Where are your hours going?
Where does AI fit in agency work?
AI fits best in the assembly steps around creative work: turning briefs into plans, first drafts, format variants, research, reporting and summaries. It fits worst where taste, accountability and client trust live: final creative decisions, approval and crisis communication. Map your process, then automate the assembly, never the judgment.
| Step | What AI does | What a person does | Level of automation |
|---|---|---|---|
| Client onboarding | Turns intake answers into a brief and house rules | Runs the kick-off, validates the brief | HIGH |
| Research | Scans competitors, trends, past performance | Picks the insight worth acting on | HIGH |
| Content plan | Drafts a month of ideas per pillar and channel | Chooses, cuts, adds the client's news | MEDIUM |
| Copy and variants | First drafts, per-network versions, translations | Edits for voice, facts and taste | MEDIUM |
| Visuals | Concepts, resizes, generated assets via design tools | Art direction, labelling of realistic AI | MEDIUM |
| Scheduling | Places drafts in the calendar with correct formats | Checks timing and context | HIGH |
| Approval | Prepares previews, collects comments | Approves. Always. | NONE |
| Reporting | Pulls data, writes commentary, builds custom views | Adds recommendations and context | HIGH |
| Community and crisis | Suggests replies, flags sentiment shifts | Decides and responds | LOW |
The 30-60-90 day rollout plan
Pilot
- 3 to 5 people on business seats
- 2 pilot clients with full house rules
- Pick 3 workflows from the map above
- Everyone logs what worked in one shared doc
- Policy and data traffic light live (Level 2)
Standardise
- Turn winning prompts into a shared library
- House rules for every active client
- First connected tools via MCP
- Short internal training, documented for AI literacy
- Agree the metrics you will track
Scale
- Roll out to the whole team with templates
- First agent workflow in production (Level 4)
- Update client contracts with the AI clause
- Review hours saved and reinvest them
- Decide your new service offers
The most common failure is skipping the middle column. A pilot that never becomes templates stays a hobby of two enthusiastic people.
A prompt pattern the whole team can use
Good agency prompts have five parts: who the AI should be, the context (house rules and data), the task, the output format, and a self-check. Teams that standardise this pattern get consistent output regardless of who is typing.
Role: You are a senior social media strategist at our agency.
Context: Follow the house rules for [Client] attached to this project. Last month's top three posts were [...].
Task: Propose [number] posts for [channels] for the week of [date], supporting [campaign goal].
Format: A table with date, channel, format, hook, caption and visual brief. Captions within each channel's limits.
Check: Before answering, verify every post against the house rules and list any rule you could not follow.
Need ready-made prompts for social media work with an AI agent? Browse the ZoomSphere MCP prompt library.
Client house rules generator
House rules are the single most effective upgrade an agency can make. Write them once per client, attach them to every AI task, and the output stops sounding generic. Gathering the inputs from the client? Send them our client intake form.
Why human approval stays non-negotiable
Because AI can be confidently wrong, because the agency is accountable to the client, and because the EU AI Act's transparency rules treat human editorial review as a safeguard. The goal is not fewer approvals, it is faster ones: AI prepares, people decide.
Build approval into the tool, not into goodwill. In ZoomSphere, posts created by an AI agent arrive as drafts in the content calendar and follow the same approval workflow as anything your team writes, so the client sees exactly what will go live. Read more about approval workflows.
How to measure if it works
- Hours per deliverable: time to produce a monthly report or a week of content, before and after.
- Edit distance: how much of the AI draft survives. Rising survival means your house rules are working.
- Approval rounds: fewer client correction rounds per post.
- Output quality: engagement of AI-assisted posts against your baseline, never just volume.
- Team sentiment: a monthly two-question pulse. Adoption dies when people feel replaced instead of helped.