How do you integrate AI tools into an agency's existing workflows? Not by starting with the tools. Start by watching how the work really gets done, find the one place where it gets stuck, and only then decide whether AI belongs there. Apply AI to a bad process and you simply get a fast bad process.
That was the uncomfortable lesson from an agency founder speaking at an AI adoption session in Singapore this week. He spent 18 months trying to make his roughly 20-person agency "AI-native" and was candid about what he would undo. His most memorable image: put a robot vacuum in a messy house and it doesn't clean anything. It drags the mess across the floor.
The timing is pointed. Digiday reports that agencies are patching AI into their existing master service agreements clause by clause rather than rewriting them. The triggers are a client wanting a seat in the agency's proprietary AI tool, a client asking for AI language, or a new AI tool turning up on a client project. Mile Marker's CEO told Digiday: "We are at the point now where we are just simply working that into existing MSAs." Contracts get patched. So do workflows. Patches on patches is how agencies end up automating chaos.
Why does AI on a bad process give you a fast bad process?
Because automation multiplies whatever it touches. If the brief arrives half-formed, an AI drafting tool produces half-formed concepts faster. If approval sits with one overbooked client services director, faster production just builds a longer queue outside her door.
Most agencies know their processes are messy. What they underestimate is how much of the mess is load-bearing. The Singapore founder compared his business to his own untidy home: chaos to a visitor, but he knew exactly where his wallet was. Agencies run the same way. How a pitch really comes together, or why one client's reports always need a second pass, lives in people's heads.
That knowledge is invisible to an AI tool. It is often invisible to new hires, too. His argument was that the first job is to make it accessible to both, before anyone buys another seat.
What is the difference between an SOP and a real workflow?
An SOP describes how the work is supposed to happen. A workflow is how it actually happens. The founder's advice was blunt: sit next to the subject-matter expert and watch them do the job. The gap between the SOP and what they really do is where the bottlenecks hide.
As the diagram shows, the SOP is a tidy checklist. The real workflow loops back on itself, detours through Slack and email, waits for someone, and has one point where everything piles up. That point, marked in red, is the only place where automation will pay off.
Every agency has its own version. The project-management tool says creative review happens on Tuesdays; in reality the ECD reviews on her phone between pitches and the team works late on Thursdays. The media SOP says the monthly report is "pull data, build deck, send". In reality an analyst spends half a day reconciling platform numbers that never quite match.
Where do agency bottlenecks usually hide?
The talk named five usual suspects: production, approval, the brief, the handover and decision-making. The rule that came with them: if you cannot identify which one is your bottleneck, do not automate anything. Here is how they tend to show up across agency types.
| Bottleneck | How it shows up | Fix first | Where AI can help afterwards |
|---|---|---|---|
| The brief | Creative shops: concepts rejected because the problem was never agreed | A short brief template the client signs off | Pre-filling category and competitor context |
| Production | Social and performance teams: the same asset resized by hand for every placement | Agreed formats and naming conventions | Automated resizing and versioning |
| Approval | PR agencies: statements stuck waiting for one senior sign-off | Named deputies and approval deadlines | Drafts that arrive ready for checking, not rewriting |
| The handover | Media and CRM agencies: context lost between strategy, planning and activation | One shared record of decisions | Summaries that travel with the work |
| Decision-making | Influencer agencies: creator shortlists debated for weeks | Clear criteria and a single decision-maker | Faster evidence-gathering for that decision-maker |
Notice the order. In every row the fix is human first and technical second. AI rarely solves an unclear brief or an absent decision-maker. It just makes their cost visible faster.
Why does this keep happening?
Systems Theory offers a plain explanation. A system's output is shaped by how its parts connect, not by how good any single part is. Speed up one step and the whole only improves if that step was the constraint. If the constraint sits elsewhere, you have simply built a bigger pile in front of it.
An agency is a system of people, briefs, clients, tools and deadlines. Dropping a powerful AI tool into one corner of it is like fitting a faster engine to a car stuck in traffic. Impressive under the bonnet, no quicker to the meeting.
The contract patching Digiday describes is the same pattern at the commercial level: add a clause each time a new tool appears rather than redesign how the relationship works. It is understandable. It also raises the stakes. If clients start asking for seats in your proprietary AI tools, they will see the process underneath. A fast bad process is much harder to hide when the client is logged in.
How do I document a workflow before automating it?
- Pick one workflow that hurts. Choose something frequent and painful, such as the monthly client report, the competitor audit or the pitch deck. Not "everything".
- Find the person who really does it. Not the head of department who wrote the SOP, but the account executive or analyst who does it every week.
- Sit next to them for a full cycle. Watch the work from trigger to delivery. Note every tool, message, wait and workaround. Ask "why" more often than "what".
- Draw the real flow next to the SOP. Put the two side by side, as in the diagram, and mark every place they differ.
- Time the waits, not just the tasks. Agency work often moves quickly and then sits for days. The longest wait usually points to the bottleneck.
- Name the bottleneck in one word. Brief, production, approval, handover or decision. If the team can't agree, keep watching. Don't automate yet.
- Fix the human part, then test the AI part. Change the process first, then pilot automation on the steps that are still repetitive, and compare the result with the old way.
Write the result down somewhere both people and AI can read it. A documented workflow is the first asset an AI system can actually use, and the first thing a new hire will thank you for.
What should an agency automate first?
Once the bottleneck is clear, the best first candidates are usually the chores nobody went into advertising to do: monthly reports, competitor audits, manual resizing, status updates. They are repetitive, they follow rules you can now describe, and the judgement sits around them rather than inside them.
That is the territory Defy Agents are built for: agentic workflows that take on repetitive delivery such as reporting, monitoring and resizing, so people can focus on the work that wins. SAMY, the global influencer agency, adopted Defy Agents to streamline performance reporting, reducing three days of work to one. Reporting is exactly the kind of job that can be described, measured and handed over once the workflow is understood.
If you want to see how agents, monitoring and data access fit together, the Defy platform lays it out. But do the mapping first. The tool is the last decision, not the first.
