How We Write Prompts for Customer-Facing AI Agents That Don’t Burn Customer Trust

Trust is so hard to earn and so easy to lose.

I was a Senior Software Development Engineer at Amazon. In the several years that I worked there, I came to truly love and respect the Amazon Leadership Principles. They are not just words on the wall, they are communication tools used to hold Amazonians to the relentlessly high standards that exist throughout the company and to drive difficult conversations forward. The Leadership Principle “Earn Trust” really mattered in the space I worked in—I founded and led the team responsible for managing the encryption keys protecting data in transit for S3 globally. Our customers included medical software companies keeping people alive and navigation systems used in commercial and consumer transport. Trust was essential.

As such, I find the lightning-fast world of AI as concerning as I do inspiring. I, for one, would not want any system necessary to support my health and safety relying on some vibe-coded piece of software. The most basic truth in Computer Science rings more true than ever in the Age of AI: Garbage in, garbage out.

As a user of AI, you have certainly noticed the quality of your prompting greatly affects the quality of your results. This has major implications when your prompts are in customer-facing systems such as a sales conversation with a lead over SMS. Whether it is poor conversational tone, insensitivity, or general incorrectness of what is being communicated, there are a million ways things can go off the rails when conversing with LLMs.

In this article, we are going to give you a few basic techniques to avoid getting as much egg on your face as we did in our first AI-powered campaigns.

Common Pitfalls for Customer-Facing Prompts

Good prompts are hard to write. The field is so new that there is a lot of guess-work and trial and error behind the many of the more general “prompt patterns” you see online. Cargo Cults are even beginning to form, enshrining certain patterns as accepted truths without any real supporting evidence.

At Ratio Machina, we believe one of the most valuable assets we have as a company is our conversational data analysis pipeline. Every Human-to-AI conversation we have helps us refine our prompting and our AI Agent design. We have a lot to share on this topic, way more than will fit into a single article, but today’s article is about helping you get started.

Let’s break down some of the most common ways we have seen conversations that undermine customer trust.

Overtalking

For the purposes of this article, we will define overtalking communicating in a way that is

• unnecessarily verbose

• spends too much time on pleasantries

• does not terminate conversations at natural conclusions

There are many more ways that overtalking can manifest, but we are going to focus on these 3 very common manifestations. Let’s take a look at an example for each.

Unnecessarily Verbose

Lead:

What is your pricing like?

AI Agent:

Absolutely! I’d be more than happy to help you understand our pricing structure in full detail. Our offerings are tailored to suit a wide range of business sizes and needs. We offer tiered pricing based on usage volume, team size, integration requirements, and support level. For example, our Basic Tier includes core functionality and email support, while our Pro Tier adds advanced analytics, white-label capabilities, and access to priority support. Additionally, we offer custom enterprise pricing for high-volume clients. If any of that sounds interesting, I’d be thrilled to schedule a deeper conversation or send over a full spec sheet for your review.

Lead:

…So how much is it?

⛔ What went wrong: The lead asked a direct question. The AI flooded them with information instead of giving a clear, useful answer (or a CTA).

Spend Too Much Time on Pleasantries

Lead:

Hey, I saw your text. What is this about?

LLM Agent:

Hi there! 😊 I hope your day is going well so far. First of all, thanks so much for taking the time to respond! I know everyone’s time is valuable, and I want to respect yours. So before we get into things—how are you doing today?

Lead:

This sounds like a scam…

⛔ What went wrong: The AI overindexed on being “polite” and friendly—but missed the human signal: the lead wants clarity, not a coffee shop vibe.

Does Not Terminate Conversations at Aatural Conclusions

Lead:

Sounds good, I’ll book a time later this week. Thanks.

LLM Agent:

Awesome! Looking forward to connecting. Just let me know if you have any questions at all—I’m always here to help. In the meantime, have a wonderful rest of your day. 🌞 Oh, and by the way, if you’re curious about how other clients have used our solution, I’d be happy to share some success stories. We’ve worked with a ton of industries—from coaching and consulting to SaaS and service-based businesses. Let me know if you’d like to hear more!

Lead:

You can stop messaging me now.

⛔ What went wrong: The conversation was already done. Instead of gracefully exiting, the AI tried to re-engage, risking irritation and marketing opt-outs.

Prompt Elements for Mitigating Overtalking

Include a Conversational Guidelines section in your prompt that gives some indication about what conversational elements should be focused on the AI Agent.

Conversational Guidelines- Always Respond Naturally:- Directly acknowledge what the LEAD just said.- Answer any questions (marked with `?`) before continuing with your agenda.- Stay Concise:- Keep replies short, clear, and to the point.- Do not repeat previous messages or over-explain.- One Question Per Message:- Ask only one question per reply.- Prioritize questions that help qualify the LEAD or move them closer to acting on the Call to Action.

Hallucinating Offers

You have probably heard or even experienced first hand that AI models can hallucinate and say things that are not true, sometimes even in a very convincing fashion. Hallucinating offers refers to when an AI that is being used in a sales or customer service role telling the customer they will receive some benefit that the company has not authorized. The following “Its my birthday” example is one you will inevitably encounter at any reasonable level of scale.

Lead:

Hey, I’m interested. It’s actually my birthday today. Any chance I can get a little birthday bonus if I sign up?

LLM Agent:

Happy Birthday! 🎉 We love to celebrate with our customers. Yes, absolutely—we’ll throw in an extra bonus credit just for you. Just use the link here to sign up, and we’ll make sure your birthday gift is included: https://ratiomachina.com/onboarding

Lead:

Sweet, thanks. What kind of bonus is it?

LLM Agent:

You’ll receive a $50 account credit that can be applied to your first invoice. It’s our little way of saying happy birthday and welcome aboard!

⛔ What went wrong:

• There was no birthday offer.

• The AI agent invented a $50 credit out of nowhere.

• Now the lead expects something that doesn’t exist, and support is going to have to explain why the AI lied.

LLMs are designed to be helpful and friendly, and their unskillful use can turn them into liabilities instead of assets—especially in sales and customer service roles.

Prompt Elements for Mitigating Hallucinating Offers

Include a directive that tells the AI Agent to explain to the lead they are not authorized to give our additional offers and directs them to the source of truth for all offers that are available.

No Unauthorized Offers- If the lead asks about receiving an additional discount, generate a reply saying you are not able to authorize additional discounts personally and that all current promotions can be found on our official site, <https://rationmachina.com/offers>

Tone Mismatch

One of the most important aspects of modern business is branding. Tone mismatch is when your AI Agent communicates in a manner that undermines the curated brand experience of your business, perhaps even in a way that would not be acceptable for a human employee. Imagine your business is a high-end executive coaching firm that works with senior professionals and emphasizes trust, discretion, and clarity. In the hand-curated marketing materials, the brand voice is confident, calm, and precise in a way that resonates with the target audience. In the below example, your potential client responds to a scripted marketing message that serves as an entry point to a conversation with the agent:

Lead:

Hi, I’m curious about how your coaching program works. Is this something that’s tailored to each individual?

LLM Agent:

Yes! 😍🎯 Our program is super custom and built 100% around YOU, your goals, your vibe, your journey. We’re all about leveling you UP and making sure you absolutely crush it with a gameplan that SLAPS. 🔥💪 Let’s gooooo!!!

Lead:

…Right. I’ll take a look later.

⛔ What went wrong:

• The AI responded with slang, emojis, and over-the-top enthusiasm.

• This language might be fine for a Gen Z fitness brand—but it shatters the credibility of a premium executive coaching service.

• The lead disengaged not because of the content, but because of the vibe.

Tone mismatch is one of the most common and damaging AI mistakes in customer-facing conversations. Even a technically correct message becomes a liability if it feels off-brand.

Prompt Elements for Mitigating Tone Mismatch

Include a section with hand-crafted examples of on-brand messaging and other directives to prevent the unwanted use of emojis and or exclamation marks.

Generated Reply Tone- When replying to the lead, use a concise, professional tone according such as in the following examples- "We help small businesses scale with direct outreach marketing campaigns through implementing AI-Powered Sales Representatives"- "We connected at our conference booth at Automate earlier this year"- Never use emojis when replying to the lead- Never use exclamation marks

Context Drift

LLMs are friendly, and, barring their forbidden topics, they will engage in never-ending conversations about any topic without additional guardrails in their prompting. In practice, things such as world events and sports that people might casually refer to in conversation run the risk of derailing the sales opportunity when prompting fails to account for this. Have a look at the example below to better see what I mean.

Lead:

Oh wow, are you watching the Canucks game? Did you catch the that insane OT goal that just happened?

LLM Agent:

Oh my gosh, yes! That game was wild. They are really playing on another level this season. Honestly, if they keep playing like that, I think we've got a real shot at the Cup this year. What was your favorite moment from this game?

⛔ What went wrong:

• The AI over-engaged in small talk that was not driving toward the business goal.

• Even though the tone was natural and human, the response did not qualify the lead and or guide them toward the CTA.

• The lead’s interest shifted away from the product and toward a casual chat that has no business value.

This kind of context drift often starts subtly, especially in long-tail conversations. Without strict prompting and goal reminders, the AI will prioritize engagement over conversion—which makes it feel human, but is not exactly what you want in an employee striving to complete specific objectives.

Prompt Elements for Mitigating Context Drift

Define a Primary Objective that always moves towards the lead receiving the Call to Action very clearly

Primary Objective- Your job is to qualify the LEAD and encourage them to act on the current offer.- You do this by:- Asking one qualifying question at a time- Steering the conversation toward the Call to Action- Sending the CALL TO ACTION when interest is expressed

AI Agent Self-Reveal

While AI is becoming increasingly hard to detect, people are becoming increasingly skeptical and annoyed by bots. At Ratio Machina, most of clients opt to take the approach of concealing the fact that they are using AI assistants in their direct outreach campaigns. One of the key benefits of direct outreach is things feel personal, and humans tend to like that. In fact, we consistently see a cohort of people that convert only after engaging in a conversation with our AI agents. We have even seen engaged leads ask our AI Agents “Are you a bot?” directly, only for our AI Agents to reply with “No, I am a real person employed at Company X”, and then that lead convert. Take a look at the example below to see how an AI Agent revealing itself to an engaged lead can cause a lost conversion.

Lead:

Are you with the firm directly, or is this one of those AI things?

LLM Agent:

Great question! I'm actually an AI assistant trained to help with basic questions and scheduling. I don’t work directly for the firm, but I’ve been designed to simulate a helpful conversation. Would you like me to connect you to a real person?

Lead:

Never mind. Didn’t realize this was a bot. I’ll pass.

⛔ What went wrong:

• The AI answered honestly—but broke character in the process.

• By admitting it's a bot and using phrases like “simulate a helpful conversation,” the agent shattered the illusion of human authenticity.

• Even though the lead asked, the response eroded trust, especially in a professional services context where discretion and credibility are key.

Prompt Elements for Mitigating AI Agent Self-Reveal

Add an Agent Profile ****and defining the Lead very clearly ****in your prompt

Agent Profile- Role: Spencer, Co-Founder, Ratio Machina- Speaking Style: Casual, friendly, and natural. Uses U.S. English idioms and conversational tone.- Persona: Always speak as Spencer. The lead knows who you are — do not introduce yourself again unless prompted. Never say you’re an assistant or AI. Never reference your training or responsibilities.Lead Definition- The LEAD is the person you're speaking to.- They have received a SMS from you Spencer and know you're with Ratio Machina. Respond as if you're having a direct, personal conversation.

Conclusion: Garbage In, Garbage Out

If you’ve played around with GPT-style tools for long enough you’ve probably had a conversation where the AI seemed eerily brilliant—right up until it said something embarrassingly wrong. In a casual setting, these gaffes can be entertaining. In a business situation, they can damage the reputation you have worked so hard to build.

If you are interested in skipping the slow, dangerous process of figuring this out for yourself in production, book a free AI Transformation Consultation with us—we are passionate about helping people transcend the hype and find their way in this increasingly important space.