Agentic workflows for agencies should automate first the work that repeats often, can be undone cheaply and can be checked against a source: data pulls, competitor monitoring and report drafts. Agents should assist on judgement-heavy work such as personas, copy and creative scoring. Anything that spends money, makes a public claim or carries the client relationship stays human, with the agent preparing and a person approving.
Key takeaways
- Sort every task with three tests: how often it repeats, whether a mistake can be undone, and whether the output can be checked against a source.
- Automate data gathering, monitoring and first-draft reporting; Click2View Asia cut monthly reporting from 9 hours to 1 on SOMIN.
- Use agents as assistants on persona, tension and copy work, where a strategist still makes the choice.
- Keep budgets, claims, final strategy and crisis replies human, with an explicit approval step.
- Redesign the workflow around the agent rather than bolting it onto the old one; that habit separates high performers in McKinsey's 2026 survey.
What is an agentic workflow?
An agentic workflow is a process in which a language model does part of the work by calling tools and, in places, choosing its own next step. It sits between fixed automation and a fully autonomous agent, and most agency value lives in that middle ground.
Anthropic's engineering team draws the distinction cleanly in Building effective agents. Workflows orchestrate models and tools through predefined code paths; agents dynamically direct their own process and tool use. Its advice is to find the simplest solution that works and add complexity only when needed, and it describes agents that pause for human feedback at checkpoints and stop after a set number of iterations.
For an agency that translates into a design rule. Use a fixed workflow where the steps are known, such as a weekly report. Use an agent where the path is open, such as researching a category the team has never pitched. Put a checkpoint wherever a client or a budget is affected.
Why do so many agent projects stall?
Because teams automate the impressive, judgement-heavy work first, instead of the dull work that is easy to check. The demo looks good, the output needs heavy correction, and the business case disappears.
Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. Its advice maps neatly onto three modes: use agents when decisions are needed, automation for routine workflows and assistants for simple retrieval.
McKinsey's State of AI in 2026, a survey of 1,719 participants, finds about two in ten respondents scaling AI agents across their organisation. Among its high performers, nearly three-quarters report fundamentally redesigning workflows because of AI, against one-quarter of everyone else.
SOMIN's data shows the cost of getting this wrong. In our ads-optimisation workspace, a 30-day whitespace study of 156 posts (September 2026) groups a theme around AI and marketing effectiveness: AI agencies that put speed ahead of quality, producing generic outputs and unmet business goals. The audience-research workspace holds an insight, built on 254 posts, that says it more briefly: "In the pursuit of speed, let not integrity be the casualty."
How should agencies decide what to automate first?
Score each task on volume, reversibility and verifiability. Automate when all three are high, assist when the output needs judgement to verify, and keep it human when a mistake cannot be undone or the client relationship is at stake.
| Task | Mode | Why |
|---|---|---|
| Pulling cross-platform performance data into one view | Automate | High volume and rule-based; every figure can be checked against the platform |
| Competitor ad and post monitoring | Automate | Repetitive watching across many accounts; a miss is cheap to fix next cycle |
| First draft of the monthly client report | Automate, human reads | The narrative follows the data; an account lead reads it before it is sent |
| Social-listening theme summaries | Assist | The model clusters posts; a strategist decides which themes matter to this client |
| Persona and tension discovery | Assist | Evidence-linked drafts save days; a planner validates them against client knowledge |
| Copy variants and resizes | Assist | Volume work, but brand voice and claims need a reviewer |
| Creative pre-scoring before spend | Assist | A score informs the choice; it should not make it |
| Content calendar and scheduling | Assist | Drafting is mechanical; publishing stays gated until quality is proven |
| Budget changes and campaign activation | Keep human | It spends money and is hard to reverse; the agent prepares, a person approves |
| Final strategy and the recommendation | Keep human | Accountability and the relationship sit with the agency, not the tool |
| Claims, regulated categories and legal sign-off | Keep human | Liability stays with the advertiser however the ad was made |
| Crisis responses and sensitive community replies | Keep human | Context, tone and timing change by the hour |
What should you automate first?
Start with data pulls, competitor monitoring and report drafts. They eat senior hours, they repeat every cycle and their errors are visible.
The Click2View Asia case study shows the size of the prize. With SOMIN as its reporting layer, the content agency cut monthly client reporting from 9 hours to 1, comprehensive competitive analysis from 18 hours to 3 and content strategy groundwork from 14 hours to 2, and moved from two people per workstream to one. Those hours went back into consultation, which is the work clients pay a premium for. SOMIN's Performance Reporting agent builds on the same engine.
The need is plain in SOMIN's social-listening data too. A perspective study of 129 posts (October 2026) includes a theme on content and social media management that flags time-consuming content creation and inefficient approval workflows. In the ads-optimisation workspace, SOMIN's moments-of-truth model surfaces the thought "How can I make sense of all this data?"
Where should agents assist rather than decide?
On work where most of the effort is research and the final step is taste: personas, tensions, concepts and copy. The agent proposes with evidence; a person chooses.
SOMIN's Strategy agent takes a median 13 minutes from brief to a stack of pre-scored concepts, but a strategist picks what goes to the client. The Paid Media agent scores concepts for predicted click-through before spend, and pushes the top ads into campaigns that are built paused, so nothing activates without sign-off. The full set is on our AI agents page.
The assist mode also protects juniors. A junior who reviews an agent's evidence learns why an angle works, which a junior who only formats slides never does.
What must stay human, and how do you enforce it?
Anything that spends money, makes a public claim or carries accountability to the client. Enforce it in the workflow itself, not in a policy document.
- Approval steps in the tool. Spend and publish actions wait for a named person, by default.
- Stopping conditions. Cap iterations and cost per run, so a confused agent stops instead of looping.
- Read-only by default. Give agents write access one tool at a time, as described in our guide to connecting Claude to marketing data with MCP.
- Logs a person can read. Every tool call and approval should be visible after the fact.
SOMIN's social-listening data captures why this matters to the people doing the work. Its moments-of-truth model surfaces the thought "I need to streamline our operations while keeping the human touch in our customer interactions." Efficiency and the human touch are both requirements.
FAQ
What are agentic workflows for agencies?
They are agency processes in which AI agents gather data, draft outputs or take steps through connected tools, with people approving the decisions that matter. Reporting, monitoring and research are the usual starting points.
Which agency tasks should be automated first?
Tasks that repeat often, can be checked against a source and are cheap to correct: cross-platform data pulls, competitor monitoring and first drafts of client reports.
Will AI agents replace agency strategists?
They replace hours, not judgement. Agents compress research and drafting, while the recommendation, the client relationship and accountability for claims stay with people.
How do you keep a human in the loop with AI agents?
Build approval steps into the tool for anything that spends money or publishes, cap how long an agent can run, start with read-only access and keep a log of every action.
This week, list every task on one client account for a single week and score each for volume, reversibility and verifiability. Automate the top two, run them alongside the manual version for two cycles, compare the errors, and only then switch over.
