---
title: "Give AI agents your company context: ICP and messaging"
description: "Give sales and marketing AI your ICP, messaging, voice, and proof. A practical context template, tool settings, approval rules, and measurement plan."
image: "https://bombos.ai/assets/og-en-v2.png"
---

# How do you give AI agents company context that matches your ICP and messaging?

Give sales and marketing AI your ICP, messaging, voice, and proof. A practical context template, tool settings, approval rules, and measurement plan.

*Piet Baudoin, Bombos · published October 30, 2026 · October 30, 2026 · updated October 30, 2026 · October 30, 2026 · 24 min read*

**Start small? One person can always try the plugin.** Together we map what your AI forgets. Free, and the plan is yours, with or without Bombos.

[Book your half hour](https://bombos.ai/en/contact/)

**Give AI agents company context by maintaining one approved company brief with your ICP, disqualifiers, positioning, offer, voice, and evidence. Put decision rules in the tool’s instructions, add the brief and approved examples to a shared project or knowledge source, and retrieve current prospect facts from your CRM for each task. Require source-backed claims, an explicit unknown when evidence is missing, and human approval before sending. Context guides the work; it does not establish buyer intent or prove your promises.**

![A small yellow-and-black bee rests among broad ivory and oat paint strokes, with a fine old-gold line across a quiet black-green background.](https://bombos.ai/assets/kennisbank/how-to-give-ai-agents-your-company-context/shared-context.webp)

By Bombos. Documentation checked October 30, 2026.

For a founder, head of growth, or GTM lead at a seed-to-Series-A B2B company, start with one job: qualify an account and prepare a relevant first email for review. The first task should remove repeated briefing and rewriting, then provide a foundation for the next task.

The template, rules, and scorecard below are our recommended operating design, not measured customer results. Product capabilities are attributed to the makers’ documentation. Salesforce describes context as instructions, tools, knowledge, memory, and boundaries together. That distinction is the starting point. [Salesforce context guide](https://developer.salesforce.com/blogs/2025/08/a-developers-guide-to-context-engineering-with-agentforce).

## What company context does an AI agent need for sales and marketing?

Company context for sales and marketing should tell an agent which accounts to pursue, what problem to address, what it may promise, and what evidence supports the message. Use the following schema as your approved company brief.

| Context | What to record | What the agent must decide |
| --- | --- | --- |
| Company and product | What you sell, the supported use case, product limits, and current terminology | Is this task within the product’s actual scope? |
| ICP | Required account characteristics, operating problem, supported systems, and geography | Does the account meet the must-have conditions? |
| Anti-ICP | Hard exclusions, unsupported requirements, and reasons to decline | Should this account be rejected despite other positive signals? |
| Buyer and trigger | User, champion, decision-maker, their different concerns, and observable buying events | Who should receive which message, and why now? |
| Positioning | Buyer problem, alternative, specific difference, and mechanism | What is relevant to this buyer instead of a generic benefit? |
| Offer and next step | Included work, exclusions, prerequisites, approved commitments, and CTA by buying stage | What can we offer and ask for now? |
| Voice | Approved examples by channel, preferred terminology, banned wording, and editing notes | How should this email, page, or sales brief sound? |
| Proof and claims | Exact allowed claim, evidence, scope, limitations, permission to publish, and review date | Can this statement be used for this audience today? |
| Workflow and access | Source systems, responsible people, allowed actions, and escalation rules | What may the agent read, prepare, change, or send? |

Keep the approved brief in a company-owned document or repository. Give each section a stable heading and each claim a stable ID. Keep supporting documents beside it, with links from the brief. Plain text, Markdown, and readable documents are practical choices; the structure matters more than the file extension.

Do not make every agent maintain a separate ICP. Give outbound, content, and sales-prep agents the same approved company context, then add task-specific instructions. This applies Salesforce’s distinction between shared knowledge and the instructions and actions for a particular job. [Salesforce context guide](https://developer.salesforce.com/blogs/2025/08/a-developers-guide-to-context-engineering-with-agentforce).

## How do you define an ICP and anti-ICP that an agent can actually use?

An agent-ready ICP is a set of evidence requirements and decision rules, including a valid unknown result. Company size and industry can establish account fit; they should not be treated as proof that someone wants to buy.

Separate four kinds of information:

-   **Account fit:** company type, operating model, location, and required systems.
-   **Buyer role:** who does the work, who owns the outcome, and who approves a purchase.
-   **Trigger:** a dated event that makes the problem timely. Funding or hiring can justify research without establishing purchase intent.
-   **Direct evidence:** a prospect’s stated problem, requested evaluation, or confirmed requirement, recorded with its source.

HubSpot’s prospecting documentation separately configures target companies, personas, and signals. Use that separation when designing your own qualification rules. [HubSpot prospecting agent](https://knowledge.hubspot.com/prospecting/use-the-prospecting-agent).

Here is an illustrative policy, not a description of a customer:

```
Required: a B2B sales motion and a recurring account-research task.
Required: the team uses a system the product currently supports.
Exclude: the requested use case requires an unsupported capability.
Research trigger: a public expansion announcement with a source and date.
Do not infer: confirmed research pain, budget, or intent from that event.

If an exclusion is confirmed: reject and explain the exclusion.
If a required fact is missing: mark unknown and request that fact.
If all requirements are evidenced: mark fit, with sources.
Keep fit and readiness to buy as separate fields.
```

For every qualification, request the rule applied, supporting source, date observed, unresolved facts, and next action. Avoid letting several weak positive signals cancel one hard exclusion. The purpose is a defensible decision that a salesperson can inspect.

## How do you turn positioning, proof, and voice into agent instructions?

Useful messaging instructions connect the buyer’s verified problem to a product mechanism, a supported claim, and a stage-appropriate CTA. Store each claim separately, distinguishing documented capabilities, promises, and measured results.

Copy this editable claim record and replace its values. The filled example uses documented Copilot behavior, not a customer result or a Bombos capability; the owner and review date are illustrative.

```
claim_id: COPILOT-01
buyer_problem: Replies without a knowledge lookup or tool call
claim_type: Documented capability
allowed_wording: With generative orchestration enabled and Allow ungrounded responses off, Copilot blocks a turn that uses neither a knowledge source nor a tool.
evidence_url: https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-copilot-studio
supported_use_case: Copilot Studio with generative orchestration enabled
measurement_period: Not applicable; no measured outcome claimed
limitations: General knowledge may still contribute; Web Search is a separate setting.
publication_permission: Public documentation; no customer material
owner: Knowledge lead (assign a named person)
review_due: 2026-11-30 (example date)
```

Here is an illustrative wording review against that record. [Microsoft’s documented behavior and limits](https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-copilot-studio).

| Decision | Example wording | Reason |
| --- | --- | --- |
| Approved in this example | With generative orchestration enabled and Allow ungrounded responses off, Copilot blocks a turn that uses neither a knowledge source nor a tool. General knowledge may still contribute to replies. | Preserves the conditions and limitation. Names the behavior in plain language. |
| Rejected | Copilot answers only from your approved company documents. | Promises a source restriction the setting does not enforce. Tools and Web Search can supply information, and general knowledge may still contribute. |

For voice, annotate examples by channel: short sentences, specific settings, no absolute assurance. Replace vague labels such as professional with these observable choices. HubSpot supports writing samples and custom brand guides. [HubSpot brand voice](https://knowledge.hubspot.com/branding/set-up-brand-voice-using-ai).

For US advertising, the FTC requires evidence before an ad runs. Public documentation supports a capability claim, not a claim of customer improvement. [FTC advertising guidance](https://www.ftc.gov/business-guidance/resources/advertising-faqs-guide-small-business).

Use this editable instruction block with your brief:

```
Read the approved company brief before preparing the task.
Apply ICP rules before outbound; separate facts from hypotheses.
Use current approved claims within their scope and limitations.
Keep claim IDs and prospect sources in separate review notes.
Name missing facts or proof; never fill gaps with inventions.
Follow the channel's approved voice examples and CTA rules.
Return: decision, evidence, draft, open questions, proposed next step.
Prepare only. Sending or publishing requires human approval.
```

Read also: [How to keep AI-written messaging consistent across a GTM team](https://bombos.ai/en/knowledge-base/how-to-keep-ai-written-messaging-consistent-across-a-gtm-team/).

## Should company context go in a system prompt, a project, or a knowledge base?

Put short behavior rules in instructions, reusable reference material in a project or knowledge base, and changing account facts in live system lookups. A searchable knowledge base retrieves relevant material when needed; Microsoft calls this retrieval-augmented generation, or RAG. [Microsoft RAG explanation](https://learn.microsoft.com/en-us/azure/foundry/concepts/retrieval-augmented-generation).

| Location | Put this there | Maintenance rule |
| --- | --- | --- |
| System prompt or tool instructions | Task, exclusions, evidence requirements, output format, and approval boundary | Keep short and update deliberately |
| Shared project | Approved brief, small reference set, and channel examples | Replace outdated uploads and control membership |
| Searchable knowledge base | Product documentation, claim evidence, approved case material, and segment guidance | Maintain source permissions, versions, and retrieval quality |
| CRM or other live source | Current account status, consent or suppression status, recent conversations, and open next steps | Retrieve the relevant record for the current task |

These are roles, not mutually exclusive products. Claude Projects can automatically use RAG as project knowledge approaches the context limit on eligible plans. A project can therefore already contain retrieval without your team building a custom system. [Claude Projects](https://support.claude.com/en/articles/9519177-how-can-i-create-and-manage-projects).

Our decision rule: start with a project when the approved material is small and easy to maintain. Prefer managed retrieval when documents grow, change often, or require different access rights. Consider custom RAG only when the existing tool cannot meet a demonstrated retrieval, permission, freshness, or audit requirement. For connection choices, see [which AI platforms integrate with existing software](https://bombos.ai/en/knowledge-base/which-ai-agent-platform-integrates-with-existing-software/).

Read also: [keeping company context portable across ChatGPT, Claude and Copilot with one owned record](https://bombos.ai/en/knowledge-base/how-to-keep-company-context-portable-across-ai-tools/).

## How do you add company context to ChatGPT, Claude, Copilot, HubSpot, or Salesforce?

You can add company context through these tools’ existing project, knowledge, and selling-context features before commissioning custom retrieval. Choose the route closest to the task and systems you already use. The steps below reflect documentation checked October 30, 2026; availability depends on the account and administrator settings.

| Route | Documented context mechanism | What your team still owns |
| --- | --- | --- |
| [ChatGPT Projects](https://help.openai.com/en/articles/10169521-projects-in-chatgpt) | Files and project instructions | Approved versions, project membership, and output review |
| [Claude Projects](https://support.claude.com/en/articles/9519177-how-can-i-create-and-manage-projects) | Project knowledge and instructions | Approved versions, memory choices, and output review |
| [Copilot Studio](https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-copilot-studio) | Configured knowledge sources with source-specific authentication | Source selection, access configuration, and fallback behavior |
| [HubSpot](https://knowledge.hubspot.com/prospecting/use-the-prospecting-agent) | Prospecting plays with audience, context, and review settings | Qualification rules, evidence quality, and activation decisions |
| [Salesforce Agentforce](https://help.salesforce.com/s/articleView?id=sf.data_library_setup.htm&language=en_US&type=5) | Data libraries and knowledge retrieval | Administrator setup, source preparation, and action boundaries |
| [Custom RAG](https://learn.microsoft.com/en-us/azure/foundry/concepts/retrieval-augmented-generation) | Your application retrieves evidence for the model | Retrieval engineering, permissions, refresh, monitoring, and maintenance |

The last column is our ownership recommendation, not a claim that every product lacks those controls. A project is a practical drafting route; CRM-connected execution needs additional configuration.

### How do you add company context to ChatGPT?

ChatGPT Projects combines reusable files and project instructions for related work. Select **New project**, upload the approved brief and examples, then open the project’s three-dot menu and **Project settings** to add the behavior rules. Choose **Project-only memory** to contain the project’s context. Shared projects use that setting automatically. [OpenAI Projects](https://help.openai.com/en/articles/10169521-projects-in-chatgpt).

Use **Share** to invite the team with appropriate access. Start the task inside that project. Keep the approved source outside ChatGPT too, and deliberately replace old uploads when the brief changes. A file upload is a copy, not evidence that later edits to the original have reached the project. OpenAI also states that the Google Drive app does not sync content in advance when added within a project. [OpenAI Projects](https://help.openai.com/en/articles/10169521-projects-in-chatgpt).

### How do you add company context to Claude?

Claude Projects provides project knowledge and instructions for work within a project. Open **Projects**, choose **\+ New Project**, and add the brief and examples through the knowledge area’s plus button. Use **Set project instructions**, then **Save instructions**, for the behavior rules. [Claude Projects](https://support.claude.com/en/articles/9519177-how-can-i-create-and-manage-projects).

Team and Enterprise accounts support project sharing, subject to administrator controls. Under the project’s memory settings, turn off **Use account memory** when you want account memory kept separate. Maintain approved decisions in project knowledge rather than assuming a correction in one conversation has become company policy. [Claude Projects](https://support.claude.com/en/articles/9519177-how-can-i-create-and-manage-projects).

### How do you add company context to Microsoft Copilot Studio?

Copilot Studio connects an agent to defined knowledge sources. Open the agent’s **Knowledge** page, select **Add knowledge**, and choose the relevant SharePoint source or document upload. Give each source a distinct name and a description explaining when to use it. [Microsoft setup instructions](https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-add-existing-copilot).

For a source-bound agent with generative orchestration enabled, review **Generative AI** settings and turn off **Allow ungrounded responses**. This blocks a response when the turn used no knowledge source or tool; it does not verify every statement in a response. Separately, turn off **Web Search** in the **Knowledge** section of **Overview** if only approved company sources are allowed. The same control appears as **Use information from the web** in **Generative AI** settings. When enabled, it searches all public websites indexed by Bing, beyond configured website sources. [Microsoft knowledge settings](https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-copilot-studio).

Configure the fallback to request the missing fact or hand off to a person. For authenticated SharePoint and Dataverse sources, Microsoft says the agent “only surfaces content that the specific user can access.” Uploaded files have different access behavior, so do not assume every source follows SharePoint permissions. [Microsoft source authentication](https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-copilot-studio).

### How do you add company context to HubSpot?

HubSpot separates prospecting context from brand voice. Open **Sales > Prospecting Agent > Create play**. Define accounts and personas in **Audience**, company and product information in **Context**, and the message approach in **Outreach**. In **Automation**, under **Email review**, choose **Review emails**. Use **Test** to preview a contact before activation. [HubSpot prospecting setup](https://knowledge.hubspot.com/prospecting/use-the-prospecting-agent).

For marketing voice, open **Marketing > Brand > Brand Voice**. Add your guide through **Guidelines > Upload custom guide**, and approved examples through **Writing Samples > Add/manage samples**. These are separate settings: a writing sample should not be your only record of qualification rules or approved evidence. [HubSpot brand setup](https://knowledge.hubspot.com/branding/set-up-brand-voice-using-ai).

### How do you add company context to Salesforce Agentforce?

Salesforce Agentforce uses data libraries to retrieve supporting knowledge. In setup, open **Agentforce Data Library > New Library**, name the library, and save it. Add the approved knowledge articles, files, or other supported source. Salesforce requires Data 360 setup and administrator permissions for this route. [Create a library](https://help.salesforce.com/s/articleView?id=ai.data_library_create.htm&language=en_US&type=5), [supported sources and prerequisites](https://help.salesforce.com/s/articleView?id=sf.data_library_setup.htm&language=en_US&type=5).

Connect the library with **Answer Questions with Knowledge**. Put task behavior and permitted actions in the agent configuration, then check it in a sandbox. A connected library supplies reference material; it does not independently establish that your sales process, record access, and approval rules are correct. [Connect the library](https://help.salesforce.com/s/articleView?id=ai.data_library_assign.htm&language=en_US&type=5), [sandbox guidance](https://help.salesforce.com/s/articleView?id=sf.data_library_setup.htm&language=en_US&type=5).

Bombos is another route when the same context has to reach several of these tools at once: one shared memory under ChatGPT, Claude, Copilot and your own agents, instead of a separate project or library per tool. Ask us to show it on your actual task, including the sources each answer rests on.

## How do you connect company context to CRM research and follow-up?

CRM context needs a defined path into the task and a controlled path back into the account record. Apten describes both directions in its B2C implementation experience: CRM events provide input, and relevant conversation outcomes return to the CRM. Its advice about splitting complex logic across stages is experience-based, not a controlled comparison. [Apten implementation notes](https://www.apten.ai/blog/ai-sales-marketing-agents-for-b2c-mid-market-and-enterprise-companies).

For a B2B workflow, configure separate stages: research the account, apply the ICP rules, prepare a draft, obtain approval, and record the outcome. For incoming replies, distinguish qualification, nurture, and a handoff to the responsible salesperson or an agreed callback. Make the handoff actionable: include the account, conversation, evidence, open question, and owner.

Specify which systems each stage needs. CRM records may be enough for account preparation; a support history, email thread, calendar, or internal knowledge base should be connected only when needed for that task. Define which source owns each field. Usermaven similarly presents shared, reliable GTM data as a foundation for its agents. [Usermaven’s product account](https://usermaven.com/blog/top-ai-agents-for-go-to-market-strategies).

Before enabling CRM writes, define the record match, permitted fields, duplicate handling, and what happens if a write fails. Preserve the unresolved task rather than marking it complete. Keep research hypotheses separate from confirmed customer facts. See [agents versus traditional workflow automation](https://bombos.ai/en/knowledge-base/ai-agents-or-traditional-workflow-automation-when-to-choose-each/) for deciding which steps need judgment and which need fixed rules.

Read also: [How do you automate inbound lead qualification and routing at a post-seed B2B startup?](https://bombos.ai/en/knowledge-base/automate-inbound-lead-qualification-and-routing-at-a-post-seed-startup/)

Read also: [How AI drafts shared-inbox replies from verified company and customer context](https://bombos.ai/en/knowledge-base/ai-draft-replies-in-a-shared-customer-inbox/).

## How do you keep company context current and resolve contradictions?

Current company context needs an owner, an approved version, and a rule for choosing between conflicting sources. Salesforce explicitly identifies conflicting instructions and outdated information as context failure modes. [Salesforce context guide](https://developer.salesforce.com/blogs/2025/08/a-developers-guide-to-context-engineering-with-agentforce).

Use this maintenance policy:

-   Give each source an ID, owner, version, approval status, effective date, review date, and access group.
-   Label evidence as verified, inferred, or unknown. Describe why it has that status instead of relying on a model’s self-reported confidence percentage.
-   Keep draft strategy outside approved retrieval. An idea from a call is not an approved positioning change.
-   Resolve conflicts by field: the approved claims register governs external claims; the designated CRM field governs current account status. Recency alone does not make a source authoritative.
-   When two authoritative sources disagree, stop the affected decision and ask the named owner. Log the resolution and reason.
-   Remove expired or superseded material from active retrieval and replace stale project uploads. Keep an archive outside the active source set.
-   After a change, check that every connected tool retrieves the new version and no longer serves the old claim.

Record corrections with their scope. A rewrite for one prospect should not silently change the company-wide ICP. Have the owner approve reusable changes, then distribute the new brief and repeat the relevant evaluation cases. This is our operating policy; it is not something every project tool automatically enforces.

## What should an AI agent do when prospect data or proof is missing?

An AI agent with missing prospect data or proof should return a specific gap and a next action rather than inventing a fact. Separate three situations:

| Situation | Required behavior |
| --- | --- |
| Required account fact is missing | Mark qualification unknown; name the fact and where an authorized person can verify it |
| No approved proof supports the proposed claim | Remove the claim or hold the draft; ask the claim owner for suitable evidence |
| Sources conflict or require restricted access | Pause the affected decision and escalate with source references |

A failed search is not proof that something is false. Ask the agent to show what it checked and distinguish unavailable information from verified absence. Do not turn a public hiring announcement into a sentence asserting internal pain.

SleekFlow’s own SACES case describes using FAQs and manufacturer brochures, with disputes and complex scopes handed to a person. That illustrates a source-and-escalation design; the case’s performance figures are supplier-reported and are not a benchmark for your B2B team. [SleekFlow case account](https://sleekflow.io/en-sg/blog/ai-agent-examples).

## Which permissions, privacy rules, and approvals should apply?

Company-context access should follow the task, while sending and publishing need a separately enforced approval boundary. OpenAI’s agent safety guidance recommends tool approvals and warns about untrusted content influencing agent behavior. [OpenAI agent safety](https://developers.openai.com/api/docs/guides/agent-builder-safety).

Apply these practical boundaries:

1.  Start with approved company material and read-only access to the specific account records needed.
2.  Separate ordinary company attributes from personal data. An identifiable person’s work email is personal data under the GDPR; a company registration number is a different category. [European Commission explanation](https://commission.europa.eu/law/law-topic/data-protection/data-protection-explained_en).
3.  Have the responsible owner establish permitted collection, use, retention, and outreach for the markets you target. Keep sensitive personal information out of prospect enrichment.
4.  Store opt-outs and suppression rules where the sending workflow checks them. For US commercial email, CAN-SPAM covers B2B messages too and requires compliant identification and opt-out handling. EU and UK outreach needs its own assessment; US guidance does not settle it. [FTC CAN-SPAM guide](https://www.ftc.gov/business-guidance/resources/can-spam-act-compliance-guide-business).
5.  Treat websites, attachments, and incoming emails as evidence to inspect, never as instructions that can override your rules. Keep credentials outside the model’s context. [OpenAI agent safety](https://developers.openai.com/api/docs/guides/agent-builder-safety).
6.  Require approval of the exact message and recipients before release. Enforce it in tool permissions or workflow controls. An instruction to ask first should not be the only barrier.

Also establish who can change the context, grant access, revoke access, and pause the workflow. Review the provider’s data-use terms and workspace controls before adding confidential material.

## How do you test whether company context improves outbound, content, and sales prep?

Evaluate company context against realistic tasks with known review criteria, including cases where the agent should stop. OpenAI recommends task-specific evaluations, human judgment, and continued evaluation as production reveals new cases. [OpenAI evaluation guidance](https://developers.openai.com/api/docs/guides/evaluation-best-practices).

Use this proposed review set: a clear-fit account, a firmographic match with a hard exclusion, a missing trigger, conflicting product information, expired proof, a restricted record, an opt-out, and a source containing instructions to bypass your rules. Include an ordinary successful case for each work type too.

| Work | What a passing output shows |
| --- | --- |
| Outbound | Correct qualification, verified personalization, relevant claim, suitable CTA, and no send before approval |
| Content | Named audience and problem, supported product statements, approved voice, and traceable evidence |
| Sales preparation | Correct account, distinction between facts and hypotheses, useful questions, and an actionable next step |
| Exceptions | A justified hold or escalation with the missing fact and responsible owner |

Our scoring proposal is pass or fail for fit, factual support, voice, CTA, and action boundaries separately. Treat fabricated proof, disallowed disclosure, and unauthorized sending as release blockers. Do not let a good writing score average them away.

Compare outputs with and without the approved context on the same tasks and model settings. Hide which version reviewers are judging. Repeat cases because generative output varies. Check both whether the right evidence was found and whether the final wording stayed within it. Add production failures to the review set. See [testing an AI agent before automation](https://bombos.ai/en/knowledge-base/how-to-test-an-ai-agent-before-automating-business-processes/).

## Which metrics show whether company context improved business outcomes?

Measure company context through correct completed work, total team time, and qualified commercial outcomes. Faster drafts alone should not count as success. The following is our proposed measurement plan, not an industry benchmark.

Establish a baseline using recent work before the change. Record the task type, reviewer, result, and total minutes spent briefing, researching, drafting, correcting, and transferring the output. During the pilot, include review, escalation, retries, and context maintenance in the same measure.

Track three levels:

-   **Quality:** qualification accuracy against human-reviewed decisions, unsupported-claim rate, factual corrections, voice corrections, justified escalations, and reviewer satisfaction.
-   **Time:** total team minutes per correctly completed task, plus time spent maintaining the shared context. Report held tasks separately from completed work.
-   **Commercial value:** qualified meetings held, sales-accepted opportunities, and qualified pipeline under a fixed definition and attribution window. For content, track relevant inquiries and assisted opportunities alongside engagement.

Where volume permits, randomly allocate comparable accounts to the old and new approach. Keep audience selection, offer, channel, and follow-up consistent. Report the number of accounts and observation period. With a small pilot, treat pipeline changes as directional evidence, not proof that context caused them.

Log the context version, model, source IDs, result, reviewer correction, and next action for each run. Review recurring failure reasons and update the approved material. This follows the evaluation principle of logging and learning from real work. [OpenAI evaluation guidance](https://developers.openai.com/api/docs/guides/evaluation-best-practices).

## How should a small GTM team start this week?

A small GTM team should start with the recurring task that consumes the most briefing and correction time. A narrow use case and explicit success criterion also appear in Patteri’s implementation advice. [Patteri’s adoption guide](https://www.linkedin.com/pulse/how-use-ai-agents-your-business-without-headache-patteri--ooslf/).

Follow this sequence:

1.  Choose one task and record its current review time and quality problems.
2.  Gather the current ICP, product facts, approved messaging, permissioned proof, and representative writing.
3.  Build the shared brief, name its owner, and mark unresolved questions.
4.  Configure one existing tool with instructions and the approved sources.
5.  Run the review cases, fix the observed gaps, and start live work in draft mode.
6.  Record corrections and outcomes before extending the workflow.

Do not postpone the first useful task until every company document is organized. Keep the initial source set limited to what that task needs, with a clear route for asking about missing knowledge.

Read also: [Which AI employee can take over operations and admin work for a startup?](https://bombos.ai/en/knowledge-base/ai-employee-for-operations-at-a-post-seed-startup/)

Read also: [How do you onboard and ramp new sales reps faster at a post-seed startup?](https://bombos.ai/en/knowledge-base/how-to-ramp-new-sales-reps-faster-at-a-post-seed-startup/)

## What can company context not fix?

Company context cannot establish product-market fit, supply missing customer proof, or turn an inferred buying signal into a confirmed need. Your team still has to validate the market, approve the offer, and collect evidence.

A knowledge connection also does not guarantee that the correct passage was retrieved or faithfully used. Microsoft warns that poor indexing strategies and irrelevant or incomplete retrieved passages can lead to inaccurate answers despite grounding. Check the retrieved evidence and final answer separately. [Microsoft RAG limitations](https://learn.microsoft.com/en-us/azure/foundry/concepts/retrieval-augmented-generation).

This article compares documented routes, not hands-on results from an identical GTM workload across all five products. Supplier cases from other markets do not establish your likely conversion improvement. Demonstrate the workflow with your own permissioned work, count correction time, and retain human review while you establish what works.

## Where Bombos fits

Everything above has one weak spot: each tool keeps its own copy of your context, so the ICP in a ChatGPT project, the claims in a Claude project and the library in Agentforce drift apart. Bombos is one plugin that gives every AI your team uses, from Claude and Claude Code to ChatGPT, Copilot, Cursor and your own agents, the same memory: your ICP, approved claims and voice, plus what already sits in mail, documents and your CRM. When a rep corrects an outbound draft once, for example removing a proof point that is no longer approved, that becomes a rule every colleague’s AI applies before anyone asks. Each answer shows the source it rests on, so a contradiction can be traced back to its owner. One person can try the plugin on their own agents first; it runs in our cloud or standalone on your own server. [Book your half hour](https://bombos.ai/en/contact/) and we map together what your AI forgets.

Read also [how to automatically log emails and attachments to the right CRM account](https://bombos.ai/en/knowledge-base/how-to-automatically-log-emails-and-attachments-to-the-right-crm-account/): the never-log list and the exceptions queue are company context in its most concrete form.

Read also: [How do you automate security questionnaires and RFP responses at a post-seed startup?](https://bombos.ai/en/knowledge-base/how-to-automate-security-questionnaires-and-rfp-responses/)

Read also: [What is context engineering?](https://bombos.ai/en/knowledge-base/what-is-context-engineering/) and [Mem0 alternatives for team memory](https://bombos.ai/en/knowledge-base/mem0-alternatives-for-team-memory/)

## Sources

All sources below were opened and read on October 30, 2026. Dates are visible publication or update dates where available; otherwise the access date is given. Product documentation describes maker-stated behavior, not an independent performance comparison. Apten, Usermaven, and SleekFlow describe their own products or implementation experience.

1.  [Salesforce: A Developer’s Guide to Context Engineering with Agentforce](https://developer.salesforce.com/blogs/2025/08/a-developers-guide-to-context-engineering-with-agentforce). August 2025.
2.  [Microsoft: Retrieval augmented generation and indexes](https://learn.microsoft.com/en-us/azure/foundry/concepts/retrieval-augmented-generation). Updated August 21, 2026.
3.  [OpenAI: Projects in ChatGPT](https://help.openai.com/en/articles/10169521-projects-in-chatgpt). Updated October 30, 2026.
4.  [Anthropic: How can I create and manage projects?](https://support.claude.com/en/articles/9519177-how-can-i-create-and-manage-projects). Living documentation, accessed October 30, 2026.
5.  [Microsoft: Add knowledge to an agent](https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-add-existing-copilot). Updated October 22, 2026.
6.  [Microsoft: Knowledge sources summary](https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-copilot-studio). Updated August 4, 2026.
7.  [HubSpot: Set up and use the prospecting agent](https://knowledge.hubspot.com/prospecting/use-the-prospecting-agent). Updated October 14, 2026.
8.  [HubSpot: Set up brand voice using AI](https://knowledge.hubspot.com/branding/set-up-brand-voice-using-ai). Updated October 13, 2026.
9.  [Salesforce: Create a Data Library](https://help.salesforce.com/s/articleView?id=ai.data_library_create.htm&language=en_US&type=5). Living documentation, accessed October 30, 2026.
10.  [Salesforce: Setting Up Data Libraries](https://help.salesforce.com/s/articleView?id=sf.data_library_setup.htm&language=en_US&type=5). Living documentation, accessed October 30, 2026.
11.  [Salesforce: Connecting Data Libraries to Agents](https://help.salesforce.com/s/articleView?id=ai.data_library_assign.htm&language=en_US&type=5). Living documentation, accessed October 30, 2026.
12.  [OpenAI: Evaluation best practices](https://developers.openai.com/api/docs/guides/evaluation-best-practices). Living documentation, accessed October 30, 2026.
13.  [OpenAI: Safety in building agents](https://developers.openai.com/api/docs/guides/agent-builder-safety). Living documentation, accessed October 30, 2026.
14.  [FTC: Advertising FAQs, a guide for small business](https://www.ftc.gov/business-guidance/resources/advertising-faqs-guide-small-business). Undated guidance, accessed October 30, 2026.
15.  [FTC: CAN-SPAM Act, a compliance guide for business](https://www.ftc.gov/business-guidance/resources/can-spam-act-compliance-guide-business). Undated guidance, accessed October 30, 2026. US scope.
16.  [European Commission: Data protection explained](https://commission.europa.eu/law/law-topic/data-protection/data-protection-explained_en). Undated guidance, accessed October 30, 2026. EU scope.
17.  [Apten: AI Sales/Marketing Agents for B2C](https://www.apten.ai/blog/ai-sales-marketing-agents-for-b2c-mid-market-and-enterprise-companies). June 1, 2025.
18.  [Usermaven: Top AI agents for GTM strategies](https://usermaven.com/blog/top-ai-agents-for-go-to-market-strategies). Updated October 29, 2026.
19.  [SleekFlow: AI agent examples in business](https://sleekflow.io/en-sg/blog/ai-agent-examples). Updated August 24, 2026.
20.  [CA Dharmajan Patteri: How to Use AI Agents in Your Business](https://www.linkedin.com/pulse/how-use-ai-agents-your-business-without-headache-patteri--ooslf/). June 24, 2025.

## What does your AI forget?

Start small? One person can always try the plugin.

1.  01You tell us which AI you use and where it keeps starting over.
2.  02We show what a shared memory changes there, on your own examples.
3.  03You keep the plan, with or without Bombos.

![Piet Baudoin](https://bombos.ai/assets/team/piet-klein.jpg)

**Piet Baudoin**You talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe

## More about starting, testing, and adjusting

-   [How do you request an AI automation platform for your business?](https://bombos.ai/en/knowledge-base/how-to-request-an-ai-automation-platform-for-your-business/)
-   [Can employees adjust AI workflows without a developer?](https://bombos.ai/en/knowledge-base/can-employees-adjust-ai-workflows-without-a-developer/)
-   [Can I start with an AI automation pilot?](https://bombos.ai/en/knowledge-base/can-i-start-with-an-ai-automation-pilot/)
-   [How do you implement an AI agent for administrative processes?](https://bombos.ai/en/knowledge-base/how-to-implement-an-ai-agent-for-administrative-processes/)
-   [How do you start a pilot for administrative automation with AI?](https://bombos.ai/en/knowledge-base/how-to-start-an-ai-administrative-automation-pilot/)

[All questions in the knowledge base →](https://bombos.ai/en/knowledge-base/)


---

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://bombos.ai/#organization","name":"Bombos","url":"https://bombos.ai/","logo":"https://bombos.ai/assets/bombos-bee.png","description":"Bombos is the memory layer under every AI your team uses. One shared memory for Claude, ChatGPT, Copilot, Cursor and your own agents, that learns how you work.","sameAs":["https://www.linkedin.com/company/bombos-ai","https://www.youtube.com/@truebombos","https://www.instagram.com/truebombos/","https://bsky.app/profile/bombostrue.bsky.social","https://www.threads.com/@truebombos","https://mastodon.social/@Bombos","https://www.tiktok.com/@truebombos","https://t.me/truebombos","https://www.pinterest.com/truebombos/","https://github.com/bombos-ai","https://linktr.ee/bombos.ai","https://bombosai.substack.com","https://hashnode.com/@bombosai","https://hackernoon.com/u/bombos","https://dev.to/bombos"],"email":"info@bombos.ai","areaServed":["NL","Europe"],"knowsLanguage":["nl","en"],"contactPoint":{"@type":"ContactPoint","email":"info@bombos.ai","contactType":"sales","url":"https://bombos.ai/en/contact/","availableLanguage":["nl","en"]}},{"@type":"WebSite","@id":"https://bombos.ai/#website","name":"Bombos","url":"https://bombos.ai/","inLanguage":["nl-NL","en"],"publisher":{"@id":"https://bombos.ai/#organization"}},{"@type":"Person","@id":"https://bombos.ai/#piet","name":"Piet Baudoin","url":"https://bombos.ai/over-ons/","worksFor":{"@id":"https://bombos.ai/#organization"},"sameAs":["https://www.linkedin.com/in/piet-baudoin","https://x.com/marketeerpiet","https://dev.to/marketeerpiet","https://medium.com/@piet_7573"]},{"@type":"Article","@id":"https://bombos.ai/en/knowledge-base/how-to-give-ai-agents-your-company-context/#article","headline":"How do you give AI agents company context that matches your ICP and messaging?","description":"Give sales and marketing AI your ICP, messaging, voice, and proof. A practical context template, tool settings, approval rules, and measurement plan.","datePublished":"2026-10-30T00:00:00.000Z","dateModified":"2026-10-30T00:00:00.000Z","author":{"@id":"https://bombos.ai/#piet"},"publisher":{"@id":"https://bombos.ai/#organization"},"image":"https://bombos.ai/assets/og-en-v2.png","mainEntityOfPage":"https://bombos.ai/en/knowledge-base/how-to-give-ai-agents-your-company-context/","isPartOf":{"@id":"https://bombos.ai/#website"},"inLanguage":"en"},{"@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Bombos","item":"https://bombos.ai/en/"},{"@type":"ListItem","position":2,"name":"Knowledge base","item":"https://bombos.ai/en/knowledge-base/"},{"@type":"ListItem","position":3,"name":"Starting, testing, and adjusting","item":"https://bombos.ai/en/knowledge-base/#starting"},{"@type":"ListItem","position":4,"name":"How do you give AI agents company context that matches your ICP and messaging?","item":"https://bombos.ai/en/knowledge-base/how-to-give-ai-agents-your-company-context/"}]},{"@type":"FAQPage","@id":"https://bombos.ai/en/knowledge-base/how-to-give-ai-agents-your-company-context/#faq","mainEntity":[{"@type":"Question","name":"What company context does an AI agent need for sales and marketing?","acceptedAnswer":{"@type":"Answer","text":"Company context for sales and marketing should tell an agent which accounts to pursue, what problem to address, what it may promise, and what evidence supports the message. Use the following schema as your approved company brief."}},{"@type":"Question","name":"How do you define an ICP and anti-ICP that an agent can actually use?","acceptedAnswer":{"@type":"Answer","text":"An agent-ready ICP is a set of evidence requirements and decision rules, including a valid unknown result. Company size and industry can establish account fit; they should not be treated as proof that someone wants to buy."}},{"@type":"Question","name":"How do you turn positioning, proof, and voice into agent instructions?","acceptedAnswer":{"@type":"Answer","text":"Useful messaging instructions connect the buyer's verified problem to a product mechanism, a supported claim, and a stage-appropriate CTA. Store each claim separately, distinguishing documented capabilities, promises, and measured results."}},{"@type":"Question","name":"Should company context go in a system prompt, a project, or a knowledge base?","acceptedAnswer":{"@type":"Answer","text":"Put short behavior rules in instructions, reusable reference material in a project or knowledge base, and changing account facts in live system lookups. A searchable knowledge base retrieves relevant material when needed; Microsoft calls this retrieval-augmented generation, or RAG. Microsoft RAG explanation."}}]}]}
```
