Steve Waddington ONET.COM.AU

PROJECTS / AI × CREATIVE

AI Agents for a Small Video Production Studio

PERSONAL AND TEAM AI AGENTS FOR A SMALL CREATIVE STUDIO: AN ACCEPTED PROOF OF CONCEPT

The client asked not to be identified and is referred to throughout as "the studio".

Project status

Proof of concept accepted at the one-month review. The studio is now considering a move to Claude Team or Enterprise plans.

Summary

The studio is a small creative business producing custom video content for business clients, run by a three-person creative team. Over two weeks on site, followed by a month of off-site support, I set up two layers of AI assistance: a personal assistant agent for each team member, and a shared team agent that handles production logistics. Alongside them, the studio began using AI image and video generation in place of some live shoots.

Measured the way we agreed at the start of the project, production output rose by a third and marketing output by 52%. The studio's own count of finished deliverables shows a bigger result, up from four a week to six over the same period. The studio had hoped for more, expects further gains, and accepted the result as a successful proof of concept.

Business problem

Competitors' AI-produced marketing was drowning out the studio's message. To stay competitive the studio needed to keep its marketing and promotion output up, and it couldn't do that by taking hours away from client production.

In a studio this size the two aren't really separate. Much of the marketing material comes from live shoots: the studio shoots a short run for a client, then uses it to pitch the concept for full production, and it also stages mock-up shoots for its own advertising. Every live shoot costs time and money whether it ends up as a client deliverable or a pitch. The studio's longer-term aim is to replace as much of that live production as it sensibly can with AI generation. It picked marketing as the starting point because it was the biggest problem and the easiest to solve.

Approach

The work had two parts.

Personal assistant agents. Each team member got their own AI assistant, running in the Claude desktop app on an individual Claude Pro subscription, plus assistants set up for particular tasks. The aim was to take routine work off each person and give that time back to production. It was also a way for each person to get used to working with an agent on their own terms before the team-level system arrived.

The assistants were built from a common template, then shaped around each person in the one-on-one meetings: the assistant's name and character, how blunt it should be, and the jobs it would do. The template sets firm limits. An assistant can read its owner's email but never send, delete or file it, and it can manage its owner's own calendar when asked but not invite anyone. It may fill in a web form, but stops and shows what it has entered before anything is submitted. It never makes payments, enters credentials or publishes anything. It treats everything it reads in emails, attachments and web pages as material to summarise, never as instructions to follow, which is a basic defence for any agent with access to an inbox. Each assistant keeps its own memory files (contacts, key dates, to-dos, things it's waiting on, where things are kept) and updates them when it's corrected, so the same mistake doesn't keep coming back.

Day to day, each assistant runs a scheduled morning briefing covering the day's calendar, upcoming deadlines, email that needs attention and anything still outstanding, and its owner can email it notes from away from the desk to file. The assistants also track the team's schedules, so everyone knows what everyone else is doing that day and that week. The template is available to download below.

A team production agent. A shared agent, running on a Claude Max subscription, acts as production assistant, scheduler, tracker and asset manager. It moves production assets to and from the studio's production asset server and keeps the manifests and asset registers up to date. It started from a sanitised version of the identity file I use for my own agentic system administrator, Maxi, adapted to production work, and in week 1 I worked with it on site to build the skill files that define its production tasks. Alongside it, AI graphics and video generation went into the production workflow, starting with marketing material and extending to client deliverables after the proof of concept.

Tools

Why a cloud and local mix

The choice came down to time and cost. Cloud-hosted models were used where the local hardware wasn't up to the job, or where doing it locally would have pushed the proof of concept over budget. The studio also wanted to find out whether local models could replace cloud subscriptions cost-effectively. That hasn't been decided yet, and the team's view is that it's too close to call.

Why ComfyUI

ComfyUI is brittle and isn't generally regarded as a stable production platform. In this case the creative team already knew it from personal projects and workflows, and were comfortable enough with its drawbacks. That gave it the flattest learning curve and the fastest route into the production workflow. Once the workflows were bedded down, they stayed stable for the rest of the proof of concept.

Implementation

The engagement ran for two weeks on site, followed by off-site support through to a review at the end of the month.

Week 1. A briefing with the team leader, agreement on outcomes and deliverables, building an asset register, one-on-one meetings with each team member, and the initial implementation, including building the production agent's skill files with it on site.

Week 2. Onboarding the agents, a progressive handover of production logistics to the team agent, and a review of how each agent was performing in its role.

End of week 2. Sign-off on the deliverables. This confirmed the proof of concept setup was in place and closed the on-site work. Everyone understood there would be off-site support afterwards, and I was happy to provide it to see the project succeed.

Week 3. Daily off-site support and fine-tuning. Most of the work was fixing agent skills when problems came up and shoring up workflow pipelines that proved brittle in real use.

Week 4. Support dropped to two calls.

End of month. Review, at which the proof of concept was accepted.

Two weeks was enough to build the system and hand it over. Bedding it in under real production took about another two weeks, and the support load fell quickly once it had. Anyone planning a similar engagement should budget for that second stage rather than treating the hand-over as the finish line.

Outcome

Production

The studio used to spend three days shooting, editing each day's footage within that week's cycle. It now spends one day on combined shooting and AI production, and puts the time saved into editing and mixing, post-production and pre-production planning. On six typical two-minute shorts, the shooting and production stage dropped from three days to one, and similar gains are expected on longer productions. High production values still need expert editing and mixing, so the AI work speeds production up but doesn't replace the editors.

Measured the way we agreed at the start of the project, production time across the whole cycle fell by 25%, which means a third more output for the same hours. That figure is drawn from the studio's asset register and timesheets, and is roughly made up of a 200% increase in still graphics, about a third of the previous time spent on some video production, and general workflow efficiencies. I've kept it deliberately conservative. The extra pre-production time isn't counted in it either, though it should increase output over the longer term.

The studio's count of finished deliverables shows a bigger result: up from four a week to six over the measured period, with further gains expected. The studio's view was that lower costs and higher output would mean a stronger bottom line.

Marketing

Marketing output rose 52%, from 158 production assets a week to 241. A production asset here is any file used in the finished marketing output: sound files, graphics, video clips, scripts and production notes. The working environment is a lot more cluttered than those tidy figures suggest, but the count is consistent before and after, and it measures volume. Whether the extra output is cutting through the competitors' noise is the next thing to measure.

Personal assistant agents

The time the personal agents freed up wasn't measured objectively. Subjectively, nobody on the team wants to give up their agent PA.

What I'd do differently

The review showed measurable productivity gains, and to me it also showed the two-part approach was right. The personal agents produced the benefit that was hardest to measure, but they were also what got the team to accept agentic workflows and work them into their day so quickly. People who had already settled into working with their own agent found the team agent much easier to take on.

Next time I would put more weight on that first layer, and stagger the team agent's rollout until the personal agents were fully bedded down, rather than bringing both in within the same two weeks.

Next steps

The studio is considering moving from individual Claude subscriptions to Team or Enterprise plans, and is still deciding whether local generation can replace cloud subscriptions. AI generation is now going into client deliverables as well as marketing, which is the studio's longer-term aim: replacing as much costly, time-consuming live production as quality allows.

Download

Personal assistant agent template (plain text, Markdown formatting). This is the template the studio's personal assistants were built from, with the studio's details removed. Fill in the bracketed sections, and replace the example tasks with your own.