Key takeaways
Company context is a reviewed description of a business that an AI assistant reads before it works: what the company does, how it approaches its market, what it offers, who it serves and how it communicates. Without it, even a well-written prompt produces generic research and messages. With it, every supported task starts from the same foundation instead of a fresh brief, and the team corrects one description rather than every prompt.
- Generic AI output is usually a context problem rather than a prompt problem: an assistant cannot know your audience, offer or voice unless it is given them.
- Describe the company once, in a structured and reviewed form, and reuse it across tasks instead of pasting a brief into every prompt.
- Company context organises working information; it is not the same as training new model weights for each customer.
- Company facts change, so context needs an owner, a review habit and a simple way to refresh it.
- People still decide which supported outputs are approved for reuse, and reuse depends on the workflow.
What company context means
Company context is the working knowledge an assistant needs to do a task the way your company would: what the business does, how it approaches the problem it solves, what it sells, who it sells to and how it speaks. A new colleague absorbs this over weeks. An AI assistant starts every conversation without it, unless it is written down and supplied.
It is different from a prompt. A prompt describes one task: research this account, draft this email, summarise this call. Context describes the company the task belongs to, so it stays the same from one task to the next. Good work needs both: a clear instruction and a reliable picture of the business behind it.
Why AI output turns generic without it
Language models learn from broad public text, so when a detail is missing they tend to fill the gap with the most typical answer for your category. Ask for a cold email without saying who you sell to, what problem you solve or how you sound, and you get a cold email that could have come from any company.
Messages that could come from anyone
Claims and phrasing that would fit every company in your category.
Research that misses the angle
Accurate facts about an account, with no link to what you offer it.
A different voice in every draft
A tone that shifts with whoever wrote the prompt that day.
The same brief, pasted again
Time spent restating the company before every task.
The usual workaround is a long brief pasted into every prompt. It works for a while, then drifts: different people keep different versions, a product change reaches some briefs and not others, and nobody is sure which description is current. The output inherits that inconsistency.
What good company context includes
Useful context is specific, structured and short enough to review. Write down the parts below in plain words, one part at a time, so each can be checked and refreshed on its own.
The company
What you do, for whom and in which markets.
The approach
How you solve the problem and what makes your way of working distinct.
The offer
Your products or services, who each is for and what each is not.
The conviction
The point of view about your market that the company stands behind.
The audience
The roles and companies you serve, and what they care about.
The message and expression
The few points every piece should carry, the tone, and words to use or avoid.
In Pebbles Ai, Company Intel organises these as Business knowledge (Company, Approach and Offerings) and Brand voice (Conviction, Audience, Message and Expression). It is built from the company website, company information, uploaded files and pasted text, and every assistant in the workspace works from the result.
Keep each part factual and current. A short, accurate description beats a long one with outdated claims, because the assistant cannot tell which sentences are still true.
Context is not model training
Providing documents or describing your company is not the same as changing a model's weights. Training adjusts the model itself. Context is information the assistant reads when it works, so it can be corrected the moment the business changes.
Pebbles Ai keeps two kinds of working information. Company Intel holds the reviewed business description. Agent memory keeps short facts that you state or confirm in conversation, at three levels: Your memory, Project memory and Company memory. Company memory reaches colleagues' assistants, project memory stays in chats filed under that project, and switching off "Remember my conversations" in Settings stops the assistants reading and saving memories. Documents, CRM records and drafts are not copied into memory.
Neither is a dedicated model trained for each customer. The comparison below sets the four ideas side by side.
Review what becomes reusable
Company information changes: offers are renamed, audiences shift and messaging is refreshed. Treat context as a document with an owner, not a one-off setup.
Admins can refresh Company Intel with "Add what's missing" or "Start fresh", and create, edit and delete entries as the business changes. Check extracted answers against their source, correct the brief and review outputs before approving them for reuse.
The Library gives supported content a review state, Pending review or Approved, and the Approved Library is available on Team and above. It does not guarantee that every workflow automatically retrieves every approved asset, so check how the workflow you rely on uses it.
Prompt, company context, memory and training compared
Four ideas that are often mixed up. Only the first three are things a team supplies while it works.
| Prompt | Company context | Agent memory | Model training | |
|---|---|---|---|---|
| What it holds | The instruction for one task | A reviewed description of the company, its offer and its voice | Short facts you state or confirm in conversation | Patterns learned from a large body of data |
| How long it lasts | One request | Until someone edits or refreshes it | Until you correct it, ask the assistant to forget it or switch memory off | Until the model is retrained |
| Who it reaches | The person who wrote it | Every assistant in the workspace | You, one project's chats or colleagues' assistants, depending on the level | Everyone who uses that model |
| How you change it | Write a new prompt | Admins edit entries or refresh with "Add what's missing" or "Start fresh" | Ask the assistant in plain words to correct or forget a fact | Retrain or fine-tune the model |
| In Pebbles Ai | Assistant chat | Company Intel | Agent memory | Model weights do not change for each company |
Context and memory are working information the assistant reads. Neither changes a model's weights.
Worked example:
grounding an outreach brief
A team is preparing a campaign and wants its messages to reflect the same audience, offer and voice, whoever writes the prompt.
- Add the company website, company information and any files or pasted text to build Company Intel.
- Review the Business knowledge (Company, Approach and Offerings) against its source and correct anything out of date.
- Review the Brand voice (Conviction, Audience, Message and Expression) rather than accepting the extraction unchanged.
- Build the campaign in the campaign builder, choosing a persona for each piece, so the drafts start from the reviewed description instead of a pasted brief.
- Review the drafts, then move a supported output into the Approved Library for later use.
Approved reuse remains workflow-dependent. This example does not imply universal retrieval, automatic freshness or full document-management functionality.
Common mistakes
Accepting the first extraction unchanged
Extraction is a starting point. Check every answer against its source before the team relies on it.
Writing plans as facts
Describe what the company offers today. The assistant repeats whatever the context claims.
One long brief for everything
Keep facts, audience and voice apart so each part can be reviewed and refreshed on its own.
Nobody owns it
Without a person responsible for keeping it current, context drifts as the business changes.
Treating approval as automatic reuse
Approving an asset does not mean every workflow retrieves it. Check how the workflow you depend on uses the Library.
Practical checklist
- Are the company facts current and attributable to a useful source?
- Does the description say what the company offers today, not what it plans?
- Is the intended audience described precisely enough for this task?
- Has the team reviewed the voice and message rather than accepting extraction unchanged?
- Is one person responsible for keeping the company description current?
- Is the output approved for this use, and does the chosen workflow support Library reuse?
Key terms
- Company context
- A reviewed description of a business that an AI assistant reads before it works: what the company does, who it serves, what it offers and how it communicates.
- Prompt
- The instruction for one task. It changes with every request, while company context stays the same between them.
- Company Intel
- The admin area in Pebbles Ai that holds Business knowledge and Brand voice, built from the company website, company information, uploaded files and pasted text.
- Business knowledge
- The factual side of Company Intel, organised as Company, Approach and Offerings.
- Brand voice
- How the company communicates, organised in Company Intel as Conviction, Audience, Message and Expression.
- Agent memory
- Short facts the Pebbles Ai assistants save when you state or confirm them, kept at three levels: Your memory, Project memory and Company memory.
- Model training
- Changing a model's weights with new data. Supplying company context or memory does not do this.
- Approved Library
- Supported content a team has reviewed and approved for reuse, available on Team and above.
Common questions
What is company context in AI?
Company context is the information an AI assistant reads about your business before it starts a task. It covers what you do, what you sell, who you sell to and how you sound. It is not the instruction for one task, and it is not model training. Your team should check it and keep it up to date.
Why does AI write generic content for my company?
Usually because it does not know enough about you. When an assistant lacks your audience, offer or tone, it fills the gaps with the most typical words for your industry. The fix is to give it a clear, checked description of your company once. Then add a clear instruction for each task.
How do I give an AI assistant context about my company?
Write a short company profile once and let the assistant read it before every task, instead of pasting a brief each time. In Pebbles Ai, Company Intel builds this profile from your website, company information, uploaded files and pasted text. Every assistant in the workspace works from it. Check and correct it before you rely on it.
What should a company profile for an AI assistant include?
Include what you do and for whom, how you solve the problem, what you sell, who your buyers are, the view of your market you stand behind, your key messages and your tone. Keep it short and factual. Pebbles Ai splits these into Business knowledge (Company, Approach and Offerings) and Brand voice (Conviction, Audience, Message and Expression).
How do I stop rewriting my company brief for every AI task?
Save the brief once as shared company context, so every task starts from it. In Pebbles Ai, Company Intel holds the checked description, and agent memory keeps short facts you state or confirm in chat. When something changes, update the description once instead of fixing every prompt.
How should I check company information extracted by AI?
Compare each answer with where it came from, such as a web page or a file. Check that offers, audiences and claims are true today. Remove plans written as facts, and add anything the extraction missed. In Pebbles Ai, admins can edit or delete entries, or refresh with "Add what's missing" or "Start fresh".
How do I write outreach that matches our company voice?
Describe your voice before you write: who you write for, the main message, the tone, and words to use or avoid. Check that description, then draft from it. In Pebbles Ai, the campaign builder lets you choose a persona for each piece, and Write with AI in Smartbox uses your brand voice. Read each draft before it goes out.
Is giving an AI company context the same as training a model?
No. Training changes the model itself. Company context is information the assistant reads while it works, so you can correct it at any time. In Pebbles Ai, Company Intel and agent memory are both this kind of context. Pebbles Ai learns your company as you work, but the model itself does not change for your company.
What is the difference between company context and AI memory?
Company context is a full, checked description of your business that admins look after. Memory is a set of short facts an assistant saves when you state or confirm them in chat. In Pebbles Ai, memory has three levels: Your memory, Project memory and Company memory. Company memory reaches your colleagues' assistants. Documents, CRM records and drafts are not copied into memory.
How can a team reuse approved AI-generated content?
Check the content first, then mark it approved so the team knows it can be reused. In Pebbles Ai, the Library gives supported content a review state: Pending review or Approved. The Approved Library is on Team plans and above. Approval does not mean every workflow picks the content up on its own, so check the workflow you use.
In Pebbles Ai
- FoundationCompany IntelBusiness knowledge and brand voice that every assistant works from, built from your website, company information and files.
- FoundationAgent memoryShort facts you share about yourself, a project and your company, carried into later chats and corrected in plain words.
- CapabilityOutreach contentBuild email and LinkedIn campaigns with a persona for each piece, ask an assistant for a message, reply or follow-up in chat, save emails as drafts in your connected mailbox and review the work as a team.
About this guide. Written by the Pebbles Ai team and checked against the current product before publication. Product details describe the web app on the update date shown above; each linked platform page carries the latest detail. Guides are revised when the product changes, and the date changes only with a verified revision.