AI for Internal Communications: Updates, Questions, Tracking

AI for internal communications works best as an agent that drafts updates from approved facts, answers repeated employee questions from a current source of truth, and keeps a record of what was announced to whom. A person still approves anything that goes out and owns the policies behind it. The gain is consistency and memory, not a higher volume of messages.

Key takeaways

• Start with one narrow role: a drafting and answering agent for a single audience, with a named human approver.
• Employee answers should come from a small set of approved documents, and the agent should say so when it cannot find one.
• Keep an announcement record (what, to whom, through which channel, when, approved by whom) so that nobody has to reconstruct it from memory.
• Permissions matter as much as prose: the agent should only read and post where its role allows.
• Judge the agent by observable outcomes, such as fewer repeated questions reaching managers and fewer conflicting versions of the same message.

What does AI for internal communications actually cover?

Internal communications is less about writing than about keeping a company telling the same story to many groups. A policy changes, a system moves, a leader leaves, a shift pattern is adjusted. Each change has to reach the right people, in words they understand, through a channel they actually read, and then the follow-up questions have to be answered the same way every time.

An agent can take on three parts of that work:

• Drafting: turning approved facts and a short brief into a message for a specific audience, in the company's own voice.
• Answering: responding to recurring employee questions in chat or email, using current documents and escalating what it cannot support.
• Tracking: recording which announcement went to which group, what it said, and who approved it.

What it should not do is decide what the company's position is. Positions, tone on sensitive topics, and anything touching employment terms, redundancies, health or legal exposure stay with people. The agent prepares, retrieves and remembers. Humans decide and sign.

How can an agent draft company updates without sounding generic?

Generic drafts come from generic inputs. If the only instruction is "write an announcement about the new expense process", the result will be polite, vague and slightly wrong. A good drafting agent works from a brief with fixed fields, plus a standing style reference.

A workable brief contains:

• The fact: what changed, effective when, and what stays the same.
• The audience: which team, location or role, and what they already know.
• The action: what, if anything, people must do, and by when.
• The source: the policy or decision document that backs the message.
• The approver: the person who must sign off before it goes out.

The style reference is a short set of approved past messages and a plain list of house rules: how the company addresses people, which words it avoids, how long updates should be. The agent drafts from the brief, and it flags gaps instead of filling them. If the brief does not say when a change takes effect, the draft should contain a visible question for the approver, not an invented date.

It also helps to have the agent produce variants from one approved core. The same change might need a short chat message for a warehouse floor, a fuller email for office staff, and a two-line note for managers with talking points. All three trace back to one approved fact sheet, which is what keeps them from drifting apart. Teams that already run marketing workflows will recognize the pattern, and the same discipline appears in AI agents for in-house marketing teams, where briefs and approvals govern what gets published.

How can an agent answer recurring employee questions safely?

Most internal questions are repeats: how to request leave, who approves a purchase, where the latest policy lives, what the new schedule means for a particular role. Answering them by hand takes time from managers and HR, and the answers vary with whoever happens to reply.

An answering agent should use retrieval: it looks up passages in approved documents and builds its reply from them, rather than answering from general knowledge. [AWS describes retrieval-augmented generation](https://aws.amazon.com/what-is/retrieval-augmented-generation/) as grounding a model's output in an authoritative knowledge source outside its training data, which is the behavior you want here. The practical rules around it are simple:

• Every answer points back to the document it came from, so the employee can check it.
• Documents have owners and review dates, and the agent ignores or flags anything past review.
• If no approved source covers the question, the agent says so and routes it to a named person instead of guessing.
• Sensitive categories (pay disputes, grievances, medical matters, immigration, disciplinary issues) go straight to a human and are not answered by the agent at all.
• Employee questions are logged, so the company can see which topics keep recurring and which documents are unclear.

That last point is an underrated benefit. A list of the questions people keep asking is a map of where your communication is failing. If ten people ask what a new rule means for night shifts, the announcement did not say it.

There is also a security side. An agent that reads shared documents and messages can be steered by text hidden inside them. [OWASP's prompt injection prevention guidance](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) covers treating retrieved content as untrusted input and limiting what the model can do with it. For an internal agent, that means it should read from a curated library, not from every inbox and chat room, and it should not be able to send messages to broad audiences without approval.

How do you track what was announced to whom?

This is the part most companies handle worst, and it is where an agent with memory is most useful. Weeks after a change, someone asks whether the contractors were told, whether the second site got the same notice, or which version the sales team saw. Nobody can answer quickly, because the record lives in scattered inboxes and chat threads.

An announcement record fixes this. Each entry holds a small set of fields:

• Title and a short summary of the message.
• The exact text that was sent, and the source document behind it.
• The audience: teams, locations, roles, and any groups deliberately excluded.
• The channel and the time it went out.
• The approver, and whether the message was revised after approval.
• Open follow-ups: questions received, corrections issued, acknowledgments still missing.

With that record, the agent can answer questions like "which groups have not yet received the revised travel policy" or "what exactly did we tell the night crew about the schedule change". It can also catch conflicts: if a new draft contradicts a message sent earlier, it flags the difference to the approver before anything goes out.

The record should be treated as a business record, with an owner, a retention rule and access limits. Who can see the full history of announcements, including drafts that were never sent, is a permissions decision, not a convenience. [OWASP's authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) recommends least privilege and checking permissions on every request, and the same logic applies to the agent's own access: it should reach only the channels, folders and audiences its role requires.

What does this look like in practice? An illustrative scenario

This is an illustrative scenario, not a real case.

A distribution company with an office team and a shift-based warehouse crew changes its rule for swapping shifts. The operations director writes three lines of facts and attaches the updated policy. The agent drafts a short message for the warehouse chat, a fuller email for office managers, and a talking-points note for shift supervisors. Each draft cites the policy and leaves one visible question: does the rule apply to temporary staff?

The director answers the question, edits one sentence, and approves. The agent posts to the approved channels and writes the entry in the announcement record, including the audiences and the time. Two days later, an employee asks in chat whether swaps need a supervisor's sign-off. The agent answers from the policy, quotes the relevant line, and links the document. Another employee asks whether swaps count towards overtime. No approved source covers it, so the agent says so and sends the question to the HR contact with the thread attached.

At the end of the week, the agent gives the director a short summary: which questions came in, which one went unanswered, and which team has not yet acknowledged the message. The director updates the policy to cover overtime, and the agent drafts the clarification for the same audiences, linked to the original entry in the record.

What is a sensible step-by-step plan for a first role?

A first engagement should be small enough to run for a few weeks and judge honestly. A plan that works for most companies:

• Choose one audience and one topic family, for example office staff and HR policy, or shift crews and scheduling.
• Name the approver for outbound messages and the owner for each source document.
• Gather the approved documents and retire anything outdated, conflicting or unowned. This is usually the most useful step on its own.
• Write the brief template and the style reference from real past messages that the company considers good.
• Define what the agent must escalate and never answer, and name the human who receives each category.
• Connect only the channels the role needs, with posting restricted to drafts or approved messages.
• Start the announcement record from the first message, even if it is just a structured list.
• Run in review mode first: the agent drafts and suggests answers, people send them, and the team compares.
• Set an observable success criterion, such as whether managers still field the same repeated questions, or whether any two audiences received conflicting wording.
• Review the question log regularly and fix the documents behind recurring confusion.

Only after that review period is it worth letting the agent answer directly for the lowest-risk topics. Posting to broad audiences without approval should stay off for much longer, and often permanently. Several of these steps overlap with other internal functions: the document hygiene and escalation design are the same ones used in AI agents for IT helpdesks, and a company that also runs recurring reports can reuse the pattern described for operations teams.

How Be Human approaches internal communications agents

Be Human designs, implements and operates the agent as part of a defined role. A first engagement sets out the audience, the source documents, the approval boundaries and the success criterion, and decides which channels the agent may read and where it may post. Managed operations can then cover monitoring of answers and escalations, recurring evaluations against real past questions, integration maintenance, permission reviews and controlled improvements to the briefs and the announcement record. The company keeps control of the content, the approvals and the access, and the framework for judging risk can follow guidance such as the [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework), which helps teams name and review the risks of a specific use.

Frequently asked questions

Can an AI agent send company announcements on its own? It can, technically, but for most companies it should not. The agent should draft and prepare, and a named person should approve anything that reaches a broad audience. Autonomous sending can be considered later for narrow, low-risk, templated notices.

What should an employee-facing agent refuse to answer? Anything involving individual pay, performance, discipline, health, legal exposure or grievances should go to a person. The agent should also decline when no approved document supports an answer. A clear handoff, with the question passed on and the employee told who will respond, is better than a confident guess.

How do we keep the agent from repeating outdated information? Give every source document an owner and a review date, and let the agent use only current, approved material. When a policy changes, update the document first and then draft the announcement from it. Reviewing the log of questions also shows where old wording is still circulating.

Does this replace the internal communications team? No. It takes over the repetitive parts: first drafts, repeat questions and record keeping. The team keeps the judgment work, such as deciding what to say, when to say it and how to handle sensitive moments, and it gains a clearer record of what has already been said.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.