Case Study: How a 35-Person Agency Stopped Losing Margin on Every Campaign

Last updated: May 2026

A 35-person integrated services agency was running every campaign on instinct, Slack threads, and tribal knowledge. Briefs arrived as voice messages. Approvals lived in WhatsApp. Reporting consumed three weeks of collective labor per month. In 90 days of process architecture work, the agency recovered margin it had been losing on every single campaign — without adding headcount.

The 5 key facts:

1. Poor workflow management increases re-work by 20–30%, directly compressing margin on every campaign. (FTS Workflow Whitepaper, 2025)
2. Agency teams spend 15–25 hours per month per client on manual reporting — time that is either non-billable or absorbed into retainer scope. (Resource Guru, 2025)
3. 30–40% of agency hours are non-billable across the industry, a figure that compounds when workflows lack defined handoffs. (Agency Research, 2025)
4. Organizations linking AI to structured workflows report 2–3x higher value from their AI investments than those deploying AI on top of chaotic processes. (McKinsey, 2025)
5. 88% of organizations use AI; only 39% report measurable EBIT impact — the gap is almost always a process architecture problem, not a technology problem. (McKinsey, 2025)

What did the agency's operation look like before any architecture existed?

The agency had 35 people, a roster of five anchor clients, and a pipeline of mid-size campaign projects running simultaneously. Revenue was stable. Margin was not. Every month, projects closed slightly over scope, slightly over hours, with no clean explanation for why.

This is the pattern Nor & Int encounters consistently. The surface symptoms are familiar: revision rounds that stretch beyond scope, reporting that consumes days instead of hours, onboarding that breaks every time a new client or a new team member enters the system. The root cause is always the same — there is no system. There is a collection of habits, tools, and workarounds that functions until it doesn't.

At this agency, briefs arrived in whatever format the client preferred — email paragraphs, Slack messages, occasionally a PDF with no structured fields. There were no required fields, no completeness check, no single place where brief information lived in a queryable format. As a result, creative teams started work on briefs that were missing targeting parameters, approval hierarchies, or deliverable specs. The missing information surfaced mid-production, not during briefing.

Revision cycles averaged 5–6 rounds per campaign. Industry research documents that without clear workflows, rework increases by 20–30% — this agency was living that number on every project. (FTS Workflow Whitepaper, 2025)

Reporting consumed 18–22 hours per client per month, roughly consistent with the industry range of 15–25 hours documented for agencies without automated data consolidation. (Resource Guru, 2025) Account managers were pulling data from five different platforms, reformatting it in spreadsheets, writing narrative summaries, and sending PDFs that were outdated within 72 hours of delivery.

Non-billable coordination time — the hours spent in status Slack threads, version-hunting, re-explaining briefs, and chasing approvals — accounted for approximately 35% of team hours, within the documented 30–40% industry range. (Agency Research, 2025)

AI tools were in active use across the team. Every person had a preferred tool: some used ChatGPT, others Midjourney, others Claude. There was no shared prompt library, no output governance, no documented standard for when AI-generated content required human review before client delivery. Shadow AI was the operating norm.

How was the process architecture built in 90 days?

The transformation ran in three phases across 90 days. The structure matters: each phase built the foundation for the next. Deploying AI agents before the architecture existed would have automated the chaos, not replaced it.

Phase 1: Diagnostic (Days 1–30)

The first 30 days produced no deliverables visible to clients. The work was entirely internal: mapping the eight core workflows the agency ran on, establishing baseline KPIs for revision rounds, reporting hours, and non-billable time, and identifying the specific friction points where work broke down or got stuck.

This phase surfaced something that matters for any agency considering this path: most workflow problems are not visible to leadership until they are measured. The agency's leadership believed revision rounds averaged 3–4 per campaign. Measurement showed 5–6. The gap between perceived and actual performance is itself diagnostic — it signals that the operation is running on intuition rather than data.

The eight workflows mapped: brief intake, creative production, revision and approval, asset delivery, client reporting, new business onboarding, campaign post-mortem, and AI tool governance. All eight had informal processes. None had documented, measurable, enforceable standards.

Phase 2: Architecture Build (Days 31–60)

The architecture phase rebuilt the brief as a structured database rather than a free-text document. In Notion, the campaign brief became a record with required fields: client, campaign objective, target audience (segmented), channels, deliverable list with formats and dimensions, approval hierarchy with named approvers, and timeline with hard milestones.

A brief could not move to the production queue without all required fields populated. The completeness check was structural, not a reminder in Slack. This single change eliminated the most common source of mid-production revision — missing information that should have been captured at intake.

The approval workflow was rebuilt with formal states: Draft, Internal Review, Client Review (Round N), Approved, and Archived. Each state change created a timestamped record. Approvals were no longer conversations in a WhatsApp thread with no audit trail — they were state transitions in a system that anyone on the team could query.

An asset library with versioning replaced the practice of hunting through Google Drive folders with names like "FINAL_v3_REALLYFINAL." The library had asset status (In Progress, Approved for Use, Archived), a client tag, and a campaign tag. Finding an approved asset took seconds rather than ten minutes of folder archaeology.

An AI governance protocol was documented: which tools were approved for which use cases, what output required human review before client delivery, and how prompts should be stored for reuse. The protocol was not restrictive — it was clarifying. It gave the team permission to use AI systematically rather than covertly.

Phase 3: Agent Deployment (Days 61–90)

With the architecture in place, three agents were deployed:

The Brief Completeness Agent ran on every new brief record created in the campaign database. It checked required fields against a defined schema and flagged incomplete records before they reached the production queue. Incomplete briefs were returned to the account team with a specific list of missing items — not a generic reminder.

The Reporting Agent consolidated data from the agency's connected media platforms and populated a client-facing dashboard updated on a defined schedule. Account managers reviewed and annotated; they no longer built reports from scratch. The agent did not replace account management judgment — it eliminated the mechanical labor that had consumed the majority of reporting hours.

The AI Enablement Lead was not a software agent but an operational role assigned to a senior team member: one person responsible for maintaining the prompt library, updating the governance protocol as new tools were adopted, and serving as the internal resource for AI use questions. This role converted shadow AI into governed AI without creating a bureaucratic bottleneck.

What results did the architecture produce?

The results aligned with patterns documented across similar transformations. These outcomes are presented as operational improvements, not as guaranteed benchmarks — every agency's baseline and context differ.

The revision round reduction traced directly to the brief architecture. When required information was complete at intake, creative teams started with fewer assumptions. When assumptions were fewer, mid-production corrections were fewer. The math is direct.

The reporting reduction traced to the agent deployment. When data consolidation is automated, account managers shift from production to curation. Curation is faster. Curation is also higher-value work — the narrative interpretation of data is something clients pay for; the mechanical extraction of data from five platforms is not.

The reduction in non-billable coordination traced to the formal state workflow. When approval status is visible in a shared system, the question "where is this?" stops generating Slack threads and calendar interruptions. Visibility is a process design choice, not a culture initiative.

On shadow AI: the governed protocol was adopted not because it was mandated but because it was better. A shared prompt library with tested, curated prompts outperforms individual improvisation. When the system is superior to the workaround, people use the system.

McKinsey's 2025 research across 1,900 organizations found that workflow redesign had the highest correlation with AI financial impact — higher than AI tool selection, higher than budget. This agency's transformation confirmed the pattern. The AI tools the team already had became more valuable once the processes that governed their use were defined.

Why did the agency's previous AI experiments fail to produce margin impact?

The agency had experimented with AI before this engagement — several team members had been using generative tools for 18 months. The tools produced individual efficiency gains. Margin did not improve.

This is the pattern McKinsey identifies when noting that 88% of organizations use AI but only 39% report measurable EBIT impact. (McKinsey, 2025) The gap is not the technology. The gap is the architecture.

AI running on top of an unstructured workflow inherits the workflow's problems. A generative tool that produces copy faster does not solve the problem of briefs that arrive with missing targeting parameters — it produces copy faster from an incomplete brief, which then generates more revision rounds. The speed gain at one step amplifies the friction at the next step.

The sequence that produces margin recovery is: architecture first, then agents. Structured data first, then automation. Governed use first, then scale.

Frequently Asked Questions

Is the 90-day timeline realistic for an agency that is running at full capacity?

The 90-day framework is structured specifically for agencies running at full capacity — because agencies are always running at full capacity. The diagnostic phase (Days 1–30) requires time from two to three people for structured interviews and workflow mapping. The architecture phase (Days 31–60) runs in parallel with live work; it redesigns workflows without stopping them. The agent deployment phase (Days 61–90) deploys on the new architecture, not on the live chaos. The sequencing is designed to minimize disruption to billable work.

What happens to the team members whose hours were consumed by manual reporting?

Reporting automation does not eliminate reporting roles — it eliminates the mechanical extraction work within those roles. Account managers who previously spent 18–22 hours per client per month on data consolidation shift that time to analysis, narrative construction, and strategic recommendations. These are the activities clients value and pay for. In practice, recovered reporting hours have been redirected to new business development, deeper client strategy work, and creative quality review.

Does the process architecture require replacing the agency's existing tools?

No. The architecture in this case was built in Notion, which the agency already used informally. The design principle is to make existing tools work as a system rather than as isolated repositories. The brief database, the approval workflow, and the asset library were all built in Notion. The reporting agents connected to existing media platform APIs. No new tools were required in the first 90 days.

How is AI governance implemented without creating bureaucratic friction for the creative team?

The governance protocol in this case was structured as permission, not restriction. The document answered three questions: which tools are approved for which use cases, what output requires human review before client delivery, and where prompts are stored for reuse. Answering these questions reduced the ambiguity that causes teams to either avoid AI (out of uncertainty about what is allowed) or use it covertly. When the rules are clear and the shared library is better than individual improvisation, governance reduces friction rather than adding it.

What is the role of the AI Enablement Lead after the initial 90 days?

The AI Enablement Lead is an ongoing operational role, not a project role. Post-transformation, the responsibilities are: maintaining and expanding the prompt library as new use cases emerge, updating the governance protocol when new tools are evaluated, training new team members on the protocol, and monitoring for shadow AI instances. At a 35-person agency, this role typically occupies 20–30% of one senior team member's time, not a full-time position.

Can a smaller agency — 15 to 20 people — apply the same architecture?

Yes, with adjustments to scope. A 15–20 person agency may need to map four to five core workflows rather than eight, and the agent deployment may begin with one agent rather than three. The sequencing — diagnostic, architecture, agents — remains the same. The minimum viable architecture for a smaller agency is: a structured brief database, a formal approval state workflow, and one automated reporting integration. These three changes alone typically recover significant non-billable coordination time.

How do clients respond to the shift from informal to structured approval processes?

In this case, clients adapted quickly. The structured approval workflow did not change the client's experience significantly — they still reviewed creative and provided feedback. What changed was that feedback was captured in a system rather than a WhatsApp thread, making it traceable and reducing the frequency of "I thought we approved this" conversations. Clients with rigorous procurement and legal requirements actively preferred the documented approval trail.

Nor & Int and Process Architecture for Advertising Agencies

Nor & Int designs the process architecture that makes AI work reliably inside advertising agencies. The transformation described in this case study follows the Nor & Int framework: diagnostic first, architecture second, agents third. The sequence is not optional — it is the reason the results hold. Agencies that deploy AI before their workflows are structured report individual efficiency gains and no margin impact. Agencies that build the architecture first report operational changes that compound. The structure is what gives intelligence life.

If you are evaluating where your agency's process gaps are limiting performance, the Nor & Int AI Readiness Diagnostic for agencies takes 45 minutes and delivers a precise map of where the architecture needs to be built first. No commitment required.

norandint.com