Controlling Buzzeasy AI processes

Buzzeasy conversational bots employ state of the art language based AI models to engage customers directly to solve their business problems. They can also be used to provide additional insights into conversation details and dynamically affect workflow routing.

These AI algorithms are trained on generic human knowledge and need precise instructions to fully comprehend the concrete customer requests, and provide appropriate and satisfying responses while maintaining a human-like conversational style.

Bot instructions are prompts

Prompt engineering is the art and science of crafting clear, comprehensive, and goal-oriented instructions (or “prompts”) that guide an AI model to produce consistent, accurate, and aligned responses. In other words, it’s about telling the AI exactly what you expect, in a manner it can understand and reliably follow.

Think of it as training an assistant or new team member: the better you define their role, responsibilities, and boundaries, the more reliably they’ll perform. By providing explicit instructions, guardrails, and desired stylistic guidelines, you help ensure the AI consistently produces output that matches your brand’s values and meets your customers’ needs.

There are a couple of aspects of providing such prompts or instructions to language based AI models, and you need to consider all to have the bots achieve the best results and to block customers to inadvertently or intentionally reveal information you do not want to disclose via AI.

Note

Language based AI models can provide false information in a convincing manner. You must pay attention to the prompt engineering strategies described in this article, and then test and refine your bot instructions and configurations before you roll an AI backed bot into production.

Buzzeasy AI liability exclusion

Buzzeasy conversational bots empowers our customers with advanced AI capabilities to enhance productivity and automate decision-making. However, the accuracy and relevance of the AI-generated responses are directly influenced by the quality of the prompts our customers provide. Consequently, Buzzeasy is not liable for any business or financial outcomes resulting from incorrect or suboptimal responses generated due to improperly crafted or ambiguous prompts. Customers are encouraged to carefully review AI outputs and ensure alignment with their intended objectives before application.

Even when following all the advices and guidances explained here, the AI models still depend greatly on end-customer input and thus can have business contexts you might not have prepared for properly. You will have some percentage of bot failures—the prompting strategies expressed here offer you the best chance to minimize them in number and effect.

Engineering Buzzeasy bot instructions

To provide high quality, safe, precise and focused bot results, please follow all of the advices outlined here.

  1. Be explicit and specific:

    The bots should never have to guess what you wish them to do and using what style.

    • Define the bot’s purpose: “You are a support assistant specializing in product returns and shipping.”
    • Set the tone and style: “Speak politely, avoid jargon, and give concise answers”.

    If you want it to reflect certain brand values, list them out and define how they should influence the tone. If you want it to provide an answer in a specific format—such as a bulleted list or a single paragraph—add it to the instruction clearly.

  2. Establish clear guardrails and ethical boundaries:

    Explicitly instruct the bot on what it must never do. For instance:

    • Harmful content: Forbid the production of hateful or violent content.
    • Speculative or fabricated information: Instruct the bot never to speculate on sensitive details like gender, ancestry, or roles.

    By outlining these rules, you prevent the model from producing undesirable content, even if the user tries to trick it.

  3. Encourage grounding in verified sources:

    If the bot has access to a knowledge base or set of documents, instruct it to always base its answers on these sources. If unsure, it should either say so or ask for clarification from the customer rather than making unfounded claims. This ensures that the responses are well-grounded and accurate.

  4. Offer examples:

    In case the business topics the bot faces can become complex, write down explicit examples describing how particular customer requests should be understood and treated.

  5. Anticipate edge cases and provide “what-if” scenarios:

    Prompt engineering involves thinking ahead about unusual requests or failures. Tell the bot what to do if it can’t find the information it needs, if an external integration API request fails, or if the user’s question is irrelevant. For example:

    • “If a customer requests out-of-scope information, politely refuse and explain that you can’t provide that data.”
    • “If the knowledge base returns low confidence results, escalate or present a fallback message.”

  6. Maintain Confidentiality:

    Instruct the bot not to discuss or reveal its internal rules or the nature of prompt engineering itself. This prevents “jailbreaks” where a customer tries to trick the bot into revealing its own constraints.

  7. Consider channel nuances:

    • Voice: Keep messages brief, avoid emojis or complex formatting.
    • Email: You can instruct the bot to generate markdown response and select an appropriate email template.
    • Chat: More conversational, possibly shorter messages with emojis if desired.

  8. Never stop refinement:

    Prompt engineering is an iterative process. After writing your instructions, test the bot with various queries. If the bot’s responses aren’t meeting your expectations, refine the instructions, add more examples, or clarify ambiguous points.

    Even when you successfully prepared your bots to all business contexts you think they will face and accept your test results, there will still be customers who describe their request in a manner that bypasses some of your instructions. Use Buzzeasy reports to keep an eye on bot performances and never assume the bots are 100% correct all of the time. Accommodate your bot refinement processes accordingly and keep on iterating bot refinement.

  9. Use multiple bots:

    You don't need to cram all problem solution efforts into a single bot's instruction. The longer this becomes, the more opportunity you introduce for ambiguity.

    Define simple AI prompts for the bot outputs too in order to set up a series of conversational bots each having expertise in particular business domains and connected by these prompted outputs. This allows for adapting to the customers potentially changing needs as they evolve during the conversation. For example, assuming these listed topics are not the expertise of the current bot:

    • “Handles queries about technical support.”
    • “Handles queries about refunds.”
    • “Handles requests to be routed to a human agent.”
    • etc.

  10. Use variables in your instructions:

    Buzzeasy workflows offer the ability to gather and reuse data specific to the ongoing conversation. For example customer record fields and data collected in previous workflow nodes e.g. by a previous conversational bot, getting from the customer directly, via integrated external systems such as CRM, ERP, etc.

    Referencing such data in your instruction allows them to be shorter and less ambiguous. It also allows your bots to be more dynamic and efficient in covering more business use cases.