What should an agency automate first? Not whatever the newest tool promises. Before automating anything, answer three questions: how complex is the decision, how much damage can a bad output cause, and what measurable outcome improves if you automate it? Start only when the answers are clear. The quick wins sit where the reward is high, the risk is low and the expertise needed is moderate.
Clients are asking similar questions, more nervously. Digiday reported this month that holding companies and independents are launching AI agents for planning, buying and measurement while stressing human oversight to reassure clients worried about agents going rogue. Dentsu's Carat is building a planning and buying agent for the upfronts with NBCUniversal, FreeWheel and Newton, and Tinuiti has launched a Bliss Point MCP server that opens its measurement layer to AI agents. WPP Media, quoted in the piece, put it plainly: "governance and human accountability are central to agentic media buying."
What should you ask before automating anything?
One Singapore agency founder, after 18 months of trying to make his own agency AI-native, offered a filter at an AI adoption session this week that fits on a single slide. As the diagram shows, it comes down to three questions:
- How complex is the decision? Is it a rule a junior could follow, or a judgement call that takes ten years of experience?
- How much damage can a bad output cause? A clumsy internal summary is an annoyance. A wrong bid change, or an off-brand post to a client's whole audience, is a phone call nobody wants to make.
- What measurable outcome improves if we automate it? Hours saved, turnaround, error rate, revenue. If you cannot name it, you will never prove it worked.
His rule was to start only when all three have clear answers. That sounds obvious. In practice, most agencies start with the tool and invent the justification afterwards.
Which agency tasks are quick wins, and which should you leave alone?
The founder's guidance was clear. High-risk, high-expertise work: don't touch it. Look instead for high reward, in revenue or productivity, combined with low risk and moderate human expertise. Those are the quick wins.
Here is how that might play out on a typical agency task list.
| Task | Decision complexity | Damage from a bad output | Measurable outcome | Verdict |
|---|---|---|---|---|
| Drafting the monthly performance report | Low to moderate | Low, if a person reviews it before sending | Hours per report, turnaround | Quick win |
| Resizing assets across formats | Low | Low, and easy to spot | Time per asset, error rate | Quick win |
| Competitor and social monitoring | Moderate | Low | Time to insight, coverage | Quick win |
| Campaign setup and QA | Moderate | Medium, because budget is at stake | Setup time, error rate | Automate, with human approval |
| Moving budget between channels | High | High, because money moves | Hard to attribute cleanly | Human-led; AI can advise |
| Crisis response and client strategy | High | Very high, and reputational | Hard to measure | Don't touch |
Look at what lands in the quick-win rows: the chores. Monthly reports, resizing, monitoring, competitor audits. Exactly the jobs that drain senior hours and that nobody went into advertising to do.
Why are clients worried about AI agents going rogue?
Because the agentic pitch often skips question two. An agent that can move money or publish content can do damage at speed. Clients have read the headlines and want to know who is accountable when it gets something wrong.
The agencies Digiday spoke to are answering with governance: human oversight, approvals, clear accountability. That is now a selling point, not a compliance footnote. An agency that can show a client its three-question filter, and the tasks it deliberately keeps human, sounds more trustworthy than one boasting about how many agents it runs.
For independents, this is an advantage. You do not have to automate the high-risk work to compete. You have to automate the right work well and explain why you left the rest alone.
Why does automating everything at once backfire?
The impulse behind "if I find the right tools, I can get control of my day back" is understandable. The result is often the opposite: six new tools, six logins, six half-finished workflows and a team that spends its day checking machines.
Cognitive Load Theory explains why. Working memory can only juggle a few things at once. Every new tool, exception and decision about whether to trust an output adds load, and past a point people make worse decisions, not better ones. Simplicity is not a nice-to-have. It is how attention survives a busy agency week.
A short decision rule takes that load away. Three questions, applied every time, replace dozens of debates about which shiny tool to try next. The quieter the rule, the more it gets used.
How do you go from a good idea to a working automation?
The founder closed with five questions to leave the room with: Problem, Ambition, Readiness, Roles and Results. Add one final step and they make a practical pilot plan.
- Problem: what are we actually trying to fix? Name the bottleneck in one sentence. "Reporting eats three days a month per client" is a problem. "We need AI" is not.
- Ambition: how far does this function need to go? A light-touch assistant may be enough. Full automation costs more and needs more people to run it.
- Readiness: which of our four dimensions is weakest? Context, people, process or platform. Fix the weak one before you scale.
- Roles: who builds, who uses, who approves? Make the subject-matter expert the approver, so nothing reaches a client without a named person's sign-off.
- Results: what will prove it worked? Set the baseline and the target before you start, and count the time spent checking output.
- Pilot one workflow for one client. Collect feedback weekly and adjust before rolling it wider.
Why can't agencies grow as fast as they can build?
The point from the talk that stuck: you can now build at 100x speed, but you cannot grow at 100x speed. Growth needs user feedback, and people are slow. A workflow built in an afternoon still needs weeks of real use, correction and trust before a team relies on it.
That is the case for starting with one quick win and doing it properly. SAMY, a global influencer agency, adopted Defy Agents to streamline performance reporting and reduced three days of work to one: low risk, moderate expertise, a measurable outcome. On the Defy platform, agentic workflows take on repetitive delivery such as reporting, monitoring and resizing, so people can focus on the work that wins.
Then ask the three questions again for the next task. Simple, repeated, compounding.
