[← All articles](https://noodleseed.com/blog)[AI guides](https://noodleseed.com/ai/blog) Founder notes · LLMday # Your AI can answer. Can your business stand behind the answer? Five questions before your AI assistant changes a booking, issues a credit or makes a promise. A practical guide to customer trust and business rules. [Fahd Rafi](https://www.linkedin.com/in/fahdrafi/) Founder, Noodle Seed 3 October 2026 · Adapted from LLMday, Redwood City ![Noodle Seed editorial illustration: Your AI can answer. Your rules decide.](https://noodleseed.com/blog-artwork/v1/your-ai-can-answer-cover.jpg) Imagine a customer asking your assistant for a service credit. The assistant checks the account, prepares a $12,000 credit and sends it to your billing system. The response never arrives. So it tries again. Now your customer has received $24,000. This is a fictional example I used in my LLMday talk. The amount makes the consequence easy to see, but the same problem can start with an everyday booking or refund. The assistant did something useful, then repeated it because it did not know what had already happened. That is the part of the AI conversation I think businesses need to spend more time on. An impressive answer can earn attention. A change to a customer’s account asks the business to take responsibility. ## Three things to take away - Decide what your assistant may promise and change before customers depend on it. - A customer’s confirmation should apply to the exact details they reviewed. - When a result is uncertain, preserve the request, check what happened and give someone ownership of the next step. ## The answer becomes your promise When an assistant says a table is booked, a customer may arrange their evening around it. When it says a refund is complete, they may stop looking for a solution. Your business inherits those expectations even if the words came from a model. This is why I draw a boundary I call the deploy line: where generation ends and responsibility begins. On one side, the assistant can suggest an option or prepare a change. Across that line, it uses your business’s authority to do something that matters. For a business owner, that boundary should be understandable. You should be able to name what the assistant can do, which customers or accounts it can act for, when someone needs to approve a decision and who handles an unresolved result. Those choices are part of the service you are offering. > The deploy line is where generation ends and responsibility begins. ## Give “yes” something concrete to mean A customer agreeing to a booking should see the date, the price and the terms they are agreeing to. A team member approving a credit should see the amount, the account and the reason. A simple “go ahead” only helps when everyone is talking about the same proposed change. If the assistant finds a different price after the customer confirms, it needs a new decision. If the account changes, the earlier approval should not quietly travel with it. Good service includes making those changes visible before acting. You also need to decide which actions require a person. Answering an opening-hours question and issuing a large credit carry different consequences. Start with rules that reflect your actual business, including spending limits and an escalation path someone can use. ## A missing response needs a careful next step In the fictional credit example, the billing system may have applied the credit before the connection failed. A missing response tells the assistant that it does not know the outcome. It does not tell it that nothing happened. The customer experience should reflect that uncertainty. Keep the request and its details. Check the business system for the result. If the result cannot be confirmed, give a named team or operator enough context to investigate. An honest “we’re checking” is more useful than a premature confirmation. ![An illustrative customer journey: request a change, review its exact details, check the business system, then confirm the result or hand an unresolved outcome to a person.](https://noodleseed.com/editorial/deploy-line/customer-promise.svg) Scroll horizontally to view the full diagram. A conceptual service journey. Confirmation follows evidence from the business system. An unresolved outcome has a human owner. Customers should not need to repeat the whole story when something goes wrong. If a person picks up the conversation, they should know what was requested, what was approved and what the assistant attempted. That handoff is part of the product’s quality. ## Five questions before you let an assistant act I use five questions to make this discussion concrete. You can take them into a conversation with your team or an AI supplier and apply them to one actual task. 1. **Whose request is this?** Can the system verify the customer or employee and the account they are asking it to change? 2. **What is it allowed to do?** Do your business rules and the person’s approval cover this exact action? 3. **Which access is it using?** Is it connected to the right business account with the permissions needed for that task? 4. **What happens if the response disappears?** How does it check the outcome and avoid turning an uncertain result into a second change? 5. **Who can explain it afterward?** Is there a useful record of the request, approval, attempted action and confirmed result? Ask to see the answers in a worked example, including something going wrong. A successful demonstration tells you what can happen on a good day. A careful recovery tells you much more about the service your customers will receive. ## Responsibility needs an owner [Watch the captioned LLMday excerpt](https://noodleseed.com/editorial/deploy-line/03-who-pays.mp4) From the talk: a buyer’s question about the cost of a mistake. This question does not establish a reimbursement commitment. Captions are included in the video. One of the questions raised in the talk was whether the cost of an assistant’s mistake would be reimbursed. It is a reasonable question for a buyer to ask. It also brings the discussion back to the business arrangements around the software: who investigates, who can correct a change and what the agreement actually covers. That question is not a promise of reimbursement. It is a reminder to discuss responsibility before an incident. Keep the operating rules, support path and contractual commitments clear enough that a business owner can understand them. My suggestion is to start with one useful action. Choose something customers already ask your team to do. Define its limits, the details a customer reviews, the evidence needed to call it complete and the person who owns uncertainty. That is the kind of progress I want us to make with AI: a better customer experience supported by business rules people can explain. The words can feel effortless because the work behind them has been thought through. ## Take the questions to your team [Download the two-page deploy line field guide](https://noodleseed.com/editorial/deploy-line/deploy-line-field-guide.pdf) and use it to review one consequential action. For engineers implementing these boundaries, read [the technical companion on noodleseed.dev](https://noodleseed.dev/insights/the-deploy-line). ## About the author [Fahd Rafi](https://www.linkedin.com/in/fahdrafi/) Founder, Noodle Seed Fahd Rafi is the founder of Noodle Seed. This article develops the argument from _The Deploy Line: When Software Becomes Infinite_, presented at LLMday in Redwood City on 1 October 2026. ## Your business. A conversation away. Get early access for what’s next from Noodle Seed. [Get early access](https://noodleseed.com/waitlist) --- Source: https://noodleseed.com/blog/your-ai-can-answer [All public pages in Markdown](https://noodleseed.com/llms.txt)