not. Several clients, several people, one invoice, and a review loop that has to stay private until the client says yes. The creative part of AI video is the easy half; the half that decides whether it works in a shop is billing, access and where the drafts live. This is the setup we would run, with the arithmetic at the end. The short version is on the agencies use case.

One workspace, one wallet, real seats

Credits live on the workspace, not on individual accounts, so a team spends from one balance instead of five personal cards. Members are admins or members, invitations go out by email, and seats and API keys come with the plan:

• Hobby $15 for 1,800 credits — 2 seats, 10 API keys
• Creator $49 for 6,000 credits — 5 seats, 20 keys, every video engine, 2× rate limits, priority video queue
• Studio $149 for 20,000 credits — 15 seats, 50 keys, 4× limits, clips up to 30 seconds
• Team $249 for 36,000 credits — 25 seats, 100 keys
• Agency $499 for 80,000 credits — 50 seats, 200 keys

Annual billing takes 20 percent off, monthly credits roll over forever, and credits never expire — which matters more than the headline number when half your months are pitch months. Anything larger goes through Enterprise at [email protected]. The pricing page has the full comparison.

Ads mode and one Element per client

Ads mode in the video tool is the same render pipeline with the brief pre-shaped. Templates such as "UGC testimonial" cover the formats that actually run — vertical for TikTok and Reels, square for feed, wide for YouTube and web — and the aspect ratio is a setting like any other.

The client's product enters as an Element: a reusable object with a description and reference photos, @mentionable in any prompt, up to 50 per account. Build one per product, plus one style Element per brand carrying the palette and tone of voice, and a new campaign becomes an @mention and a sentence instead of a re-briefing. It is also how a junior's prompt inherits the art director's rules without a meeting.

Here is a complete Ads-mode prompt for a fictional client, a cold-brew brand:

A woman in her thirties pulls a bottle of @coldbrew from a fridge in a bright apartment kitchen, condensation beading on the glass, handheld camera following the bottle up to the counter, hard morning light through a window with a soft bounce on her face, candy pop palette, dynamic tempo, commercial look, vertical frame with headroom for captions.

The director's rack — film type, era, tempo, camera gear, lens, aperture, palette, lighting — folds into the prompt server-side, so the brand look is a saved set of choices rather than a paragraph everyone copies slightly wrong. Enhance rewrites a rough draft into a director-grade prompt for free, which is a useful equalizer when the person writing the brief is not the person who shoots.

Review happens in exports, not in public

Say this part out loud to the team: client work never goes to the community. Publishing there is a public act — title, cover, likes, comments, views — and it is for your own reel, not for a campaign under embargo.

Client review runs on the private library and on exports. Every render is saved automatically with its full recipe, visible to the workspace, so the account manager can see the takes without asking the editor for a link. Approved cuts leave as downloads, into the deck, the shared drive, or the edit. If you cut in Premiere Pro or After Effects, the eroq panel generates from inside the app and imports renders straight into the open project, saved under ~/Documents/eroq; it is distributed unsigned today, so it installs in developer mode.

The corollary is a good habit anyway: one review link, one round of notes, then a Remix of the exact recipe rather than a fresh prompt that quietly changes three other things.

The API for the repetitive third

Some of the work is not creative at all — forty variants of the same spot with a different end card, or the same three shots in three formats. That part belongs in a script. The REST API is OpenAI-shaped at https://eroq.ai/v1 with a Bearer key: images/generations, videos/generations as an async job you post and then receive on a signed webhook, audio/speech, and films for whole storyboards, where one render call shoots every scene. Rate limits are per key — 60 chat requests a minute, 20 images, 6 video or speech — multiplied by your plan.

Per-key limits are the reason to issue one key per client or per campaign: a runaway job on one account does not starve the others, and usage per key is the closest thing to a per-client cost line you will get without an accountant. The docs cover the endpoints, and there is a CLI and an MCP server if the automation lives in a coding agent instead of a cron job.

What a campaign costs

A six-short campaign, three 5-second clips per short on Motion One at 100 credits per 5 seconds, plus one cover still at 10 credits:

• Per short — 3 × 100 + 10 = 310 credits
• Six shorts — 6 × 310 = 1,860 credits

That is about $18.60 at the entry pack rate of $10 for 1,000 credits. Shoot the same campaign at 1080p on Kling 2.5 Turbo (120 credits per 5 seconds) and it is 6 × (3 × 120 + 10) = 2,220 credits. On the Agency plan ($499 for 80,000 credits) that is 80,000 ÷ 2,220 ≈ 36 campaigns a month; on Team ($249 for 36,000) it is 36,000 ÷ 2,220 ≈ 16.

Add takes to taste — a second take on the hook shot of each short is 6 × 120 = 720 credits on Kling — and remember that failed renders refund themselves and blocked requests are never charged, so the estimate does not need a padding line for accidents.

FAQ

Can each client have their own budget?

Credits are pooled per workspace, so the cleanest split is one workspace per client when budgets must not mix, and per-key usage inside a shared workspace when they only need to be reported. Seats scale with the plan, from 5 on Creator to 50 on Agency.

Can we work on adult brands?

Within the acceptable use policy, and adults only. Motion One is the uncensored video engine and is available on every account; the premium engines are moderated upstream and are SFW only. Real identifiable people without consent are banned outright on every engine, which rules out the deepfake requests you will occasionally be asked for.

How do we hand off to the client's editor?

Download the renders, or import them straight into Premiere Pro or After Effects with the panel. Everything stays in the workspace library with its recipe, so a change request two weeks later is a Remix of the approved take rather than a rebuild.

Setting up a shop? Get an API key, create the workspace first, and make each client's product an Element before anyone writes a prompt.