Ad Agency AI Operating System: A Lightweight Version in Notion

An AI operating system for an ad agency is a shared workspace where the agency's brand-voice rules, brief templates, AI prompts and approval steps live in one place, so every draft starts from the same inputs and passes the same checks. A simple version is five connected Notion databases, built in a few days.

It is not a product you buy or an autonomous agent that runs the agency. It is closer to a well-kept kitchen: ingredients labelled, recipes written down, and a rule that nothing leaves the pass until someone checks it. The AI is one cook among several, and the system is what makes its output consistent and safe to send.

Follow me on Instagram@sagnikteaches

The five pieces, and what each one does

Resist the urge to build twenty databases. A lightweight agency system needs five.

Connect on LinkedInSagnik Bhattacharya

Each piece links to the others through Notion relations. The work tracker row for a deliverable points to the client's voice page, the brief used, the prompt used and the approval entries. That chain is what makes the output auditable: if a client asks why a headline used a certain phrase, you can trace it.

Subscribe on YouTube@sagnikteaches

Piece one: the client voice library

The voice page is the highest-value part of the system because it is the thing most agencies keep only in one person's head. A usable page has six fields:

• Who we're talking to: one sentence on the audience.
• Tone in four words: for example "plain, dry, helpful, unhurried".
• Words we use / words we never use: two short lists.
• Five approved lines: real copy the client signed off, which the AI can imitate.
• Five rejected lines with the reason: the client turned these down, and why.
• Claims rules: what must never be promised or compared, and who clears a claim.

The rejected lines carry more information than the approved ones. Illustrative: the online clothing store client rejected "Dress to impress" with the note "too pushy, we don't do slogans". That single example does more to steer a model than a paragraph of adjectives. Ask a strategist to fill the page from the last ten pieces of approved and rejected work, then have the AI summarise patterns and draft the tone words for the account lead to edit.

Piece two: brief templates that the AI can use

A brief the AI can use is a brief a new freelancer could also use. Every brief template gets the same core fields, with deliverable-specific extras:

CLIENT: [relation to voice page]
DELIVERABLE: Paid social ad set (3 variants)
OBJECTIVE: One sentence, one measurable aim
AUDIENCE: Who, where they are, what they already believe
THE ONE THING: The single message
PROOF: Facts we may use (with source) and facts we may not
OFFER AND CALL TO ACTION: Exact wording and link
MANDATORIES: Legal lines, hashtags, brand rules
FORMAT: Character limits, platform, number of variants
APPROVER: Name on client side
DEADLINE: Date and time

Make "Proof" a required field. It is where AI-written copy most often goes wrong, since an assistant will happily write "voted best in the area" if the brief leaves a gap. If the field is empty, the prompt should instruct the assistant to write without claims.

A worked brief, end to end

Illustrative case: a ten-person agency has a client who runs a pest-control firm. The deliverable is a set of three paid social ads for a spring wasp-nest removal offer.

The account manager creates a row in the work tracker, picks the "Paid social ad set" brief template and fills it in: the objective is "20 booked call-outs in four weeks from a $1,500 budget"; the audience is homeowners with gardens; the one thing is "Fast, safe removal with a fixed price"; proof is "Fixed price of $85 for a standard nest, same-week visits" (the client's own figures, with a link to the client's email confirming them); mandatories include the client's licence line.

The prompt from the library takes the brief and the voice page and asks for three variants within the brief's limits for this client: headlines of up to 40 characters and primary text of up to 125. A sample of what comes back, illustrative:

Variant A. Headline: "Wasp nest? Fixed $85 removal." Text: "Spotted a nest this spring? We'll visit this week and remove it safely for a fixed $85. Book a call-out in two minutes."
Variant B. Headline: "Same-week nest removal." Text: "No guessing at the price. $85 for a standard nest, and we come this week."
Variant C. Headline: "Don't poke it. Call us." Text: "Wasp nests are best left to us. Fixed-price removal, booked online."

The first gate is the strategist's. She notices Variant C's "don't poke it" is playful, and the client voice page lists "plain, calm, practical, reassuring". She keeps it, marks it as the bold option and logs the decision. The second gate is the compliance check: all three include the licence line from mandatories. The third gate is the client's. This chain takes under an hour for the draft stage, including the review, where previously the account manager would have written the three variants alone over an afternoon.

Piece three: prompts that live with their owner

A prompt library is a table of prompts, each with a purpose, an owner, a version, a "last tested" date and an example output. Prompts decay when models change, so give each one a test routine: run the prompt on two old, approved briefs and compare the output with what the agency actually shipped. If the output is clearly worse, fix the prompt before anyone uses it.

Keep prompts short and structured. The core prompt for the ad-set deliverable has four parts: role ("you write paid social copy for [client]"), inputs (brief and voice page pasted in), rules (use only the facts in Proof, respect character counts, never invent offers), and output format (a table with headline, text and a one-line rationale). The rationale line is useful for reviewers: it shows what the assistant was trying to do, which helps to spot when it misread the brief.

For agencies working across a Notion workspace, Notion's own AI features can read pages in the workspace. As of November 2026, Notion's pricing page shows Plus at $10 and Business at $20 per member per month billed yearly ($12 and $24 billed monthly); the Notion Agent and AI meeting notes appear under Business, with only a limited trial of Notion AI on Free and Plus. Custom Agents are priced separately: free to try, then $10 per 1,000 monthly Notion credits. Plans change, so check the page before you commit. You can also keep the prompts in Notion and run them in a separate assistant, which avoids any dependence on the workspace plan.

Piece four: the work tracker and its statuses

The tracker needs status values that make the gates visible. A sensible set: Briefed, Drafting, Internal review, Fixes, Client review, Approved, Live. "Internal review" is where the human gate sits; the tracker should refuse, by convention or by a Notion automation, to move to "Client review" without an approval log entry. On paid plans, a Notion database automation can fire when a property such as Status is edited, then edit properties or send a notification to named members, which is enough to tell the strategist a draft is waiting (the free plan allows only Slack notification automations). For triggers and alerts, Notion automations for tasks and alerts covers the options.

Include a "Source of draft" property with values such as AI-assisted, human-written, and AI-translated. It costs nothing and answers the question clients increasingly ask. If your contracts mention AI use, AI clauses for agency contracts explains what to say about disclosure and approvals.

Piece five: approval gates that mean something

A gate is a question a named person answers, with a yes or no and a note. Four gates cover most agency work:

1. Brief gate: before drafting. Is the brief complete, are claims sourced, is the approver named?
2. Craft gate: after the draft. Does it follow the voice page, hit the objective, read well aloud?
3. Claims gate: before client review. Is every factual claim traceable to the Proof field, and does anything need legal or regulatory clearance?
4. Client gate: the client approves in writing, and the approval is logged with the date.

The approval log is a database with one row per gate per deliverable: the deliverable (a relation), gate type, reviewer, result, note and date. Illustrative: the wasp-nest ads pass the brief gate on Monday; fail the claims gate on Tuesday because Variant A's headline, "Fixed $85 removal", drops the words "standard nest" and the client's price list has a surcharge for roof nests; the writer changes it to "Standard nest? Fixed $85 removal."; the claims gate passes on Wednesday. The log shows the failure, the fix and the date, which is more useful in a client dispute than any amount of memory.

Setting it up in order

1. Day one, morning: create the voice-library database and fill one client page from their past work.
2. Day one, afternoon: build the brief template for the deliverable you produce most.
3. Day two: write the prompt, test it on two old briefs, and fix it until the output is within one edit of what you shipped.
4. Day three: build the work tracker and approval log, relate them, and run one live deliverable through.
5. Week two: add the second client; adjust the templates when the second client breaks an assumption.
6. Month one: review the approval log. Which gate fails most? Fix the cause upstream.

For the wider team picture, the AI rollout plan for a ten-person marketing agency covers training and sequencing, and AI content workflow for agencies goes deeper on the brief-to-draft stage that this system hosts.

Three more examples of the system in use

A car repair garage's monthly email. The brief template for emails includes "reason to write now". The garage's brief says "service reminders for customers whose cars are due in the next 60 days". The assistant drafts subject lines; the strategist rejects two for sounding like scare tactics, referencing the voice page's "no fear" rule, and the rejection is added to the rejected-lines list so the next run avoids it.

A furniture maker's landing page. The brief's Proof field lists "hand-finished, made to order, lead time 6 to 8 weeks". The assistant writes a draft that says "delivered in two weeks". The claims gate catches it because the phrase has no source in the brief. The fix goes into the prompt: "never state delivery times except from Proof."

A freelancer's overflow. The agency brings in a contractor for a surge. The voice page, brief and prompt are the contractor's whole induction, and the work tracker shows what is in flight. Because the system is the same one the staff use, the contractor's work passes the same gates.

Building a voice page from past work: a sample run

The slowest part of setup is the first voice page, so it is worth using the assistant to speed it. Collect ten pieces of the client's approved copy and five rejected ones, each with the client's comment. Strip any names and personal data. Then use a prompt like this:

Below are 10 approved lines and 5 rejected lines (with the client's reason) for one client.
1. Describe the tone in four words.
2. List 8 words or phrases the approved lines favour and 8 they avoid.
3. State in one sentence what the rejected lines have in common.
4. Suggest 3 rules a new writer should follow.
Base every point on the examples. Do not generalise beyond them. If the evidence is thin, say so.

Illustrative output for the pest-control client: tone "plain, calm, practical, brisk"; favoured words "visit", "fixed price", "this week"; avoided words "infestation", "nightmare", "invasion"; common thread of the rejected lines: "they dramatise the pest rather than describing the job"; suggested rule: "describe the service, never the fear." The account lead edits it: she swaps "brisk" for "reassuring" because the client uses that word on calls, and adds a rule the examples could not show, that the client never names competitors. Together this takes about an hour, and the page is something the whole team can read in two minutes.

Permissions and where client material may go

Agencies hold material that belongs to clients, sometimes under contract terms about tools and data. Build the permission question into the system rather than leaving it to memory. Each client page carries three yes-or-no fields: may drafts be generated with an external AI tool, may the client's unpublished figures be pasted into prompts, and is anything restricted to named staff. The brief gate then includes a check that the answer is "yes" before drafting.

In Notion, teamspaces and page sharing can restrict who sees which client's pages; Notion lists private teamspaces and granular database permissions under its Business plan on its pricing page, so a larger agency that needs that separation should confirm the plan before building. A smaller agency can achieve most of the same effect by keeping one database per client, which is a little more setup and avoids relying on filters to hide things.

A monthly hour to keep the system alive

Systems like this die by neglect, not by failure. Diary one hour a month with a fixed agenda:

1. Open the approval log and sort by result. Which gate failed most this month, and for which deliverable type? Change the brief template or prompt that caused it.
2. Check the prompt library: which prompts were not used? Retire them. Which were edited locally? Pull the improvement back into the shared version.
3. Re-run the test briefs against the main prompt. If the model behind your assistant has changed and results drifted, note it and adjust.
4. Read one client voice page with the account lead and ask whether the last month's rejected lines are on it.
5. Look at the tracker for deliverables that went live without an approval entry. Each one is a gap in the process, not a person to blame.

It is also the right moment to review which work still needs a person from start to finish. A sensitive crisis statement, a client's difficult announcement or a pitch for a new account do not belong in a drafting prompt, and the system should say so on the brief template, with a checkbox: "human-written only".

What goes wrong in a lightweight system

• Voice pages go stale. Add a review date and re-read them when a client rebrands. A stale voice page produces confident copy in last year's tone.
• Everyone keeps a private prompt. The library fails if people prefer their own. Make the shared prompt easier to find than a note on a desktop, and invite changes through the owner.
• Gates become ticking exercises. If every gate passes every time, the reviewer is not reading. Spot-check one deliverable a month by re-running the claims gate yourself.
• Confidential client material in the wrong place. Decide which clients' materials may be pasted into which AI tools, record it on the client page, and follow the client's contract.
• Complexity creep. If a database is not queried within two months, delete it.

Checking that the system is paying for itself

After a month, compare three numbers for the deliverable you moved first. Time from brief to first internal draft. The number of rounds of client changes. And the number of gate fails caught internally rather than by the client. Illustratively, if the ad-set draft stage fell from half a day to an hour and a half, and the client asked for one round of changes instead of two, the system has earned its keep even before you count the approval log's value in disputes. If the numbers do not move, look at the brief template before you look at the prompt: weak briefs, not weak prompts, are the usual cause.

Questions about building an agency AI system in Notion

Do I need a paid Notion plan to build this?

You can sketch the databases and templates on the free plan, but a free workspace with two or more members can add only a limited number of blocks, and database automations other than Slack notifications need a paid plan. Notion's AI features also differ by plan, with a limited trial on lower tiers, so check the pricing page first.

Is this the same as buying an agency management platform?

No. Agency platforms handle time, billing and resourcing. This system covers the creative and approval side: voice, briefs, prompts and sign-offs. Many agencies run both, and the Notion layer is cheap to try first.

Can the AI publish directly to clients?

It should not. Keep a human approval gate between every AI draft and anything a client or the public sees. The gates in the system exist for exactly that reason.

How long does a lightweight version take to set up?

A working version with one client takes two to three days of part-time effort: a day for the voice library and brief template, a day for prompts and testing, and the rest for the approval database. Add clients one at a time.

Further reads

• Agency Client Onboarding With AI: Questionnaire, Access and Kick-Off — Collect the voice and access details at the start of a client.
• Agency Client Retention: Spot At-Risk Retainers With AI — Use the work tracker data to spot at-risk retainers.
• Best AI Tools for Advertising Agencies: Copy Variants, Mock-Ups, Reporting — Tools to plug in around the Notion core.
• Is Notion AI Worth It Now It Comes With Paid Plans? — Whether Notion AI itself is worth paying for.
• How to Write Business Proposals Faster With AI — Reuse the voice library for pitches and proposals.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Notion pricing page (notion.com/pricing, checked 4 November 2026). The workspace design, briefs and figures are illustrative.