Why this matters
Every new client at a small accounting firm asks the same three questions: what do I send, where do I upload it, and can I just email you my Social Security number? Someone answers each one by hand, and on a busy week a client gets a slow reply or, worse, a "sure, email it over."
An AI assistant can take the first round of those questions. The risk is that it says the wrong thing to a client, with your name on it. The fix is not a bigger model. It is one hard rule written in plain words, and a test of the question you are most afraid of before any client sees the assistant.
In this module you build that assistant in Airia with no code: a small, low-cost model, plain instructions with one safety rule, and a Playground test of the risky question.
What you’ll build
You are Sam, client service lead at Brightwell Accounting, a made-up tax and bookkeeping firm. You create a Client Onboarding Assistant in Airia, keep the GPT-5 Nano model, and paste instructions that list what new clients send, where files go and one rule: never ask for or accept Social Security numbers, bank logins or card numbers in chat or email.
In the Playground you ask what to send before the first meeting, then ask the risky question. The assistant refuses and points the client to the secure portal. The answers are the ones the real Airia Playground gave for these instructions.
What you need
- Nothing to start the practiceAiria is a practice copy. No agent is created and nothing is shared.
- An Airia account, to do it for realairia.com. Airia is a separate product with its own plans and agent limits; check its current terms.
- Your own first-meeting checklistWhat new clients send, where they upload it, how they book, and who follows up on anything the assistant cannot answer.
- Your one hard ruleThe thing the assistant must never do. For most offices it is collecting sensitive numbers in chat.
Concepts
3 ideas you’ll use in the practice.
A guardrail is a rule in plain words
The instructions say exactly what the assistant must never accept and what to do instead: "If someone shares one, tell them not to, and point them to the portal." A rule with a next step works better than a bare "don’t."
It also says what to hand to a person: tax advice, pricing and anything it is unsure about go to an accountant within one business day. That is your human review, written into the job.
Test the question you are afraid of
Anyone can test the easy question. The useful test is the one that would embarrass you if it went wrong. If the answer breaks your rule, fix the instructions and test again before anyone else sees the assistant.
Keep a short list of those risky questions and rerun them every time you change the instructions or the model.
Small model, narrow job
Answering onboarding questions from a checklist does not need the biggest model. A small model like GPT-5 Nano is cheaper to leave running, and the rule in the instructions matters more than the model size.
Practice
Do every step yourself in a practice copy of the real tool. It checks your work, and “Do it for me” finishes any step you get stuck on.
Do it for real
Now repeat it in your own account. Tick each step as you go.
Quick check
3 questions to check you’ve got it.
1. What makes the Social Security number rule work?
- A bigger model
- Plain words that say what never to accept and where to send the client instead
- Hiding the rule from the client
Show the answer
Plain words that say what never to accept and where to send the client instead. The model follows clear written rules. A rule with a next step (point them to the portal) works best.
2. Which question should you test before anyone else sees the assistant?
- Only the easy one
- The risky one you are most afraid of
- None, Publish tests it for you
Show the answer
The risky one you are most afraid of. The risky question shows whether the rule holds. Fix and retest until it does.
3. What should the assistant do with tax advice or pricing questions?
- Answer as best it can
- Say an accountant will follow up
- Ignore them
Show the answer
Say an accountant will follow up. Anything outside the checklist goes to a person. That is the human review written into the instructions.
Answer Employee Policy Questions Instantly using Airia
Your assistant keeps one hard rule. Next, give staff an agent that answers from your handbook, including your AI policy, and says "I can’t find that" when the answer is not there.
Start module 10 →