Agency AI Operating System — Field Guide

Your best thinking shouldn't live in one person's head.

This handbook is for creative agency leaders who are serious about building operational intelligence, not just distributing AI tools. It covers how to structure the context, skills, and knowledge that make your whole team work like your best person, across every account, every time.

The problem this solves

Most agencies use AI the way they used to use freelancers: task by task, client by client, with nothing that carries over. On a typical account, the account lead knows the client. The senior copywriter knows the voice. The designer knows the visual language. The creative director holds the brand standards. When any one of them is out, or when the agency scales, that expertise doesn't transfer with them.

Agency AIOS is the infrastructure that distributes your agency's best thinking. Not a tool stack. Not a prompt library. A system that compounds what your agency already knows, so every project starts from your best work instead of from zero, and every person on the team can produce work that's faster, higher quality, and cheaper to deliver.

01 — How to think about it

Agency AIOS is not about individual tools. It's about building shared operational intelligence.

Every creative agency has the same structural problem: the expertise that makes the work good lives in individuals, not in systems. AI tools amplify individual output, but without shared context, they amplify inconsistency at the same rate. The answer is not more tools. It's infrastructure.

Before

Intelligence is individual

Each person on the account or the project holds a piece of the context, but no one has access to all of it. New projects start with briefing sessions. New hires shadow for months. When a senior person leaves, everything they knew walks out with them.

After

Intelligence is organizational

Client context, feedback patterns, process knowledge, and institutional standards are captured in systems every team member can access. AI runs on that context, so every output reflects your agency's standards and your client's brand no matter who produces it.

The tradeoff

More structure, less firefighting

The cost is building and maintaining the knowledge layer, capturing what you know in a form that AI and your team can use. The payoff shows up in fewer revision rounds, fewer client escalations, and a creative director who finally has time for the work only they can do.

Questions to ask before applying AI to any agency workflow
  • Is this knowledge or just information? Does this task require judgment, or does it require context? AI handles context, not judgment.
  • Who is the single person whose absence would make this task impossible? That's your knowledge concentration risk.
  • What would happen to this account if that person left tomorrow?
  • What does "on-brand" mean for this client, in specific, checkable terms, not just adjectives?
  • What does the CD always correct in outputs from junior team members? That's your first skill to build.
  • Where do new team members go wrong on this account in their first month?
  • Would a 30-minute-per-brief time saving here meaningfully free up creative director capacity for the work that actually moves clients?

"Every creative agency has the same structural problem: the expertise that makes the work good lives in individuals, not in systems. AI tools amplify individual output, but without shared context, they amplify inconsistency at the same rate."

Section 01 — How to think about it
02 — The five-layer stack

Agency AIOS is layered, not tool-shaped.

Each layer has a different job. The mistake most agencies make is treating their tool stack as their operational system, which means context lives in chat threads, voice lives in individual writers, and institutional knowledge evaporates with every departure.

Interface layer Where the team interacts with AI day-to-day: brief distillation, voice calibration, pitch structure, status reporting, and creative review. Examples: Claude Code, a brief intake prompt, a voice calibration workflow, a pitch structure assistant.
Agentic layer Reusable skills and commands that encode how your agency works: how you take a brief, how you check voice, how you build a pitch, how you scope. Examples: brief distillation skill, voice calibration skill, scope builder, pitch structure command, debrief synthesis skill.
Project layer Per-client and per-campaign workspaces where execution happens: briefs, production notes, feedback logs, and deliverables in progress. Examples: client context folder, campaign brief workspace, creative review log, production handoff notes.
Knowledge layer Your agency's accumulated intelligence: agency context built once, client context built account by account, and what you've learned across every account. Examples: client brand profiles, account history logs, process library, pitch archive, new hire context files.
Records layer Official and archived: approved brand guides, signed SOWs, final deliverables, campaign archives, and approved work that future projects can reference. Examples: approved campaign assets, signed contracts, final brand guides, case study source material, post-mortem archives.

The useful rule: work in context, compound what you learn, keep a creative director in every decision that changes the work, not every decision that produces it.

03 — Layer setup guides

Set up each layer by the job it performs.

The expandable panels below are the operating manual for each layer: what it is, why it breaks when it's missing, how to set it up, and what it looks like in daily agency use. Start with the knowledge layer. Everything else runs on it.

Layer 01

Interface Layer

Where your team interacts with AI for production work: the prompts, intake flows, and workflows that give everyone a consistent, context-loaded starting point.

What it is

The surfaces the team uses to get work done with AI: a brief intake prompt that loads client context, a voice calibration workflow that reads the client's profile, a pitch structure command that applies the agency's approach.

Why it matters

Without a shared interface, every team member uses AI differently. Outputs don't match your standards. Senior time is spent fixing what junior AI produces rather than directing it. The agency adds AI and somehow adds work.

What goes wrong when it is missing

Each person starts from scratch differently. A copywriter prompts with their own phrasing. An account manager prompts with theirs. Neither has loaded the client context. The outputs don't match each other or the client's brand, and someone senior has to reconcile them. AI adds velocity to inconsistency.

How to set it up
  1. Designate one primary AI interface for production work: the tool your team uses when they're making things, not exploring.
  2. Build intake tasks for the most common starting points: brief distillation command, tone of voice skill, pitch template.
  3. Define a review gate: everything that goes to a client gets a human read before it leaves the building.
  4. Assign a quality owner for each output type: creative director for copy and concepts, account lead for status and reporting.
  5. Log what AI produces and what gets changed; you'll find patterns that improve your skills faster than any other practice.
Agency examples
  • A copywriter writes a draft using the client's tone of voice skill. The first draft is already calibrated to the client's brand, not a generic starting point.
  • An account manager runs a brief distillation on 40 pages of stakeholder input. It extracts the actual creative problem and writes the brief in 10 minutes rather than three briefing sessions.
  • A designer produces on-brand work for an account they've never touched by loading the client knowledge file before they start creating concepts.
  • A CD uses a presentation alignment skill before every client presentation. It reads the brief history, previous decisions, and the client's stated sensitivities to produce a 10-line context summary before the call.
Layer 02

Agentic Layer

Reusable skills and commands that encode how your agency works, capturing the judgment of your best people so it can be applied consistently at any scale.

What it is

The procedures that make your agency's approach repeatable: how you take a brief, how you check voice, how you structure a pitch, how you scope a project. Each skill encodes what your best person knows so anyone on the team can apply it.

Why it matters

Most agencies have two or three people who "know how to do it right." When those people are unavailable, the work drops. The agentic layer captures that institutional judgment so it doesn't leave with the person who holds it.

What goes wrong when it is missing

Junior outputs require heavy CD revision because there's no encoded standard for what "good" looks like on this account. Pitches are rebuilt from scratch each time because the structure that worked last time lived in someone's head. Scopes are inconsistent because each account manager approaches them differently. The CD becomes the bottleneck not because they want to be, but because they're the only quality gate.

How to set it up
  1. List the five tasks where you most often think "I'll need to revise this before it goes to the client." Those are your first skills to build.
  2. For each task, write a skill brief: what context to load, what to produce, what format, what the CD would check on review.
  3. Build evaluation criteria: not just "does it sound right" but specific, checkable outputs. What would make the CD approve it without revision?
  4. Run each skill three times on real client work before relying on it in production.
  5. Update skills when the same correction appears twice; that correction is a skill refinement, not a quality problem.
Skills worth building first
  • Brief distillation: reads a raw client brief and extracts the real creative problem, the audience, the tone constraints, and the measurable success criteria.
  • Voice calibration: reads a client's brand profile and recent work samples, calibrates AI outputs to match the client's specific register.
  • Pitch structure: takes a brief and produces a first-draft narrative structure for a client presentation, in the agency's preferred approach.
  • Scope builder: takes a brief and outputs a project scope with line items, rationale, and dependencies.
  • Debrief synthesis: reads post-project notes and extracts learnings for the knowledge layer, so the next campaign on this account starts smarter.
  • Status report: reads the current project state and produces a structured client status update in the agency's preferred format.
Layer 03

Project Layer

The file system of working knowledge: what's in flight right now, for this campaign, at this account. Distinct from the knowledge layer's lasting context and the records layer's approved archive.

What it is

Per-campaign workspaces where production happens: the current brief, production notes, this round's client feedback, creative review logs, and in-progress deliverables. It answers "where does this project stand right now," not "what has this agency learned." The working room, not the archive.

Why it matters

Without a defined workspace, project context lives in email threads, Slack messages, and people's inboxes. The next person to pick up the work has no idea what decisions were made, what the client said last Tuesday, or why that headline got changed.

Recommended agency folder structure
clients/
  [client-name]/
    account-brief.md      ← who they are, business context, goals
    voice-profile.md      ← how they sound, examples, do/don't
    relationship-log.md   ← conversations, decisions, sensitivities
    history-log.md        ← what we've made, what worked, what didn't
    campaigns/
      [campaign-name]/
        brief.md          ← distilled creative brief
        production/       ← in-progress work
        feedback-log.md   ← client feedback, round by round
        review-notes.md   ← CD review notes per round
        deliverables/     ← approved outputs

agency/
  process-library/        ← how we work
    brief-intake.md
    pitch-structure.md
    scope-template.md
    review-criteria.md
  pitch-archive/          ← past pitches, what won, what didn't
  new-hire-context.md     ← what someone needs to understand us
How the team should work with it
  • Every client account has its own folder. AI reads the current campaign folder before producing anything for that client.
  • After every client interaction, a team member adds a one-line update to the relationship log.
  • After every CD review, review notes go into the campaign folder, so the next draft addresses what actually changed instead of guessing.
  • Approved deliverables move to the records layer. In-progress work stays in the project layer.
  • When a campaign hands off between team members, the incoming person reads the current brief and feedback log first, not the account's full history.
Agency examples
  • A production designer picks up a campaign mid-flight and reads the current brief and feedback log to see exactly what's approved and what's still open, without pinging the account lead to catch up.
  • After a round of client feedback, the account manager logs it directly against the campaign. The copywriter's next draft addresses round two instead of guessing what changed.
  • A campaign pauses for two weeks over the holidays. The person who picks it back up reads the review notes and resumes exactly where the last round left off, instead of re-reading a month of email.
Layer 04

Knowledge Layer

Two kinds of accumulated intelligence: agency context, which is how you work, and client context, which is what you know about each account. Distinct from the project layer's working files and the records layer's approved archive.

What it is

Agency context is how you work: culture, values, standard operating procedure, the judgment calls your best people make without thinking about them. Client context is what you know about each account, and because clients and their projects make up most of what an agency actually does, this is where most of the layer's value accumulates.

Why it matters

The knowledge layer is the difference between AI that helps and AI that produces work you have to rebuild. Without it, every prompt starts cold. With it, every prompt starts with the accumulated intelligence of everyone who has ever worked on this account.

What goes wrong when it is missing

AI outputs are generic because AI has no context to make them specific. Voice calibration is manual every time because there's no brand profile to reference. New team members produce off-brand work for months because the brand isn't written down anywhere useful. When a senior person leaves, everything they knew about an account goes with them. The CD reviews everything because they're the only accessible source of context.

What to build first
  1. Process library (agency context, build this first): how your agency works, how you scope, how you present, how you give feedback, how you onboard a new client. The unwritten rules, written.
  2. New hire context file (agency context, build this first): what someone needs to understand to work at this agency, meaning the standards, the culture, and the judgment calls that are never explained but always expected.
  3. Client brand profile (client context, build one per account): the client's business, their audience, their competitive position, what makes them distinct, and how they sound. Owned by the account lead, not just the writers.
  4. Account history (client context, build one per account): who they are, what they're trying to do, what they've tried that didn't work, what the CD always says about them. Written in the voice of an experienced team member briefing a new one.
  5. Pitch archive (agency context, grows with every account you win or lose): past pitches with notes on what won and what didn't. Patterns that appear across wins are your competitive intelligence.
The client brand profile: what it should actually contain

Most "brand voice" documents stop at voice, which is useful to the writer and designer producing the work but incomplete for the account lead directing it. A useful client brand profile contains:

  • Business context: what the client does, who their audience is, what they're trying to achieve, and where they sit against competitors. This is what an account lead references most.
  • Decision-makers and sensitivities: who signs off, what they've reacted badly to before, and what's off-limits.
  • Voice and tone: sentence rhythm, vocabulary this brand uses and would never use, and structural habits (does it open with a claim or a question, does it use subheadings), demonstrated with real examples rather than adjectives.
  • Calibration examples: 3 to 5 pieces of approved work you'd use to show a new writer what "on-brand" looks like. The more specific the better.
  • Common corrections: what the CD always changes when writing comes back from junior team members. These are the calibration targets.
Layer 05

Records Layer

The official archive: approved deliverables, signed contracts, final brand guides, campaign archives, and approved work that future projects and new team members can learn from.

What it is

The governed archive where finalized, approved work lives, separate from the working project layer, accessible to the right people, and structured so AI can actually read and learn from it.

Why it matters

The records layer is what a new team member learns from. It's what AI reads to understand what "done" looks like for this agency and this client. It's also what the client references two years later when they ask "what did we decide?"

What goes wrong when it is missing

Every new project reinvents the wheel because past work isn't findable. New hires can't learn from case studies because the approved work is scattered across personal drives, email threads, and Dropbox folders nobody manages. A client asks for the final version of a brand guide from 18 months ago and nobody can find the approved file with confidence.

How to set it up
  1. Designate a single, consistent location for approved final work, separate from project workspaces.
  2. Establish a clear rule: a draft lives in the project layer until a client approves it. Only approved final versions move to records.
  3. Build a searchable index: a simple text file that lists what's in the archive and why it matters. AI reads indices, not folder structures.
  4. After every major campaign, write a one-page post-mortem that lives alongside the records: what worked, what didn't, what to do differently.
  5. Define who can access what: client-approved work is usually sharable; signed contracts and financial documents are not.
Making records useful for AI calibration

A records layer that just stores files is an archive. A records layer that's structured for AI use is a compounding competitive asset. The difference:

  • Add a one-paragraph annotation to every major piece of archived work: what brief this answered, what made it succeed, what the client said about it.
  • Tag archives by what they're useful for: "voice calibration example," "pitch structure reference," "tone under-pressure example."
  • Build a case study source file for each client: a structured account of what you did, what you learned, and what you'd do differently. AI reads this when starting a new campaign.

"The knowledge layer is the difference between AI that helps and AI that produces work you have to rebuild. Without it, every prompt starts cold. With it, every prompt starts with the accumulated intelligence of everyone who has ever worked on this account."

Section 03 — Layer setup guides
04 — Operating patterns

The habits that keep output quality consistent without routing everything through one person.

These patterns are small enough to embed into existing workflows. Each one closes a specific gap: brand drift, cold-start briefing, context loss on handoff, and the knowledge that evaporates when a senior team member moves on.

Load contextAI reads the client file, brand profile, and relevant history before producing anything.
Produce draftFirst output calibrated to the client's voice and the brief's actual creative problem, not a generic starting point.
CD reviewCreative director refines and approves. Corrects, adjusts, adds judgment. Does not rebuild from scratch.
Client deliveryOutput goes to client with human sign-off. No AI output reaches a client without a team member reading it first.
CompoundWhat the CD changed and what the client responded to gets encoded back into the knowledge layer. The next brief starts smarter.
The brief distillation pattern

Never start a project with a raw client brief. Most clients communicate the symptom, not the problem. A 40-page deck of stakeholder input has one real creative brief inside it, and finding it manually burns senior time that could go elsewhere.

The brief distillation skill reads the raw input and extracts: the actual creative problem, the audience, the constraints, and what success looks like in measurable terms. The output is a one-page brief that the CD can approve in 10 minutes. From that approved brief, everything downstream starts aligned.

The brand compounding pattern

Brand profiles that don't get updated stop being useful within six months. Voice calibration compounds only when profiles are updated after every major delivery, not just at account setup.

The rule: after every client approval, AI adds insights to the brand profile. What did the client specifically praise? What did they push back on? What did the CD change, and why? With virtually no manual maintenance, each project builds a brand profile that dramatically outperforms the one you wrote at onboarding.

The context harvest pattern

The most expensive knowledge in your agency is the context that exists only in a senior team member's head. The context harvest pattern makes that knowledge organizational before it walks out the door.

After every major delivery, a short structured debrief: what did we learn about this client's actual decision-making process? What creative approaches work? What language do they respond to? What would we do differently? These notes feed directly into the account context file so the next project team starts with the full picture.

The new hire ramp pattern

The traditional agency onboarding for a new account: weeks of shadowing a senior, attending client calls, asking questions. The knowledge transfer is slow and expensive because it's oral: one senior person's time, one new team member at a time.

With a built knowledge layer, a new team member works with AI to understand the account: the brand profile, history log, and pitch archive. The AI pulls in the context needed to bring the new hire up to speed. Their first draft already sounds like someone who's been on the account for six months, because they have access to the same context that person has. Ramp time drops from weeks to days.

5 layers in the stack
10 implementation steps
11 use cases mapped
4 starter templates
05 — Implementation roadmap

Start with the account where the CD spends the most review time.

Don't begin by building a system. Begin by identifying one account where the quality gap between junior output and client-ready output is highest, and build the knowledge layer for that account first. Everything else follows.

Map your knowledge concentration risks.

For each of your top five accounts, ask: who is the single person whose absence would make this account unserviceable? That person is a concentration risk. Their knowledge is the first thing to capture.

Who holds it

Whose absence would make this account unserviceable within a week?

What goes wrong

What does the CD always correct in junior outputs for this account?

Invisible rules

What does a new team member need to know that nobody ever writes down?

Transfer test

If that key person left tomorrow, how long would it take to bring their replacement up to speed?

Voice check

What does "on-brand" mean for this client, in specific, checkable terms?

The account where more than two of these point to the same person is your highest-risk account and your best starting pilot.

Score each workflow for AIOS fit.

Prioritize recurring, text-based workflows where the inputs are consistent, the output is reviewable, and the CD dependency is high.

Frequency

Does this happen weekly, monthly, or per-project with regularity?

CD dependency

How many rounds of CD revision does this typically require?

Repeatability

Are the steps consistent enough to write down as a process?

Source material

Does the context AI needs already exist somewhere the team can access?

Reviewability

Can a CD spot an error in the output quickly enough that the review gate isn't a bottleneck?

Risk if wrong

What's the cost of a bad output: a revision, or a client conversation?

Brief distillation and voice calibration typically score highest on this matrix. Start there, not with pitch decks or brand strategy.

Choose one workflow to pilot.

Pick the workflow where CD review time is highest, the steps are most consistent, and a bad output won't immediately damage a client relationship. Brief distillation is usually the right first pilot.

Sit with the person who does it best and document how they do it.

Watch your most effective account manager or senior CD run this workflow. Write down the steps, the cues they look for, the judgment calls they make automatically, the things they'd tell a new hire. That document becomes the skill brief.

Build the project folder for your pilot account.

Create the account folder structure with a context file, brand profile, and history log. Even if they start empty, having the structure means the team knows where to put things, and AI knows where to look.

Build the first knowledge file: the client brand profile.

Pull your three best pieces of approved work for this client. Have AI write a brand profile that uses those pieces as examples rather than adjectives. Include the CD's notes on what makes each one work. This is the most important document in the knowledge layer.

Build the first skill: brief distillation or brand calibration.

Turn your documented process into a reusable skill: what context to load, what to produce, what format, what the CD checks on review. Write it as instructions to a very competent person who has never worked at your agency.

Add evaluation criteria before you run it on real work.

What would make the CD approve this output without revision? Write down three to five specific things. These become your quality checks: not "does it sound right" but specific, observable outcomes.

Run the pilot on a real brief.

Use the skill on an actual client brief. Log what the AI got right, what the CD changed, and why. The changes are not failures; they are your next round of skill refinements.

Measure CD review time and expand.

Track how many rounds of CD revision the piloted workflow required before and after. That number going down is your ROI signal. When the first skill runs reliably with one round of review or less, add the next workflow. Never expand before the first skill is stable.

Expected time savings by workflow type

AI-assisted times reflect elapsed time with context loaded and one light steering pass, not including CD review, client approval, or distribution. Savings vary significantly based on how developed the knowledge layer is. These are directional estimates from agencies with established context layers.

Agency WorkflowCurrent TimeWith Agency AIOSNotes
Brief distillation — raw client input → actionable creative brief2–4 hrs20–40 minHighest frequency, fastest payback. Best starting pilot.
Brand calibration — initial session setup for new campaign1–2 hrs10–20 minSavings compound; every subsequent brief on this account benefits.
Client status report — weekly or campaign milestone1–2 hrs10–15 minHigh frequency, consistent structure. Immediate quality win.
Pitch structure — first-draft narrative framework4–8 hrs30–60 minRequires established pitch archive and account context. Strong second skill.
Scope and estimate draft2–4 hrs20–40 minRequires process library with scope templates and line-item rationale.
New hire onboarding — one client account2–4 weeks shadowing2–3 days readingThe highest ROI output of a mature knowledge layer. Requires all five knowledge assets built.
Competitive research brief8–16 hrs1–2 hrsStrong use case for pitch prep and new business development.
Campaign debrief synthesis — post-project learnings2–4 hrs20–30 minThe investment that makes every subsequent project on this account more efficient.
Knowledge layer build — one client account (initial)4–8 hrs one-time5–10 min per update ongoingA one-time investment that compounds across every project and every team member for the life of the account.

Strongest early savings: brief distillation, brand calibration, client status reporting. Strongest long-term return: new hire onboarding and knowledge layer maintenance.

06 — Use cases

Use AI where the workflow is frequent and has a clear human decision gate.

The highest-leverage AI applications in an agency sit in the space between brief intake and final delivery: distilling, drafting, calibrating, and summarizing. Anything where a competent person could produce a useful first draft in 30 minutes, given the right context, is a good AI candidate.

Agency WorkflowGood AI UseHuman Decision Gate
Brief distillationRead raw client input: deck, email thread, stakeholder notes. Extract the actual creative problem, audience, constraints, and success criteria. Output a one-page creative brief.CD approves the distilled brief before any creative work begins, alongside the account lead on accounts where they hold the primary client relationship. This is the most important gate in the process.
Brand calibrationRead the client brand profile and relevant approved work samples. Generate an initial calibrated draft for a new campaign, demonstrating the brand's voice before the team starts producing at volume.CD or senior writer approves the brand calibration example before it's used as the reference for the campaign.
Copy drafting — all formatsWith client context and brand profile loaded, produce first-draft copy across formats: headlines, body copy, social, email, presentation narrative. Calibrated to the client's voice, not generic AI output.CD or senior writer reviews and refines before anything goes to the client. Output should need refinement, not rebuilding.
Pitch structureFrom an approved brief and account context, produce a first-draft narrative structure for a client presentation: the argument, the sequence, the evidence. Not the design. The logic.CD or strategist reviews the structure before creative development begins, whoever owns the pitch's narrative logic on this account. Once the structure is approved, production can proceed in parallel.
Client status reportRead current project state, milestone log, and open items. Produce a structured client status update in the agency's format: what's done, what's next, what needs client input.Account lead reads and approves before sending. Should require a 5-minute check, not a rewrite.
Scope and estimateFrom a brief and the agency's process library, produce a project scope with line items, rationale, and dependencies. The first draft that the account lead refines based on relationship context.Account lead and CD review before the estimate goes to the client, with the agency owner added for larger engagements. Financial and relationship judgment stays with humans.
Competitive context researchResearch the competitive landscape for a pitch or new campaign: what are competitors doing, what's the category conversation, what's the white space. Produce a structured brief, not a raw data dump.CD or strategist reviews for strategic relevance before using in pitch development or brief framing.
Creative review documentationFrom review session notes or a recorded discussion, produce structured creative review notes: what was approved, what needs revision, what the direction is, what the next round should address.Whoever ran the review, CD, design lead, or account lead, confirms the notes before they're distributed. Decisions in creative review have downstream consequences; verify they're captured correctly.
New hire account contextRead the client knowledge layer and produce a new-hire briefing document: who the client is, how they communicate, what they care about, what's worked, what hasn't, what to avoid.Account lead reviews for accuracy and currency before the new hire uses it as their primary onboarding reference.
Campaign debrief synthesisRead post-project notes, client feedback, and CD review notes. Produce a structured debrief: what worked, what didn't, what to do differently, and what should be encoded into the knowledge layer for this account.CD reviews and approves before the debrief learnings are encoded into the account knowledge file. Wrong learnings compound just as fast as right ones.
Meeting and call follow-upFrom meeting notes or transcript, produce structured follow-up: decisions made, action items with owners, open questions, next steps. Route to the meeting owner before it goes to the client.Account lead or meeting owner reviews before external distribution. Action items and decisions need human verification.
What AI should not be used for (and why)
WorkflowWhy to keep humans here
Final creative judgment on strategyStrategy is about choosing what to prioritize and what to sacrifice. AI can map options; it can't make the judgment call that defines what your agency believes.
Sensitive client conversationsWhen the relationship is at stake (a difficult conversation about direction, a budget negotiation, a scope dispute), human presence, tone, and judgment are irreplaceable.
New business first contactThe first impression of your agency is a human impression. AI-drafted outreach, however good, reads as AI-drafted to the people you most want to impress.
Final approval on anything client-facingThis isn't a limitation to work around; it's a feature. Human accountability for client output is not optional.

"One skill that the CD approves in a single pass is worth more than five skills that each need rebuilding."

Section 05 — Implementation roadmap
07 — Templates

Starting points for your knowledge layer and operational structure.

These templates are designed to be copied into your agency's project folders and adapted based on what you learn in the first real pilot. The client brand profile is the most important. Start there before building anything else.

Brand calibration Copy drafting New hire onboarding

Client brand profile

The most-referenced document in your knowledge layer. Every AI output that touches a client loads this file first. Without it, every prompt starts cold and sounds like the category, not the client.

View template
Client:
Account lead:
Last updated:

BUSINESS CONTEXT
What they do, who their audience is, what they're
trying to achieve, and where they sit against
competitors.

DECISION-MAKERS AND SENSITIVITIES
Who signs off. What they've reacted badly to
before. What's off-limits.

VOICE AND TONE IN THREE SENTENCES
Write how this client sounds in three sentences.
Not adjectives. Actual sentences in their voice.

RHYTHM PATTERN
Describe the sentence-length rhythm. Does it open
long and tighten? Short declarative followed by
elaboration? Use examples.

VOCABULARY SET
Words they use: [list]
Words they would never use: [list]
Phrases that are characteristic: [list]

CALIBRATION EXAMPLES
Link or paste 2 to 3 pieces of approved work that best
represent the voice. What makes each one right:

COMMON CD CORRECTIONS
What does the CD always change in outputs for
this client? [specific corrections, not adjectives]

RECENT UPDATES
[date]: [what changed and why]
Pitch structure Brief distillation Status report New hire onboarding

Client account context

The relationship intelligence that usually lives in one account lead's head. When any team member loads this before touching an account, coverage is consistent whether the account lead is available or not.

View template
Client:
Account lead:
CD:
Last updated:

WHO THEY ARE
Business context, category, market position, and
what they're actually trying to accomplish, not
what the brief says, what's really going on.

WHAT WE'VE MADE FOR THEM
Summary of work to date. What has landed well.
What hasn't.

HOW THEY MAKE DECISIONS
Who actually decides? How does feedback arrive?
What triggers rejection? What triggers approval?

WHAT THE CD ALWAYS SAYS ABOUT THEM
The unwritten brief. What every team member should
know before touching this account.

RELATIONSHIP HISTORY
Key moments, pivots in direction, difficult
conversations. Written for a new team member.

CURRENT PRIORITIES
What they're focused on right now. What they're
anxious about. What would make this a great year.

RED FLAGS
Things to watch for. Topics to handle carefully.
All production sessions Creative review CD sign-off

Brief distillation output

The approved brief is the contract every downstream output is built against. When it's clear, copy, pitch, and scope all pull in the same direction. When it's fuzzy, the CD gets called in to reconcile the divergence.

View template
Client:
Campaign:
Account lead:
CD approved: [ ] Yes  [ ] Pending

THE ACTUAL PROBLEM
In one sentence: what does the client really need
to solve? (Not the symptom they described, but the
underlying problem.)

THE AUDIENCE
Who are we talking to? What do they already
believe? What do we need to change?

CREATIVE CONSTRAINTS
What we cannot do. What we must do.
Non-negotiables from client, brand, category.

SUCCESS CRITERIA
How will we know this worked? Measurable if
possible. Client-verifiable.

THE BRIEF IN ONE SENTENCE
If this campaign works, [audience] will
[do/believe/feel something specific].

WHAT WE DON'T KNOW YET
Questions to resolve before creative work begins.
Knowledge layer update Brand profile refresh Campaign kickoff

Campaign debrief

This is how the knowledge layer compounds. Every insight captured here (what the client responded to, what the CD changed, what to do differently) gets encoded back so the next campaign on this account starts measurably smarter.

View template
Client:
Campaign:
Completed:
Debrief owner:

WHAT WE MADE
Brief summary of deliverables and what the
campaign was meant to accomplish.

WHAT WORKED
Specific creative, strategic, or process decisions
that produced good outcomes. With evidence.

WHAT DIDN'T WORK
Honest account of what failed, what we'd do
differently, and why it went wrong.

WHAT WE LEARNED ABOUT THIS CLIENT
New insights about how they make decisions, what
resonates. Specific and usable for the next brief.

BRAND UPDATES
What should be added to or revised in the client
brand profile based on this campaign?

KNOWLEDGE LAYER ACTIONS
[ ] brand profile  [ ] account context
[ ] pitch archive  [ ] process library
First-pilot checklist
08 — Common mistakes

Most agency AI failures are context failures, not model failures.

The quality of what AI produces is almost entirely a function of the context it has access to. Generic outputs don't mean AI can't help; they mean the knowledge layer isn't built yet. Here is where agencies consistently go wrong.

Building skills before building the knowledge layer

A voice calibration skill with no brand profile produces generic AI output. A brief distillation skill with no account context produces generic briefs. The knowledge layer has to come first. Skills are only as good as the context they read.

Brand profiles made of adjectives, not examples

"Bold, human, and direct" is not a brand profile. It's an aspiration. A useful brand profile contains actual sentences in the client's voice, specific vocabulary choices, rhythm patterns, and real examples of approved work. Without that, AI produces whatever "bold, human, and direct" means to it, which is not what it means to your client.

No review gate before client-facing output

AI output that goes to a client without a human reading it first is not an efficiency gain. It's a liability. Every skill that produces client-facing output needs an explicit, named human review gate, built into the process, not assumed to happen naturally.

Trying to automate creative judgment

Brief distillation, voice calibration, scope drafting, and status reporting are all appropriate AI candidates because the quality criteria are checkable. "Is the creative direction right?" is not. AI produces material for judgment; it doesn't replace it. The moment a skill tries to make a strategic creative decision, it has overreached.

Scaling before the first skill is reliable

If the brief distillation skill still requires two rounds of CD revision every time, adding voice calibration and pitch structure on top creates more overhead, not less. One skill that the CD approves in a single pass is worth more than five skills that each need rebuilding.

Treating a tool stack as an operational system

Giving everyone Claude accounts is not Agency AIOS. It is individuals using AI differently, producing inconsistent outputs, with no shared context and no compounding value. AIOS is the infrastructure on top of the tools: the knowledge layer, the skills, the review gates. Without that infrastructure, more tools just accelerate inconsistency.