How do we give our AI agents shared company knowledge?

Give AI agents shared company knowledge through governed access to approved sources, with permissions checked before information reaches the model. Keep company facts separate from agent instructions and task memory. Retrieve relevant documents, query live systems for current records, and let named people approve corrections and consequential changes.

Give your AI agents access to the same approved company sources through a permission-controlled retrieval interface. Keep product rules in their owned documents, current account facts in live systems, and unfinished work in separate task memory. Update an authoritative source once, then verify that every relevant agent receives the correction. Start with one recurring question and a named source owner. The aim is one coherent company across agents, with evidence for each answer and clear limits on who may receive it.

Three blue, olive and pink robot helpers consult one open binder while a human hand replaces a blank page

Which architecture should we use for shared agent knowledge?

Shared agent knowledge needs a common way to find authorized evidence, rather than a separate handbook uploaded to every agent. Choose the architecture by the information’s structure, rate of change and access requirements. A shared interface can query several existing systems; you do not need to move all company information into one repository.

NIST defines retrieval-augmented generation, or RAG, as pairing a generative model with a separate information retrieval system. Relevant information enters the model’s context when answering. That allows knowledge updates without retraining. Uploading documents into such a system therefore differs from teaching new facts to the model’s parameters.

Keep four responsibilities explicit. Company knowledge states approved facts and decisions. Agent instructions define the job and permitted behavior. Task state records unfinished work. Durable memory preserves selected information across sessions. LangChain’s memory documentation distinguishes thread-scoped state from longer-term memory shared across conversational threads. Neither persistence mechanism establishes that a remembered statement is company policy.

Use this recommended selection table before choosing a platform:

Information and taskStarting architectureMain limitation to check
A small, stable set of approved answersMaintained documents with ordinary search or explicit selectionManual updates and access handling must remain workable
Knowledge already managed within a workspaceNative retrieval or a permission-aware connectorAvailable sources, account requirements, synchronization and audit visibility
Unstructured documents used by several applicationsA shared retrieval index with source referencesYou own ingestion, permissions, deletion and retrieval quality
Current account status, entitlements or appointmentsA narrowly scoped live API or database queryAuthentication, field selection, outages and observation time
A handoff that must survive separate sessionsA scoped state or memory storeExpiry, source restrictions and unapproved interpretations

For a native option, Microsoft documents delegated access for its Copilot Retrieval API: it enforces the signed-in user’s permissions on each request. This is a specific product property. Verify your intended API, source coverage and account entitlement before assuming it fits.

For an indexed option, Amazon Bedrock describes its document retrieval pipeline: extract text, split it into chunks, create numerical embeddings and retrieve similar passages while retaining a mapping to the original document. A vector index supports this search. It does not decide which document is authoritative.

Example throughout this page: a post-seed software startup wants a sales agent and a support agent to answer integration questions. Keep approved product capabilities in the product documentation. Query the CRM for an account’s current onboarding stage. Store an unfinished follow-up in task state. Both agents should use the same approved capability record without receiving identical tools or account access.

MCP can provide a connection to retrieval or live tools. The July 2026 MCP authorization specification defines transport-level authorization for HTTP and makes authorization optional for implementations. A connection using MCP therefore needs its own verified access policy. MCP is not a knowledge repository or an approval decision.

Choose a native workspace when your sources already live there and its retrieval respects their permissions. Use an open protocol such as MCP when several clients must reach the same tools. In both cases, keep source records exportable with stable IDs and versions, so the retrieval layer can be replaced without rewriting the knowledge.

Which company fact do your agents currently maintain in separate copies, and who would notice if those copies diverged?

What should go into our shared company knowledge?

Shared company knowledge should contain the information needed to make a defined decision: approved facts, their limits, operating procedures and the reasons behind important choices. Include what the company intends to achieve and what it refuses to promise. Keep raw observations and provisional interpretations visibly separate from approved statements.

Begin with recurring questions your team can already answer correctly. Gather the relevant product pages, process instructions and decision records. Remove superseded duplicates from the active answer set. Leave original systems in charge of the records they own. A product capability should not acquire a second authority because marketing copied it into a convenient folder.

Use the following recommended record template for each retrievable knowledge item. Fields can sit in document metadata or a separate catalog, provided they remain attached to extracted passages.

FieldWhat the owner records
Canonical source ID and URLThe enduring identity and original location
Subject and scopeProduct, feature, audience, language and applicable conditions
Approved statementThe fact or rule the company stands behind
Rationale and boundariesWhy the rule exists, exclusions and prohibited claims
Owner and approverNamed people responsible for content and publication
Status and versionProposed, approved, superseded or withdrawn, with a version identifier
DatesEffective date, last substantive review and next review trigger
Access and sensitivityPermitted organization, users or groups, plus intended disclosure audience
Retention and dependenciesWhen to remove it and which summaries or processes depend on it

This template is an implementation recommendation. NIST’s Generative AI Profile provides the broader basis: document oversight responsibilities, provenance and changes to content. A timestamp alone does not establish approval or continuing validity.

Example: the product owner records that an integration imports selected contact fields but does not synchronize deletions. The rationale explains that deletion handling remains under review. The approved external wording describes the supported import. The excluded claim is “complete two-way synchronization.” Sales and support can use the same boundary without interpreting an internal roadmap as a delivery commitment.

Read also: deciding which customer answers belong in a public knowledge base.

Preserve decision context from meetings selectively. Before recording, agree the purpose, appropriate permission or consent, audience and retention. Tell participants before recording what is kept, for whom and for how long. Ask the owner to confirm the actual decision, rejected alternative and condition for revisiting it. Keep a brainstorming remark labeled as an observation. Do not let an automated summary silently promote someone’s suggestion into company policy.

For instructions about what an agent may do with this knowledge, define goals and guardrails for marketing AI agents. A knowledge record can explain a limitation; the agent’s execution controls must enforce its permitted actions.

If your agent knew every product fact but none of your reasons or boundaries, what would it be likely to promise?

How do we retrieve the right knowledge without overloading the agent?

Company retrieval should return the smallest sufficient body of authorized evidence for the current task, including the qualifications needed to interpret it. Select by topic, applicability and authority before asking the model to synthesize an answer. A large context window does not remove the need to choose relevant material.

Design retrieval units around meaning. For a policy, keep the rule with its exceptions and effective date. For a table, preserve column headings and units alongside the relevant row. For a procedure, retain the prerequisite and the step being answered. For scanned material, check the extracted text against the original before treating it as evidence.

Example: the integration question is “Does deleting a contact in the source remove it from our workspace?” A passage saying “contact synchronization is supported” is insufficient. The relevant unit includes the supported operations and the deletion exception. Retrieve its parent section when a short matching passage omits that condition. Avoid returning the entire product handbook merely because it contains the feature name.

Compare retrieval methods on your actual questions. Exact names and identifiers deserve explicit lookup cases. Paraphrased questions deserve meaning-based cases. Amazon Bedrock’s query documentation describes semantic search, hybrid search combining vectors with raw text, metadata filtering and reranking. Availability depends on the configured store. These are options to compare, rather than a universal recipe for better answers.

Use this recommended retrieval sequence:

  1. Identify the caller, intended audience and task. Ask for clarification if product, account or version is ambiguous.
  2. Select the permitted sources and applicable approved versions. Apply authorization before passing evidence to the model.
  3. Retrieve candidate passages, retaining headings, exceptions, source IDs and versions.
  4. Compare candidate relevance and fetch missing context. Query a live tool separately for current structured facts.
  5. Answer only the supported part. Attach the source to the claim it supports and distinguish any inference.

Specify the no-answer path as carefully as the answer path. For a separate question about deleting an entire source account, a proposed response is: “I found the approved contact-import rule, but no approved answer about account deletion. I have sent that question to the product owner.” Do not convert a low-confidence match into a confident yes.

Treat retrieval scores as ranking signals. Set acceptance rules against owner-labeled examples instead of presenting a similarity score as the probability that a business answer is correct. Compare alternative chunk boundaries and context amounts on the same questions. Retain extra complexity only where it resolves observed failures.

Which exception would disappear if your most important policy were cut into isolated paragraphs?

How do we preserve permissions across agents and recipients?

Shared company knowledge must preserve authorization from source to answer: the requesting identity, application, task and recipients must all fit the disclosure policy. Enforce access outside the language model before restricted evidence enters its context. A prompt that says “keep this confidential” is not a substitute for denying the read.

Bhatt and colleagues’ participant-aware access-control paper demonstrates exfiltration attacks and argues that material used in an interaction should be authorized for all participants. Its proposal is a security architecture, not certification of a particular deployment. The practical implication is especially relevant when an agent drafts something for someone other than the requesting employee.

Example: a founder may read internal integration incident notes. A prospect receiving a sales email may only receive approved public capability information. Generate the external draft from the externally permitted evidence set. An internal investigation can use a separate restricted context. Having the model read everything and then asking it to remove sensitive details leaves the boundary dependent on generation.

For a custom system, require a server-validated company identity and access scope. Do not let the model choose its own organization or permission filter. An autonomous scheduled agent needs an explicit service identity and purpose; it cannot inherit a hypothetical signed-in user’s rights. Review broad service accounts particularly carefully.

The OWASP RAG Security guidance recommends carrying source permissions into chunks and enforcing them at retrieval. It also covers cache isolation, invalidation and deletion. Apply those controls to derived summaries and memories as well. Otherwise, a restricted source may disappear from search while its content remains available elsewhere.

Use this recommended access review:

StageEvidence to request from the builder
Source connectionWhich identity reads which repositories, records and fields
Extraction and indexingHow source restrictions attach to chunks and derived records
RetrievalHow current authorization is checked before context reaches the model
Output and handoffHow the intended recipients restrict the evidence set
Cache and memoryHow permissions, expiry and source dependencies are retained
RevocationHow changed access or deletion blocks reuse and clears derived copies

Agree a maximum acceptable propagation delay separately for ordinary content changes and access revocations. Ask for a demonstration with an already-cached answer. If a revocation has not propagated, require a current authoritative permission check or block the affected retrieval. Document what happens to queued drafts and retained operational records. “We synchronize regularly” is not an operational boundary.

Treat retrieved documents as untrusted data. OWASP also identifies malicious instructions inside retrieved content as a RAG attack surface. Keep tool authorization independent of those instructions. A product document must never grant an agent permission to change sharing settings.

Before selecting a provider, map where document text, embeddings, inference inputs, logs and evaluation samples go. Check contractual terms for use in training, retention, processing regions, deletion and subprocessors. NIST’s profile recommends supplier assessment and explicit agreements about ownership, security and provenance. A permission-correct answer does not resolve the separate question of whether the processing arrangement is appropriate.

How do we keep knowledge current and resolve conflicting sources?

Shared knowledge stays dependable when a named owner controls each topic, approved changes supersede earlier versions, and corrections reach every dependent index or memory. Resolve conflicts by authority and applicability before recency. A newer informal comment should not override an approved policy simply because it has a later timestamp.

Define precedence in the knowledge contract. For the example startup, the product owner could approve the capability specification, while account-specific commitments remain with the authorized account owner. An agent should identify which rule applies to the question. It should not combine conflicting statements into an invented compromise.

Keep both effective date and recording date. A document edited today might describe a future release. An older record might still govern an account’s agreed terms. Put an explicit current-version pointer beside each topic, and retain superseded material outside the default answer set when there is a legitimate reason to keep it.

Use this example correction procedure:

  1. Support reports that the approved import description is incomplete. The report links the disputed statement and observed behavior.
  2. The product owner checks the underlying evidence and approves a corrected capability record, with its effective scope.
  3. The update supersedes the previous approved version. The retrieval process records successful indexing or reports failure.
  4. Dependent summaries and queued drafts are invalidated or reviewed. Both sales and support are checked against the corrected question.
  5. The operator records when the correction became usable and investigates any continuing old answer.

Every entry should have an owner and review date, but schedule reviews according to the consequence and volatility of the topic. A change in product behavior should trigger review immediately through your agreed process. A routine review date is a reminder to inspect the source, rather than evidence that inspection happened.

Freshness also needs observable operation. Record the source version, last successful ingestion, most recent failure and observation time for live queries. Show those details to the reviewer where they affect the decision. Define whether a source outage permits a labeled older answer or requires a handoff. Treat uncertain permissions more strictly than a tolerable delay in low-risk product wording.

NIST’s Generative AI Profile recommends documenting provenance and human oversight roles. Apply that principle by giving people authority to correct the record and challenge the system’s interpretation. Agents can prepare the work, while accountable owners decide what the company stands behind.

Who can correct a wrong company answer today, and how would they verify that every agent has stopped using it?

Can AI agents write back to the shared knowledge base?

AI agents can be given tools to propose or perform knowledge changes, but reading, drafting, approving, publishing and deleting should have separate permissions. Begin with proposed additions in a review area. Broaden write access only for changes whose consequences, validation and recovery your team can explain.

Write-back is useful when it leaves a reviewable record. It becomes risky when an inference is published as an approved fact or another agent treats an unreviewed summary as evidence. Preserve the distinction between an observation, a proposed interpretation and an authorized decision through the whole workflow.

Use this recommended starting policy:

Proposed operationInitial boundary
Record an unanswered questionAppend to a restricted improvement queue with source and task references
Suggest a corrected paragraphCreate a draft against the source version; preserve the original
Summarize a meeting decisionRequest confirmation from the decision owner before approval
Change approved policy or product claimsRequire the named owner’s approval of the exact change
Publish externally, delete records or change accessUse a separate authorized operation with explicit review

Example: after the integration question, the support agent proposes a new FAQ entry. Its proposal contains the question, suggested answer, supporting source version and unresolved uncertainty. The owner approves or edits it. Publication writes a new version, then triggers retrieval checks. Other agents should not retrieve the draft through the approved company knowledge route.

Protect against concurrent edits. Require the write operation to name the version it read. If that version has changed, stop and reconcile instead of overwriting another correction. Give retried requests an operation identifier so the same proposal does not create duplicate entries. Keep an audit record of proposer, approver, previous version and resulting version, with a workable rollback.

OWASP recommends independent tool authorization and explicit confirmation for high-risk RAG-influenced actions. Implement those checks at the write tool. For the review process itself, use human approval in AI marketing workflows to define what the reviewer sees and which exact action the approval permits.

How do we keep different agents consistent across customer channels?

Agent consistency means preserving the same approved facts, limits and commitments across channels while adapting the explanation to the question. A support answer and a sales email can use different wording. They should not disagree about what the product does or what the company has promised.

Give agents a common evidence contract: source identity, approved version, applicable scope, supported statement and unresolved questions. Keep their roles separate. Sales can explain fit and propose a conversation. Support can investigate a permitted account issue. Neither role should expand a capability because the other agent used enthusiastic wording.

Example handoff: the sales agent passes the buyer’s integration question, the approved import-only capability record and a note that the buyer’s required deletion workflow still needs a fit decision. The support agent retrieves the current record under its own permissions and checks any relevant account state. It receives references and unresolved work, rather than an unsupported conclusion that the integration is fully suitable.

If permissions differ, pass only references and context the receiving agent may access. Recheck the evidence when a queued handoff is resumed after a product update. A shared conversation transcript should not become a route around narrower source permissions.

For partner-facing workflows, connect the knowledge contract to turning product integrations into shared growth. Agree which integration claims both parties approve, who updates them and which follow-up each side owns. Consistency has commercial value when a buyer can move between conversations without discovering a different product in each one.

Would a buyer speaking to your sales agent, support agent and founder meet the same company?

How do we know whether our shared knowledge is trustworthy?

Evaluate shared agent knowledge by separating source retrieval, use of evidence, answer correctness, permissions and freshness. A fluent answer can fail any of those checks. Require evidence from representative company questions and failure scenarios before expanding access or allowing consequential actions.

The RAGAs research paper distinguishes finding relevant context, using it faithfully and producing a useful generation. Its automated metrics support evaluation; they do not establish that your company’s answer is current, authorized or commercially appropriate. Add owner judgments and task-specific acceptance conditions.

Create a company-owned question set with expected evidence, applicable audience and acceptable outcomes. Include ordinary wording, paraphrases, exact product identifiers, questions requiring two sources, conflicting versions, unsupported questions and attempts to retrieve restricted content. Include tables and extracted documents where your corpus uses them. Keep some cases separate from the examples used to tune retrieval and instructions.

Use this example evaluation record for the integration scenario:

CheckQuestion for the reviewer
Retrieval coverageDid the returned context include the approved import rule and deletion exception?
FaithfulnessDid every factual statement follow from that context without extra capabilities?
Correctness and applicabilityWas the rule valid for this product version and account situation?
Citation supportDid each source actually support its attached claim, including qualifications?
AbstentionDid an unsupported promise remain unanswered or reach the right owner?
AuthorizationWere restricted notes excluded before model input and external drafting?
FreshnessDid a correction or revocation change behavior within the agreed boundary?
Operational usefulnessCould the recipient act appropriately after reasonable review effort?

Inspect retrieval before rewriting the prompt. If the approved passage was absent, investigate source coverage, parsing, permissions or ranking. If it was present but misread, inspect the instructions and synthesis. If both worked but the answer was wrong, inspect the source’s authority and applicability. These failure classes need different repairs.

Report the denominator for each measure. Retrieval coverage concerns questions with known answer evidence. Correct abstention concerns unsupported questions. Permission checks concern prohibited reads and disclosures. Keep failed, refused and unresolved runs visible. A single overall accuracy figure can hide the failure that makes the application unsuitable.

Repeat the relevant checks after changes to sources, access rules, extraction, retrieval settings or models. Record the configuration with the result. Compare against the previous version on the same cases, then inspect ordinary use for questions your evaluation missed. Measure employee lookup and review effort separately from answer quality. Faster drafts are useful only if the total work and resulting decisions improve.

Which failure would make you stop the system even if its average answer score improved?

Where should a post-seed startup start, and when should it wait?

A post-seed startup should begin with one recurring knowledge task, a bounded corpus and a named operator. Add a second agent to prove that the same approved source can serve different roles. Postpone a wider rollout when source ownership, disclosure rules or ongoing maintenance remain unresolved.

For the example startup, a first scope could cover integration capability questions using approved product documentation. Keep account-specific exceptions and external sending under review. Select the question set, record the current manual lookup effort and agree the pass condition in advance: on the question set, the approved passage is retrieved, each claim is supported by its cited source, unsupported questions are referred to the owner, no restricted read occurs, and lookup effort falls compared with the manual baseline. Meet this condition before connecting more repositories.

Run the proposed route beside the existing process. Observe unanswered questions, source corrections, permission changes and handoffs. Have someone other than the builder identify a stale answer, stop retrieval and restore manual handling. Accept the operational handover along with the useful answers.

Choose ordinary search or a maintained FAQ when those meet the need. Use a fixed workflow when the information and route are predictable. Delay the agent when nobody can judge the answer, the team disagrees about policy, or a forthcoming system replacement makes the connection temporary. More automated retrieval cannot resolve an undecided company position.

If the remaining difficulty is identity, security or reliability across systems, use the decision framework for when a GTM workflow needs AI engineering. Assign that component to someone who can verify the boundary. Preserve portable source records, version history and operating instructions so future changes do not depend on one builder’s memory.

How can Poldermarketing help connect knowledge to go-to-market work?

Piet Baudoin of Poldermarketing helps post-seed startups connect their go-to-market work to usable information: approved product explanations, evidence for lead qualification, CRM records and clear follow-up. A concrete starting point is one recurring question whose answer currently gets reconstructed across several tools.

Piet works remotely in Dutch and English. Alongside positioning and findability in Google and AI answers, he builds and runs AI automations with care: lead follow-up, enrichment and qualification, CRM setup, reporting, content distribution and tool integrations. Agree the source owner, allowed actions and maintenance before expanding the workflow. Read how Piet scopes, builds and runs the work, or request a no-cost growth scan to identify a useful starting point.

Questions people ask

Do we need to train a model on our company documents?

Usually, start with retrieval: provide approved information when the agent needs it. NIST defines retrieval-augmented generation as pairing a model with a separate retrieval system, allowing knowledge updates without retraining. Consider fine-tuning separately for a demonstrated behavioral need. Do not use it as a substitute for source ownership, current records or access controls.

Can agents using different models share the same knowledge?

Design a common retrieval interface that returns permitted passages with source identifiers, versions and dates. Different agent applications can call that interface if their integrations support it. Keep authorization on the server and specify the response format. A shared interface does not make different models equally reliable; evaluate each agent's use of the evidence.

Is a vector database required for shared company knowledge?

No. Begin with the least complex route that meets your task and access requirements. A small approved document set may need ordinary search, an existing workspace may provide retrieval, and current account facts may need a direct API query. A vector index is one implementation choice for finding relevant material, rather than the whole knowledge system.

Does sharing an agent also share its knowledge permissions?

Treat agent sharing and source access as separate configuration decisions. Check whose identity performs each read, which permissions the application uses and who receives the answer. Do not assume that a colleague who can open an agent may receive every source available to its creator. Repeat the check for scheduled runs without an active user.

Should agents remember everything from company conversations?

Set a purpose and retention rule before capturing conversations. Keep personal preferences, temporary task state and proposed summaries separate from approved company decisions. Ask the responsible person to confirm a decision record before publishing it as policy. Preserve its source and access restrictions, and provide a way to correct or remove derived memories.

Related questions

Sources

  1. NIST CSRC (2025): Retrieval-augmented generation, glossary definition
  2. Microsoft Learn (accessed 10 November 2026): Security and authentication for Microsoft 365 Copilot APIs
  3. Amazon Web Services (accessed 10 November 2026): How Amazon Bedrock knowledge bases work
  4. LangChain documentation (accessed 10 November 2026): Memory overview and cross-session namespaces
  5. Model Context Protocol (28 July 2026): Authorization specification
  6. Amazon Web Services (accessed 10 November 2026): Configure and customize knowledge-base queries and response generation
  7. Bhatt et al., arXiv (18 September 2025): Enterprise AI Must Enforce Participant-Aware Access Control
  8. OWASP Cheat Sheet Series (accessed 10 November 2026): RAG Security
  9. Es et al., EACL and ACL Anthology (March 2024): RAGAs, Automated Evaluation of Retrieval Augmented Generation
  10. NIST (July 2024): AI 600-1, Generative Artificial Intelligence Profile

About the author: Piet Baudoin

I am Piet. I help startups with their brand, positioning and go-to-market, and I set up AI automations that turn customer notes into draft content and follow-up tasks. I discuss the choices with you so the work helps your business and its buyers. Want to apply this to your company? Try the growth scan or book a conversation.