AI for agencies: a practical guide

This guide is for owners and account and ops leads at marketing, paid media, SEO, and creative agencies of 5 to 100 people. The agency problem is specific: the same work repeats for every client, but no client's data, voice, or budget may touch another's. By the end you'll have seven workflows running in type.com, each with copyable instructions, a schedule, and a review rule.

Structure does most of the safety work: one private Space per client, holding only that client's connections, instructions, and memory. The workflows are written once as shared skills, so client twelve gets the same quality bar as client one.

The examples follow a fictional agency, Northlight Media (26 people, 14 retainer clients), and three of its clients: Pellwood Outdoor (outdoor gear on Shopify, Meta, and Google), Calder Dental Group (lead generation on Google Ads), and Brisa Coffee (a coffee subscription). For a full list of agency skills, see the best AI skills for agencies. In-house teams should read AI for marketing teams, and SEO agencies will find search-specific workflows in AI for SEO.

type.com's model is one Space per department rather than one agent per job. At an agency, the unit that needs a wall is the client. Every channel can use every connection in its Space, and channels aren't a permission boundary, so with two clients in one Space, a question about one can read the other's ad account.

A Space keeps four things apart. Connections are assigned per Space, so Pellwood's Meta account is invisible to Brisa's Space. Space instructions belong only to their Space, and Space memory is kept per Space. A private Space is limited to the people you add, and workspace memory draws on public activity, one more reason to make every client Space private.

Add two private internal Spaces: Agency Ops (owner, ops lead, finance) for retainers, time, and margins, and New Business for prospects until they sign. Never merge client Spaces: a merge can't be undone, and the destination's members get every moved connection.

The connections step decides whether the wall holds. For Meta, connect with a system user token assigned only this client's ad account and only ads_read. For Google Drive, select only the client's folder. If a platform can only connect through an agency login that sees every client, such as a manager account, use exports instead.

In type.com, a Read-only setting tells the agent not to make changes but doesn't block them at the connected service, so the account or token itself must be unable to edit. That makes the approval rules below enforceable.

Then create a Library document called Client profile with everything that differs by client: KPI and target, committed spend per platform, each campaign's daily budget floor and ceiling, the signed scope and exclusions, and the client approver. Every skill here reads it first.

When: the first two weeks, then quarterly. Owner: the account lead. Inputs: the signed statement of work, brand book, best emails and ads, kickoff calls, and website.

Usually the voice lives in the account lead's head and the copywriter learns it through rejected drafts. Writing it down from evidence fixes that, if the AI reads only this client's material. At an agency, a broad search for brand guidelines returns other clients' brand books.

Pellwood's guide came from the 2024 brand book, its 40 top emails, and six call transcripts. It surfaced a claims rule the copywriter hadn't known: "waterproof" only for products rated 10,000 mm or higher. Maya Chen approved it on Sep 30, and Pellwood's Grant Ellis signed off on Oct 2.

Review rule: the account lead edits both documents and gets client sign-off on the voice guide. Until then both stay marked DRAFT, and the copy skill won't run.

Shortcut: the Client Brand Voice skill in the type.com Skills Library packages the voice half. It works on one named client per run, searches only sources scoped to that client, and excludes anything that looks like another client's.

When it runs: Mondays at 7:00 AM CT, one scheduled automation per client Space, posting to reporting. Owner: the account lead, who should also create it, since scheduled runs use the creator's credentials. Create it with Create, then Automation; it starts paused, so Save, Test (which runs privately as you), then Enable.

Under deadline, report commentary turns into a list of every metric that moved. The instructions below force the opposite: revenue from the store rather than the ad platforms, a threshold for calling a change meaningful, and at most three causes, each with evidence.

Pellwood's Sep 28 to Oct 4 draft led with a 3.9x blended return against a 3.5x target ($18,420 spend, $71,840 Shopify revenue) and named three causes: fatiguing Trail Layers ads on Meta, brand search limited by budget on Oct 2 and 3, and a GA4 tracking gap on Oct 1.

Review rule: the account lead checks the three biggest numbers against the platforms, rewrites the commentary, and sends by noon. AI never emails the client. For the general pattern, see how to automate a weekly team report.

Shortcut: for month-end, the Client Report Builder skill builds the monthly report from exports with a script, computes blended CAC and return, never sums platform-claimed return, and ranks next month's changes with dollar figures.

When: daily at 8:00 AM CT, one automation per client Space, posting to paid-media. Owner: the media buyer. Inputs: the ad accounts and the Client profile's commitments, tolerance, and campaign floors and ceilings.

Two things make a pacing check usable. Proposals must add up exactly to what's left of the commitment, inside each campaign's floor and ceiling, or nobody can apply them. And the check must find the cause, because a budget raised last week keeps overspending until its owner knows.

On Tuesday, Oct 6, Calder Dental had spent $4,960 of its $24,000 October commitment against $3,871 expected, on track for $30,752. The change history showed the Implants Search budget raised from $350 to $500 on Sep 30. The check proposed $732.31 a day across three campaigns, Brand held at its $150 floor. Nikhil applied it in Google Ads and replied "applied 9:14 AM".

Review rule: the media buyer approves or edits each change, applies it in the platform, and confirms in the thread. The next run reads the budgets back.

Shortcut: the Budget Pacer skill does this math in a script from exports and says when floors and ceilings make the commitment impossible. The Ad Account Audit skill is a good first-week hygiene pass on a new client's account.

When: on demand in creative. Owner: the copywriter, reviewed by the account lead. Inputs: the brief, brand-voice.md, the Client profile, and the client's best ads from the last 90 days.

The hard part isn't volume. It's copy that sounds like the client and makes no claim they can't support. So the skill writes from what has worked for this client, labels each variant's angle so the set doesn't converge, and checks every line against the client's rules.

In Pellwood's Trail Layers refresh, the check flagged a variant calling the Ridge Shell jacket "waterproof". It's rated 5,000 mm, below the guide's 10,000 mm rule, so the fix suggested "water-resistant".

Review rule: the account lead approves voice and strategy, and the client signs off on anything published. Regulated clients such as Calder Dental also get their own compliance review. See AI agent approval workflows. Shortcut: Client Brand Voice's review mode returns line-level suggested edits against one client's guidelines.

When: as requests arrive by forwarded email or a Slack mention. Owner: the account lead. Inputs: the request and the Client profile's scope.

Scope creep arrives as quick favors done before anyone checks the statement of work. Checking at request time leaves real choices: quote it, swap it, or do it once on purpose. After the work is done, you can only absorb it.

On Tuesday, Oct 6, Brisa's Lucia Ortega asked for four TikTok cutdowns of a founder interview. Brisa's scope covers 12 static Meta ads a month and excludes video editing. The check marked it out of scope, estimated 10 to 14 hours, and Maya chose to quote it.

The monthly half runs in Agency Ops, because it needs time entries, contracts, and loaded costs. On the first Monday at 9:00 AM CT, an automation computes per client the effective hourly rate (fee divided by hours), delivery margin (fee minus loaded cost, divided by fee), deliverables shipped against scope, and renewals due within 45 days.

Shortcut: the Retainer Scope Tracker skill packages the monthly review with a script, catches over-delivery even when hours look fine, and ranks retainers to renegotiate.

When: per pitch, in the New Business Space. Owner: the new business lead. Inputs: the prospect's shared files, their site, your benchmarks, and a case study document of results clients approved for use.

A conflict check comes first, because pitching a current client's competitor can cost you that client. And every number must trace to the prospect's data or a stated assumption, because their team will check your math.

When a prospect signs, create a client Space and copy in the signed scope and audit. Don't move the pitch channel: its threads, including pricing discussion, would follow.

Shortcut: the New Business Pitch Builder skill turns a prospect's exports into prioritized gaps, confidence-adjusted sizing in gross profit, and a 90-day plan.

Agency margin comes from doing the same excellent work for every client without rebuilding it. A skill is that method written down; the Space supplies the client. Then a new hire can run client twelve's report at client one's standard.

Build the skill once under Library, Skills, then add the same organization skill to each client Space. Keep client values out of it. The docs recommend that reporting skills read a shared definitions file and keep different clients' definitions separate, which is the Client profile's job. Creators and admins edit a skill; anyone else can suggest a change for the owner to review.

At Northlight, Jordan suggested a call-quality line for clients whose KPI is qualified leads. Theo approved it on Oct 1. It now runs for Calder Dental and every lead-gen client, and Pellwood's report is unchanged.

Copies you import from the Skills Library don't update automatically, so treat them as starting points you own. Shortcut: for paid search, the Google Ads Command Center skill routes Google Ads work across monitoring, search terms, and reporting using approved definitions, and keeps proposals separate from execution.

Pin this in each pod's internal channel.

Write these rules into each Space's instructions and skills, then enforce them with the permissions in the tools themselves. Instructions alone are not a security control.

Know who runs what. AI acts with the effective access of the person and conversation that started the work. Scheduled runs use the creator's credentials, so when someone leaves, recreate their automations under a current owner and remove their connections and Space access. Keep credentials out of prompts and skills.

Agency-wide logins need their own check. Once a month, the MCC Security Audit skill can run from Agency Ops as a read-only scan of your Google Ads manager account tree, flagging unexpected manager links, unknown users, pending invitations, and accounts not mapped to a client. Every fix is a manual step.

Pick two pilot clients with clean data and one skeptical account lead whose edits will sharpen the skills fastest. Widen only when the step before has held up.

Take the baseline before turning anything on, from last month's time entries, pacing results, and copy revisions. The 30-day goal is the same or better work in fewer hours per client, not more output.

When a workflow drifts, fix the skill rather than the thread, so every client gets the correction. Read how shared AI memory works for teams for what memory does and doesn't carry, and see the best AI skills for agencies for the full list of agency skills.