What Is an AI Foundations Document?

An AI Foundations document is a short, structured reference document that tells an AI system what your company is, how it communicates, and what it produces, so the AI’s outputs reflect your company rather than a generic business.

Without it, AI produces generic outputs. With it, AI produces outputs that sound like your company, use your industry’s vocabulary, and meet your quality standards.

The five types of AI Foundations document

1. Company overview (150 to 200 words)

What the company does, who its customers are, what makes it distinctive, and what its operational context is. This is the broadest Foundations document and is loaded into every shared workspace.

What it contains:

• Company name and primary business (what it does in plain language)
• Customer description (who buys from the company and in what context)
• Distinguishing characteristics (what makes this company different from a generic company in its sector)
• Revenue scale and team size (for calibrating the appropriate level of operational complexity in outputs)
• Geographic context if relevant

What it does not contain: marketing language, mission statements, or aspirational descriptions. The company overview is operational, not promotional.

Sample sentence: “Pacific Crest Distribution is a 45-person HVAC parts wholesaler based in Phoenix serving commercial mechanical contractors and facilities management companies across the Southwest. The company has supplied the same contractor base for 22 years and is known for same-day availability on standard stock items.”

This one sentence tells AI more about how to calibrate its outputs than a three-paragraph mission statement.

2. Brand voice guide (200 to 300 words)

How the company communicates: the tone, the vocabulary conventions, the register for different audiences, and the specific phrases the company uses and avoids.

What it contains:

• Tone description (specific adjectives that are accurate, not aspirational: “direct and solution-focused” not “warm and collaborative” if the company is actually direct)
• Vocabulary to use (the specific terms the company uses for its products, services, processes, and customer relationships)
• Vocabulary to avoid (the generic business language, the competitor’s terminology, the jargon that does not reflect how this company actually communicates)
• Audience-specific calibration (how the tone shifts for key accounts vs transactional accounts, for internal staff vs external customers)

What it is not: a copy of the brand guidelines document from the marketing team. The brand voice guide for AI purposes is operational: it specifies the specific words and phrases that produce company-appropriate AI outputs, not the design principles of the brand identity.

3. Customer communication standards by tier (200 to 300 words)

The specific communication conventions for each customer tier: how the company communicates differently with key accounts, growth accounts, and transactional accounts.

What it contains:

• Tier definitions (what defines each tier: revenue, strategic importance, relationship depth)
• Tone calibration per tier (how the directness, formality, and relationship warmth shifts by tier)
• Structure conventions per communication type (the opening, the body, the closing for routine updates, exception notifications, and relationship communications)
• Specific examples of phrases that reflect each tier’s conventions

Why tier-specific standards matter: the company that sends the same tone of back-order notification to the commercial contractor who represents 12% of revenue and the transactional account who places one order per year is missing a significant relationship management opportunity. The AI that knows the difference produces the right calibration without the account manager having to specify it each time.

4. Exception and vocabulary guides (150 to 250 words each)

Function-specific reference documents that define the vocabulary for specific operational situations or specific industry contexts. Each guide covers one domain.

Examples:

Exception vocabulary guide (for customer service): the specific language for the most common exception types (back-orders, delivery delays, quality issues, billing discrepancies). For each exception type: the preferred framing, the disclosure convention, the resolution language, and the vocabulary that reflects sector expertise rather than generic customer service language.

Regulatory vocabulary guide (for healthcare billing): the precise regulatory terminology for the company’s payer communication context: the appeal argument structures, the denial code response vocabulary, the compliance language that distinguishes a professionally written appeal from a generic one.

Technical vocabulary guide (for manufacturing): the specific terminology for the company’s quality documentation, production scheduling communications, and customer technical specifications: the vocabulary that signals manufacturing expertise to the customer or regulator reading the output.

Each vocabulary guide is narrow in scope (150 to 250 words per domain) because narrow scope produces more consistent AI referencing than broad, comprehensive vocabulary documents.

5. Workflow specifications (150 to 200 words each)

The operational specifications for the AI-assisted workflows in each function: the inputs the team member provides, the output format AI produces, the quality standards the output must meet, and the human review gate.

What it contains:

• Workflow name and trigger (what initiates this workflow)
• Input format (what the team member provides, described as specifically as possible to produce consistent inputs)
• Output format (the structure, length, and format of the AI’s output)
• Quality standards (the specific criteria the output must meet before being used)
• The human review gate (what the team member checks before the output is sent or filed)

Why workflow specifications matter: when the workflow specification is in the shared workspace, any team member who opens the workspace and provides the inputs receives a correctly structured, quality-appropriate output, regardless of their AI experience level. The workflow specification makes the quality consistent. Without it, quality varies with who is prompting.

How to build an AI Foundations document

Who builds it

The people who know how the company’s work should sound and be structured: not AI, and not the IT team.

• The company overview is built by the managing director or COO
• The customer communication standards are built with the customer service lead or account management director
• The vocabulary guides are built with the function leads who know the professional vocabulary of their domain
• The workflow specifications are built with the team member who runs the workflow most frequently

How long it takes

The build process

1. The practitioner or AI system owner prepares a structured interview guide for each document type: the specific questions that extract the operational detail the document needs.

2. The interview session runs for 45 to 60 minutes with the relevant function lead. The practitioner asks the prepared questions, probes for specificity (not “professional tone” but “what specific words characterise our tone: give me an example sentence”), and documents the answers in real time.

3. The practitioner or AI system owner produces the first draft of the document from the session notes. The draft is 200 to 400 words in the appropriate format.

4. The function lead reviews the draft, identifies any vocabulary that does not reflect actual practice, and approves.

5. The document is uploaded to the shared AI workspace. The AI system owner tests it against a sample workflow to confirm it produces the expected calibration improvement.

The AI Foundations document vs the AI Foundation — the distinction

The AI Foundations document (singular) is a single context document: the company overview, or the customer communication standards, or the payer vocabulary guide.

The AI Foundation (as used in the Phos engagement framework) is the full collection of context documents, workflow specifications, and workspace configuration that together make AI produce company-specific outputs consistently.

The Foundation is the complete system. The Foundations document is one component of it.

The four-phase Phos engagement begins with Phase 1: AI Foundations (the sprint that produces the complete Foundation for the company).

This phase includes building all the relevant Foundations documents, configuring the shared workspace, and testing the documents against the team’s primary workflows.

The AI Foundations document is the individual deliverable. The AI Foundation is the operational system that the documents, the workspace configuration, and the trained team produce together.

For context on how the Foundations documents relate to the broader context pack, see what is an AI context pack and how to give AI full business context.

Common questions on AI Foundations documents

”Can AI help build the AI Foundations documents?”

Yes, with the right process. The practitioner or AI system owner runs the structured interview session with the function lead and produces rough notes.

These rough notes can be input to the AI workspace with a structured format request: “Turn these notes into a [document type] for a [company description].”

The resulting AI-produced draft requires the function lead’s review and correction, because AI cannot know what the company’s actual communication conventions are.

The AI helps with the structure and the format. The human provides the specific content that makes the document accurate.

”What is the difference between an AI Foundations document and a system prompt?”

A system prompt is the instruction text that precedes every AI session and specifies how AI should behave. An AI Foundations document is a content document: a reference that provides context about the company, its work, and its standards.

In practice: the system prompt is the instruction (“When I provide customer notification data, produce a notification in the format specified in the Customer Communication Standards document loaded in this workspace”).

The AI Foundations document is the reference the system prompt points to.

Both are part of the complete Foundation. The Foundations documents contain the operational knowledge. The system prompts specify how to apply it.

”How often should AI Foundations documents be updated?”

At minimum, reviewed quarterly and updated when: the company’s communication standards change, new vocabulary conventions are adopted, the improvement loop identifies a quality gap that traces to Foundation inaccuracy, or a new function or product line requires new vocabulary.

The AI system owner’s weekly improvement loop review includes the question: “Did any output this week fail because the Foundation document did not accurately reflect how we communicate?” If yes: the document is updated before the next week’s sessions.

The AI Foundation build is Phase 1 of every Phos engagement

An AI Foundations document is a short, structured context document that tells AI what your company is, how it communicates, and what good work looks like in your specific operational context.

Without the Foundation, AI is a capable general-purpose tool. With it, AI is a capable company-specific tool — one that sounds like your company, uses your industry’s vocabulary, and meets your quality standards. Building the Foundation takes 4 to 8 hours. It is the most important single investment in an AI implementation.

Phos AI Labs builds the complete AI Foundation as Phase 1 of every engagement: the context pack built, configured, and tested before the first team member is trained. Thirty minutes, no deck. Start here.