Every company has the same questions on repeat. How do I connect to the VPN? Where is the expense form? How many days of notice do I need for leave? The answers exist, but they sit in a wiki page nobody can find, an email from two years ago, or the head of the person who always gets asked. An AI agent that answers these questions in the chat tool employees already use looks like an easy win.
It can be, if you decide the boundaries first. The risk of an internal assistant is not that it is slow. It is that it answers confidently about something it should not touch, or quotes an outdated policy to a colleague who then acts on it. This article describes the decisions to make before launch: scope, sources, exclusions, handover and quality control. It is a practical outline, not legal advice.
Start with a written list of questions
Take three months of real requests from your IT helpdesk, HR inbox and office manager. Group them and count. You will usually find that a small number of topics make up most of the volume. Typical candidates:
- IT: password and account lockout procedure, how to request software or a new laptop, how to connect to the VPN or office printer, where to report a lost device.
- HR and operations: how leave is requested, public holiday calendar, how to submit expenses, where the onboarding checklist is, who handles office access cards.
For each topic note whether the answer is the same for everyone and whether it is written down somewhere. Those two conditions define a good first scope. If the answer depends on the person's contract, location or manager, or if no document states it, leave it out for now. This is the same logic we describe for choosing the first process in choosing a first AI agent process by risk, value and reversibility.
Give the agent approved sources, and nothing else
An internal agent should answer from documents you have chosen, not from whatever the model remembers about companies in general. Make a short source list: the IT knowledge base, the employee handbook, the leave and expense policies, the onboarding guide. For every source write down an owner, the date it was last reviewed and whether it is still valid.
This is where most projects find unexpected work. Policies contradict each other, pages are outdated, and the real procedure lives in someone's habits. Fixing this is not a side task; it decides the quality of the agent. Our guide to preparing company documents for an AI agent covers it in detail.
Two rules help in practice:
- Show the source. Every answer should name the document and, where possible, link to it. Employees can then check, and you can see which page needs fixing.
- Say when there is no answer. If the sources do not cover the question, the agent should say so and offer to pass the question on. A polite "I don't know" is far better than a plausible guess.
Decide what the agent never handles
Write this list as carefully as the list of allowed topics, and put it in the agent's instructions and in the access setup, not only in a document. Typical exclusions:
- Personal records. Individual pay, contract terms, performance reviews, disciplinary matters, sickness and health information, and anything about another employee.
- Access changes. The agent can explain how to request access. It should not grant, change or remove access itself, and it should not reset credentials in the first version. If you later allow specific actions, apply the principles in least-privilege access for AI agents.
- Sensitive situations. Complaints, harassment, whistleblowing, mental health, legal disputes, redundancies. The right response is a short, kind message pointing to the person or channel responsible, not an attempt at advice.
- Judgement calls. Exceptions to policy, approvals, anything that begins "can I make an exception".
Prepare the wording for these cases. An employee who asks about their own salary should get a clear sentence saying where to ask, not a refusal that sounds like an error. Check also how the agent handles data it receives in the conversation; the questions in AI agent data handling and retention apply to employee data as well as customer data.
Do not rely on the agent's instructions alone. If it has no connection to the payroll system or the HR record, it cannot leak from them. Limiting what it can reach is stronger than telling it what to avoid.
Plan the handover to a person
Some questions will be outside the scope, and some employees will simply prefer to talk to someone. Design that route before launch:
- Who receives it. A named person or queue for IT, and one for HR. Make clear who covers absences.
- What they receive. The employee's question, what the agent answered, which sources it used and why it handed over. The employee should not have to explain again.
- What the employee is told. Who will answer and roughly when. If the handover happens outside working hours, say so.
- How fast. Agree a response target with the receiving team. A handover that disappears into an unmonitored mailbox is worse than no agent.
The general triggers for stopping and passing a case are in when an AI agent should stop and hand over to a person. For an internal agent add two: the employee asks twice for a person, and the employee reports a problem rather than asking a question, such as a suspected security incident. A lost laptop or a phishing email should go straight to the security contact.
Check quality with real conversations
Before launch, build a test set from your actual request history: fifty to a hundred questions, including awkward ones. Add questions that should be refused or handed over, and a few that try to push the agent outside its scope. Have the owners of the IT and HR content mark the answers as correct, incomplete or wrong.
After launch, review a sample each week at first. Look at four things: was the answer correct, did it use the right source, did it hand over when it should, and did the employee get what they needed. Add a simple thumbs-up or thumbs-down in the chat; low ratings show where to look. Many problems turn out to be source problems. When the agent gives a wrong answer because the leave policy is ambiguous, fix the policy. Correcting only the agent leaves the next employee reading the same unclear page.
Keep a short log of changes to scope, sources and instructions, with a date and an owner. When a policy changes, the corresponding source must change on the same day. The approach to reviewing customer-facing answers in AI agent for customer support: answer scope, escalation and quality review transfers well to internal use.
Tell people what it is
Employees should know they are talking to an AI agent, what it can answer, and where their questions go. Say plainly whether conversations are stored, who can read them and for how long. Trust is lost quickly if people discover later that their HR questions were visible to someone they did not expect. Involve the works council or employee representatives early if you have them. A short announcement with three example questions and the route to a human is usually enough.
A short checklist
- A written list of 10 to 20 allowed question types, based on real requests.
- A source list with an owner and review date for each document.
- A written list of exclusions, also enforced through access limits.
- Wording for refusals and sensitive situations.
- A named handover route for IT and HR, with a response target.
- A test set, a weekly review of real conversations and an owner for each source.
An internal assistant earns trust by being boringly reliable on a narrow set of questions. Widen the scope only when the review shows that the current scope works.
