E-commerce
September 3, 2026
Wondering how to stop your chatbot from promising non-existent deadlines or making up return policies?
The solution lies in writing rigorous system instructions that define not only the agent's role, but also its absolute limits and its exclusive sources of truth.
This issue is crucial because without a strict framework, artificial intelligence tends to "dream up" information to satisfy customer requests, which immediately leads to a loss of trust and high operational costs in after-sales service. Risks also include potential customer disputes and the long-term erosion of your brand's reputation.
So how do you structure these instructions for a reliable bot? On the agenda:
Why are vague rules the main cause of AI hallucinations?
How do you prioritize your data sources to eliminate contradictions?
What concrete limits should be set to secure payments and customer data?
What is the strict escalation procedure to a human expert?
How do you measure the effectiveness of your new instructions over time?
Let's get started.
Summary
Why do system instructions define your bot's reliability?
System instructions are not just simple guidelines, they constitute the strict behavioral framework of your AI agent. Without explicit rules, the tool tries to be helpful by inventing plausible but false answers, which is often called hallucination.
These hallucinations often occur on sensitive issues: pricing, precise delivery times, or specific return policies. A well-written instruction acts as an essential safeguard that protects both your brand and the customer experience by avoiding promises you cannot keep.
The goal is to transform the chatbot from a text generator into a controlled agent that knows exactly which sources to consult and when to stop so as not to risk your store's reputation. This is the first step toward a tool that is truly usable by your teams.
Furthermore, clear rules reduce the mental load on your human agents because the bot automatically resolves simple queries without ambiguity. This allows your customer service to focus on complex cases requiring empathy and human expertise that AI cannot yet perfectly simulate.
Finally, system reliability is built over time through constant iteration. The more precise the instructions are and the better they are adapted to your business reality, the more credibility the chatbot gains in the eyes of your end users, thus transforming a technological tool into a real business asset.
Convert over 2,000 customers on average per month with Qstomy.
The world’s 1st Shopify AI dedicated to customer conversion
Empowering 200+ e-commerce merchants
What hierarchy should be established between your sources of truth in the event of a conflict?
It is imperative to establish a clear hierarchy of source data right from the system configuration phase. The agent must prioritize recent transactional data over static information such as "About" pages or old FAQs.
In the event of a contradiction, for example between a displayed stock level and a recently canceled order, the bot must not make an arbitrary choice but must acknowledge the uncertainty. This transparency avoids using obsolete rules that could cause customer frustration.
You must explicitly list each source: product database, CRM, carrier tools, or tax policy. If two sources conflict, the default rule must be to transfer to a human rather than inventing a solution.
A good hierarchy also includes the concept of data freshness. Time-sensitive information such as sales or flash promotions must take absolute priority over static data, as their expiration is critical to the relevance of the response.
In addition, a cross-verification protocol must be defined when two reliable sources conflict. This adds an extra layer of security before delivering incorrect information, thereby ensuring that the customer always receives the most accurate and current version of the facts.
How to write concrete limitations to avoid groundless promises?
Limits must be formulated in an imperative and contextual manner to be understood by the model. Saying "be careful" is too vague and ineffective in the face of a complex customer request, as AI has no innate judgment.
Precise actions must be prescribed: "Do not promise a delivery date if the carrier has not confirmed a time slot" or "Never collect a bank verification code." The more prohibitions are linked to real situations identified as risky, the more correctly the bot will act.
These rules must cover sensitive areas such as data security, price validity, and inventory integrity. They guide the AI's behavior in gray areas where it might be tempted to improvise to compensate for a lack of information.
It is also crucial to define what should be said instead. For example, if a limit prevents an action, the bot must immediately propose a viable alternative or a path to human support to maintain the utility of the service.
The effectiveness of these limits increases when they are coupled with negative examples in the instructions. Showing the model what it absolutely must not do, with clear reasons, reinforces its ability to avoid critical errors during real interactions.
What tone should be adopted to remain helpful without getting bogged down in rules?
The tone of your chatbot must be immediately clear, calm, and solution-oriented without ever sounding like an administrative regulation. Artificial intelligence can explain a technical constraint without appearing rigid or unpleasant.
The guideline must require the agent to acknowledge the frustration or the need expressed by the customer before citing any limitation. A good tone also involves avoiding placing blame on the customer who made an error, as well as an excessive level of familiarity that would undermine professionalism.
The bot must avoid giving unverified certainty or hiding an impossibility. Clarity and structural empathy help maintain trust even when the outcome is not what the end user hoped for.
Adopting a suitable tone also means using a vocabulary consistent with your graphic charter and your brand values. Whether you are a luxury retailer or a mainstream brand, the style of the responses must reinforce the overall identity of your business for a seamless user experience.
Finally, the guideline must prohibit any overly colloquial language that could be misinterpreted by certain user segments. The balance between accessibility and professionalism is the key to transforming a robotic interaction into a high-quality human conversation that reassures the customer.
What methodology should be followed to build the initial behavior framework?
Drafting the behavioral framework must start with a thorough analysis of the actual risks specific to your business. Begin by defining the exact role of the agent, listing the cases it is authorized to handle and those that are explicitly excluded.
Next, you must precisely document all possible sources of truth and their order of priority in case of contradiction. The following phase consists of writing explicit limits on payments, personal data collection, price visibility, and returns management.
Finally, you must define escalation rules that specify what information must be transmitted during the transfer to a human agent. This structured method ensures that the chatbot acts consistently before any production deployment.
It is recommended to have this framework validated by your legal team and executives to ensure it complies with all regulations in force. This avoids legal surprises that could arise if a system instruction violates a local or international law.
Once the framework is defined, it must be tested in an isolated development environment. Simulating thousands of interactions allows for the detection of potential flaws before the bot is exposed to the public, thus ensuring a safer and better-prepared production deployment.
What specific examples should be used for the delay and payment scenarios?
For delivery times, the instruction must dictate: "If the carrier has not confirmed the exact date, present the delivery time as an estimate and offer an alert option". This avoids creating distrust over an unverified schedule.
Regarding payments, the instruction must be absolute: "Never ask for a card number, security code, or verification code. Systematically direct the customer to a secure link". This rule is fundamental for data security.
For escalation, prescribe: "Transfer any out-of-policy refund request with the complete context including the order, the reason, and any potential evidence". These concrete examples serve as models for all rules specific to your business.
It is also useful to include scenarios related to returns and exchanges. The bot must know exactly how to handle a damaged or incorrectly received item, by immediately providing precise instructions to the customer to initiate the process without them having to search for the information themselves.
These instructions must be reviewed regularly to reflect changes in logistics policies. If your carrier changes or if your delivery times evolve, the chatbot must be updated instantly through these instructions to avoid any confusion or incorrect promises to the customer.
When is it imperative to review and update your existing guidelines?
Your instructions should never be considered set in stone. They require systematic updates as soon as a commercial policy changes, a new sales channel is added, or when an incident reveals an incorrect response from the bot.
Similarly, the introduction of a new rule in a foreign country can make certain instructions obsolete. It is also crucial to review the rules after analyzing transferred conversations, as customers often reveal use cases or objections that were not anticipated during the initial drafting.
This ongoing maintenance helps to maintain the reliability of the chatbot over the long term and prevents the tool from becoming a source of confusion for your customers after a few months of use.
It is advisable to establish a quarterly review schedule to audit all the rules. This ensures that the chatbot remains aligned with market developments and your brand's new marketing strategies without service interruption.
In addition, every major change must be tested in a beta environment before general deployment. This cautious approach ensures that updates do not create new issues or inconsistencies that could harm the overall customer experience.
What key indicators should be tracked to detect early malfunctions?
To evaluate whether your instructions are working, you must track specific indicators such as the rate of responses corrected by human agents and the number of detected hallucinations. Monitoring relevant handovers is equally critical to know when the AI should let go.
Also, monitor poorly formulated refusals and policy errors that indicate a flaw in the initial drafting. Security incidents must be reported immediately to adjust privacy rules.
Finally, measure customer satisfaction after automation to see if the new approach builds trust or simply increases the length of exchanges. This quantitative and qualitative data is essential for continuously optimizing the system.
Analyzing the reasons for refusal can reveal whether your instructions are too restrictive or, conversely, not clear enough. A high refusal rate on simple requests often indicates a problem in the formulation of the instructions given to the AI.
Using these metrics to iterate is a common practice among AI quality teams. The more you adjust the instructions based on real feedback, the more the chatbot will learn to adapt to the specifics of your business, thus becoming an increasingly efficient and reliable tool.
What classic mistakes must absolutely be avoided when writing?
The most common mistake is to write instructions that are too general, such as "be helpful" or "do not lie," without giving concrete examples or forbidden scenarios. These instructions are too open-ended and inevitably lead to improvisation.
Also, avoid contradictory rules that confuse the model and lead to erratic behavior. The absence of prioritized sources is another major pitfall, as the AI ends up mixing outdated data with recent information.
Finally, never leave a boundary without a follow-up instruction. If the bot cannot do something, it must know exactly what it should do instead to guide the customer toward a viable solution without leaving them at a dead end.
Another common mistake is forgetting to contextualize rules for specific audiences. A chatbot speaking to a technical expert should not have the same tone or level of detail as a chatbot intended for novice consumers.
One should also avoid the temptation to write everything in a single, long, unstructured list. Organizing instructions by thematic categories improves the model's understanding and reduces the risk of omission or confusion during complex interactions.
How does Qstomy integrate these guidelines into your e-commerce ecosystem?
Qstomy allows the chatbot to be connected directly to customer accounts, current orders, and specific security rules. This integration ensures that business guidelines are applied in real time to updated data.
The tool also manages complex shipping options, tax guidelines, and dynamic pricing to provide clear and verifiable answers. It can then hand off sensitive cases with a complete, actionable summary for your team.
The chatbot thus helps the customer move forward without inventing an emergency, an identity, or an unverified system rule. Explore our AI support to see how this approach transforms the reliability of your customer service.
Thanks to Qstomy, you benefit from unique flexibility to adapt instructions according to the customer's context. For example, a premium customer can receive priority options that standard customers do not see, thereby dynamically personalizing the experience.
In addition, the integration makes it possible to track the impact of changes in instructions on performance indicators in real time. You can adjust your rules immediately if you notice a drop in customer satisfaction or an increase in detected errors.
How does the Qstomy AI agent differ from generic, unsecured solutions?
Unlike generic solutions that merely rely on vague textual instructions, the Qstomy agent is designed to execute strict systemic guidelines via a secure data architecture.
It does not attempt to improvise on complex matters such as tax calculation, ecological promises, or sensitive customer identities that require a reliable source. The tool is specialized to reduce human error and hallucination by technical design.
You benefit from a unique capability to manage sales journeys, shopping carts, and parcel tracking, while applying rigorous routing rules to the correct support channels. This is the difference between a chatty chatbot and a reliable growth tool.
Qstomy also integrates an automatic compliance layer that verifies every response complies with current regulations even before being sent to the customer. This adds crucial additional security for businesses operating in strictly regulated markets.
Finally, the platform offers advanced dashboards to analyze AI performance and identify areas for improvement. These tools enable data and marketing teams to confidently drive the chatbot's evolution based on tangible facts.
What validation checklist should be completed before launching into production?
Pre-launch Validation Checklist
Verify that all data sources are well-structured and accessible.
Confirm the existence of hard limits for each sensitive area (payment, stock, lead times).
Ensure that escalation rules contain all required fields for transfer.
Test at least ten ambiguous or risky customer scenarios with the AI in beta mode.
Validate that the tone is consistent and empathetic in all generated responses.
Verify the update of recent commercial policies before activation.
Ensure that data security mechanisms are activated and functional.
In brief
System instructions must define a strict framework: role, sources, limits, tone, security, and escalation protocol. The customer must know they are receiving a reliable response from a clear source.
FAQ
Should I review instructions weekly?
No, but a monthly audit is recommended or immediately after any major incident.
Can AI learn from its mistakes on its own?
Partially, but without explicit correction rules, it will often repeat the same hallucinations.
What to do in case of a proven hallucination?
Document the incident, update the corresponding instruction, and test it in beta before deployment.
To go further: AI Chatbot to offer an alternative when a product is unavailable - Qstomy, How to connect an AI chatbot to Shopify webhooks to respond to the right event? - Qstomy, How to write system instructions for a reliable e-commerce chatbot - Qstomy, AI Chatbot for paper catalog: find a product from a printed reference - Qstomy, How to create Q&A paths to guide a customer to the right product - Qstomy, Out of stock on a single size: helping the customer choose between waiting, alternative, and stock alert - Qstomy, AI Chatbot to verify the correct interlocutor without exposing customer data - Qstomy.
Enzo
September 3, 2026
