How to build a brand spec your AI agents can actually enforce

Most multi-location operators have a brand voice document. Almost none have a brand spec. The difference matters more than it sounds, because the brand voice document is something a copywriter reads. The brand spec is something an AI agent enforces. If your AI marketing tools are generating output that drifts from your brand — and most are — the gap between "we have a brand voice document" and "we have an AI-enforceable brand spec" is the reason.

This guide walks through what an AI-enforceable brand spec contains, how to author one in 4-8 hours starting from your existing brand voice document, and how to validate it works. It does not require any specific AI tooling — the spec is a portable artifact that any production system can consume.

The difference: voice document vs. brand spec

A brand voice document is prose. It says things like "our voice is direct and accountable" or "avoid jargon" or "use first names with customers." A copywriter can read it, internalize the spirit, and produce on-brand output. A model cannot, reliably, because the rules are stated abstractly and the gate (the model that scores output before it publishes) has nothing to score against numerically.

A brand spec is structured. Same intent, but expressed as machine-checkable rules: tone dimensions with weights, a forbidden-phrase table with replacements, a tone matrix with score thresholds, per-vertical compliance overlays loaded at runtime. A model — specifically the brand-voice gate model that runs before publish — can score every output against the spec on a 0.00-1.00 scale per dimension, sum the scores against weights, and decide auto-publish vs queue-for-review without a human in the loop except for borderline cases.

What an AI-enforceable brand spec contains

Six sections. Each is required; skipping any one means your AI outputs will drift along the missing dimension.

1. Core voice principles

3-7 principles stated as constraints, not aspirations. Each principle is concrete enough that a model can flag a violation. Bad: "our voice is approachable." Good: "no contractions in body copy; contractions permitted in CTAs and subject lines."

If you cannot express a principle as a constraint, it is not enforceable. Either rewrite it as a constraint or move it to the brand voice document where humans can interpret it.

2. Forbidden phrase table

A two-column table: phrase to forbid, phrase to replace with. 15-30 entries typical. The forbidden side covers marketing puffery ("comprehensive," "best-in-class," "world-class"), generic AI-narrator openings ("Welcome to," "In today's article we'll explore"), and category-specific risk phrases (your industry's regulatory hot-buttons).

The gate's forbidden_phrase_check dimension scores 1.00 if no forbidden phrases are detected, and reduces by a fixed penalty per detection. Replacement guidance helps the producer model regenerate without re-falling-into-the-trap.

3. Tone matrix with weights

The tone matrix is the heart of the spec. 5 dimensions with weights summing to 100%, scored 0.00-1.00 per output by the gate model:

• claim_compliance — whether claims are supported by stated facts; no unsupported superlatives. Weight typically 20-30% (higher for regulated verticals).
• forbidden_phrase_check — absence of phrases from section 2. Weight typically 10-20%.
• tone_match — direct + technical + accountable + appropriate warmth; no over-apology. Weight typically 20-30%.
• regional_appropriateness — adapts to context (location, vertical, jurisdiction) without misfiring. Weight typically 10-20%.
• schema_adherence — length target, structure, signature pattern, formatting rules. Weight typically 15-25%.

Weights vary by vertical. Healthcare operators typically weight claim_compliance higher (regulatory exposure). Service-business operators typically weight tone_match higher (warmth signals matter more in customer-facing copy). Pick weights deliberately — they are the dial that determines what your AI is willing to publish.

The gate aggregates scores against weights. Above the auto-publish threshold (typically 0.90 for review responses, 0.92 for blog/page copy), output publishes. Below, it routes to editorial governance for human review. This threshold is the second most important number in your spec; calibrate it after your gate has scored 30-50 outputs you have manually rated as ship-or-not-ship.

4. Per-channel adaptations

Different channels have different conventions. A LinkedIn long-form post follows different rules than a Reddit comment than a Google Business Profile review response. The brand spec must encode the per-channel deltas.

Typical channels: LinkedIn long-form, LinkedIn comment, Twitter/X thread, Reddit top-level post, Reddit comment, autoresponder email, cold outreach email, GBP review response, location page copy, YouTube long-form script, YouTube short script, paid ad copy.

Per-channel entries specify: word-count target ranges, hook structure rules (e.g., "first 2 lines must contain the comparative reframe for LinkedIn long-form"), CTA pattern (e.g., "diagnostic + quiz, not direct upsell" for cold-channel posts), and any tone deltas (e.g., "Reddit voice tolerates contractions").

5. Per-vertical compliance overlays

If your operation crosses multiple regulatory verticals (healthcare locations + restaurants + cannabis + multi-state lending), each vertical needs its own overlay loaded at runtime. The overlay is not optional content — it is enforced before the rest of the spec runs.

Healthcare overlay: HIPAA prohibitions on patient identifiers, treatment specifics, visit details. State medical board advertising rules (avoid superlatives, unsupported claims like "painless"). Specific terminology preferences ("doctor" vs "physician" vs "provider").

Restaurant/multi-unit food overlay: FTC chain rule (21 CFR 101.11) at 20+ locations under same name. FDA Menu Labeling Final Rule for caloric disclosure. State-level menu labeling overlays. Allergen disclosure language.

Cannabis MSO overlay: per-state advertising rules (recreational vs medical varies). No federal-level claims. Age-gating language. Section 280E tax-deductible-language carve-outs.

Multi-state lender overlay: TILA / Regulation Z disclosure language. State-specific lending licensing references. APR vs interest rate precision. Equal-credit-opportunity language.

The compliance overlay routes borderline outputs to the compliance officer queue, which is a different routing target than the editorial governance queue. Most multi-vertical operators conflate these queues; do not. The compliance officer reviews against regulatory exposure; the brand director reviews against voice drift. They are different people with different SLAs.

6. Voice signature phrases

5-10 phrases that signal your voice when used. Use sparingly to avoid pattern detection (formulaic feel) but use them intentionally to reinforce brand recognition.

For example, Completions uses: "operator-grade," "load-bearing," "the strategic decision is which X" (comparative reframe pattern), "the producer that drifted is the worst evaluator of whether it drifted" (gate-architecture vivid).

The signature phrases section is what gives your AI outputs voice continuity across surfaces. Without it, even on-brand outputs feel like they came from different writers.

Step-by-step authoring process

Start from your existing brand voice document. The work is translating, not inventing.

Step 1: Extract constraints from prose (1-2 hours)

Open your brand voice document. For every paragraph, ask: "what concrete rule is this paragraph trying to express?" Write down the rule in constraint form.

Example: a paragraph saying "we want to come across as warm but professional" becomes the constraints "no over-apology (cap of 1 'sorry' or 'apologize' per response)" and "include a personal-name-or-role signature (e.g., '— The {team_name} team') for customer-facing responses."

Some paragraphs will not produce constraints. They are aspirational, not enforceable. Move them to a "principles for human reviewers" section. Do not try to force-fit aspirational language into the spec.

Step 2: Build the forbidden phrase table (30-45 minutes)

List every phrase your team has flagged in past content reviews as off-brand. Add the standard marketing-puffery list (comprehensive, best-in-class, etc.). Add your industry's regulatory hot-buttons (any phrase a compliance review has flagged in the last year).

For each, write a replacement. The replacement is what the producer model regenerates with after the gate flags the original.

Step 3: Score 30-50 historical outputs to calibrate weights and threshold (2-3 hours)

Take 30-50 historical pieces of content (review responses, location page copy, social posts) — a mix of "shipped clean" + "shipped after editing" + "didn't ship." For each, manually score on the 5 tone dimensions (0.00-1.00 each). Then compute the aggregate against your draft weights.

Compare your aggregate scores to the actual ship/edit/reject decisions made on each piece. If your aggregate scores cluster correctly (clean-ships high, rejected-ones low), your weights are good. If they don't, adjust the weights and rescore until they cluster correctly.

The auto-publish threshold falls out naturally: it is the aggregate score below which you would have pulled the piece for review. Typical thresholds end up between 0.88 and 0.94 depending on stakes.

Step 4: Add per-channel adaptations (1-2 hours)

For each channel your AI agents will publish to, write the channel's deltas. Most are short — word-count target, hook rule, CTA pattern, signature pattern. The full spec for any channel rarely exceeds 15-20 lines.

Step 5: Add per-vertical compliance overlays (varies by complexity)

If you operate across multiple verticals, this step is the longest. Each vertical's overlay is its own document referenced from the main spec. Get a compliance attorney to review each overlay before it goes live; the cost of a wrong rule getting auto-applied across 50+ locations is high.

Step 6: Version + commit + lock (15 minutes)

The brand spec is a version-controlled artifact. Commit v1.0 to git (or whatever version control your brand team uses). Tag it. Reference the tag from your AI production pipeline. Future edits require version bumps + a backfill plan for any in-flight outputs that were generated against an earlier version.

Validation: does the spec actually work?

Three tests, run weekly for the first month and monthly thereafter.

Test 1: Aggregate-score distribution

The gate scores every output. Plot the distribution of aggregate scores across the last week. A healthy distribution shows ~85-95% of outputs above the auto-publish threshold and ~5-15% routed to review. If 100% auto-publish, your spec is too lax. If <50% auto-publish, your spec is too strict and the producer model is fighting it.

Test 2: Per-dimension distribution

For each of the 5 tone dimensions, plot the per-dimension score distribution. A healthy spec produces tight distributions on each dimension (most outputs scoring 0.85-0.95 per dimension). A wide distribution on one dimension means the producer is unstable on that dimension; either tighten the producer prompt or flag the dimension for human review even when aggregate passes.

Test 3: Manual override rate

Track how often a human reviewing a "queued for review" output would have made a different decision than the gate. The override rate should be <10% within 4 weeks of spec deployment. Higher means the spec is mis-calibrated and needs adjustment.

Maintenance pattern

Brand specs are not write-once. Update them when:

• A new vertical enters scope (add overlay)
• A new channel enters scope (add adaptation)
• A new forbidden phrase surfaces in production review (add to table)
• The producer model upgrades (rescore historical outputs; recalibrate weights if scores shift)
• A regulatory rule changes (update overlay; bump version; backfill in-flight outputs)

Quarterly: review the per-dimension score distributions. Tighten or loosen weights based on what's drifting.

Annually: full recalibration with a fresh 30-50 historical output set. Aggregate scoring drifts as your brand evolves; recalibration keeps the gate aligned.

What this gets you

A brand-voice gate that catches drift before it publishes, scoring every AI-generated output against constraints your brand team owns. An audit trail (every gate decision logs the per-dimension scores), so when a customer asks "why did your franchise location post X" you have the evidence. A compliance defensibility argument (every output passed an enforced compliance overlay), so when a regulator asks "how did this claim get published" you have the answer.