An AI-assisted campaign pipeline, designed around the human decision

Context

A creative and media production agency runs on two kinds of work. One is the craft its clients pay for: strategy, ideas, film, design, media planning. The other is the administration around every campaign: budget approvals, status reports, rights checks, reconciling a media plan with the approved order. This case concerns the second.

This case describes design work prepared for a proposal after one initial meeting with the agency. There was no contract, no access to the agency's systems or client data, and no discovery. Everything shown ran on fictional clients, campaigns and figures. We publish it because the reasoning (where AI sits in a campaign pipeline, and where it must stop) can be assessed by a peer on its merits.

The challenge

With no discovery behind them, the problems below are our standing hypotheses about organisations at this point of adoption, not findings about this agency. An adoption gap. The proposal was written for an agency working in an office suite with AI features built in; the hypothesis is that non-technical staff make little use of them.

Operational friction. Time goes to manual reporting, email triage and repetitive campaign administration. Generic tools. An off-the-shelf assistant knows nothing about how an agency approves a budget or clears a film for a market, so it tends to stall at the first agency-specific task.

One concern is particular to agencies. Much of what moves through a campaign pipeline belongs to clients, or is bound by talent and music contracts. A design that sends, publishes or clears anything on its own would create more risk than the time it frees.

Our approach

We proposed three strands: building every colleague's confidence with the AI tools already in place, a review of operating processes to find systemic friction, and joint development of new internal and client-facing tools.

1. 1Listen before prescribingA three-week discovery comes first: interviews yield a friction map, pattern analysis ranks use cases by impact and effort, and a tailored playbook follows.
2. 2Teach on people's own workFour learning tracks by level of responsibility. Each participant works on a workflow they already own, so the first automation built is one they need.
3. 3Show the pipeline end to endTwo scripted demonstrations share a fictional client: routine budget approvals at one end, a launch-readiness review under pressure at the other.
4. 4Write down the stopping pointsNo send and no publish stay on screen for the whole simulation, and sources are marked read-only. The written brief adds that legal clearance is never automatic.

What was designed

The enablement programme

The learning design rejects one course for everyone. Executives work on vision and governance; senior leaders on redesigning workflows and leading change; managers on directing hybrid teams of people and AI; practitioners on prompting and automating their own work. Around the tracks sit eight deliverables, among them three fortnightly workshops, an asset library the agency keeps, a six-monthly adoption survey and a final impact review.

Everyday automation, built by conversation

The first demonstration shows a practitioner, not a developer, assembling a campaign budget approval workflow through three plain-language requests to the office suite's assistant (Gemini writing Apps Script inside Google Workspace). Approval emails are read into a tracking sheet. Each row fills the approval template, which is exported as a PDF and emailed to the budget owner. A status deck updates as campaigns are added. The walkthrough is a scripted playback on fictional clients and campaigns.

Launch readiness with several agents

The second demonstration is a 90-second simulation of a harder day. A fictional campaign launches the next morning in a dozen markets, the client committee meets in ninety minutes, and nearly two hundred synthetic files (contracts, budgets, creative assets, approvals) offer no single version of the truth. Inside an isolated project with read-only access, a lead agent turns an account director's request into an auditable plan and delegates to four specialists working in parallel: evidence lineage, finance and media, creative rights, and experience checks.

The simulation surfaces contradictions with their source: usage rights that expire before launch in two markets, a media plan above the approved order, a claim without enough support. Then it stops. The responsible person chooses the remedy, and only then is a seven-slide client deck assembled, each statement linked to its evidence. A failed quality check is corrected and rerun, and a read-only re-audit is scheduled before go-live. Deterministic checks (hashes, versions, dates, budget sums) are kept apart from the model's work of classifying, extracting and summarising.

A third piece is a concept only, mocked up and not built: an agent that follows selected public social accounts during a product launch and turns the reaction into signals for strategy, creative and media teams.

Results

There are no outcome figures in this case, and that is deliberate. The proposal carried a simple capacity calculator (people trained, hours freed each week, working weeks), but its inputs were illustrative defaults, not numbers from the agency. We publish nothing from it.

What the work produced is a design that an agency's leadership, operations and client-service leads can each challenge. Were it to proceed, discovery would first take a baseline: hours spent on campaign administration, time from budget approval to issued paperwork, and use of the AI features already in place. No contracted engagement is claimed.

Governance and risk

The person conducts; the agents play. The design asks each colleague to move from performing every task to setting the intent, assigning the parts, reviewing the evidence and deciding what happens next. In the simulation, agents detect risks; legal clearance stays with the people accountable for it. Nothing is sent or published automatically, and connectors are labelled read-only.

Client and campaign data. The proposal takes a privacy-by-design position: minimise the data used, restrict access, encrypt in transit and at rest, keep ownership explicit, and never use client or campaign data to train public models. These are design positions. None has been through the agency's security review or its clients' contractual terms, which is where a real engagement would begin.

What we learned

• Decide where the pipeline must stop before deciding what it automates. In agency work, sending, publishing and legal clearance belong to named people.
• Teach automation on a workflow the person already owns. A budget approval they handle every week is a better first lesson than a generic prompting course.
• Separate what code can verify (dates, sums, versions) from what a model infers, and show the source for every flagged risk.

Client identity, locations and identifying details are withheld under confidentiality. Figures are rounded. Measured figures come from engagement records; client-reported figures are attributed, not audited; modelled figures are projections from the engagement's business case and are labelled as such.