AI Performance and Call Quality Overview

Maple's AI voice agent is designed to handle restaurant phone calls efficiently and naturally. Like any intelligent system, its performance depends on proper setup, ongoing optimization, and continuous learning. This FAQ covers what you can expect from AI call handling, how quality is measured, common issues you may encounter, and how the system improves over time.

How accurate is the AI at handling calls?

Maple continuously tracks and improves AI call handling performance. Based on platform-wide quality metrics, the AI failure rate has shown consistent improvement — with ongoing improvements as the model is updated.

A "failure" is defined as any call where:

• The customer cannot place an order

• The customer cannot be transferred to a human agent

• The call drops due to AI behavior

This means the vast majority of calls are handled successfully, and the success rate continues to climb as the AI model is refined.

How does the AI improve over time?

The AI improves through several mechanisms:

1. Continuous model development — Maple's engineering team works on AI model improvements every day. New model versions go through QA testing and stress testing before deployment.

2. Real customer interactions — The system learns from actual calls and becomes more accurate and responsive to your business-specific inquiries over time.

3. Knowledge base enrichment — The more entries you add to your knowledge base, the better the AI can answer questions independently. Accounts with sparse knowledge bases (e.g., only 7 entries) will see more calls transferred to staff rather than handled by the AI.

4. Post-launch auditing — During the first week after going live, the support team audits incoming calls to identify misunderstandings or issues. The team will reach out via your preferred method (email or text) to discuss problematic calls and determine corrections through FAQ and knowledge base updates.

5. Ongoing cancellation and issue analysis — Maple reviews AI-related issues weekly to identify patterns and drive targeted improvements.

Think of the AI as a new employee — it may need some guidance early on, but once it's trained on your business operations and any initial bugs are ironed out, it performs excellently.

Why does the AI keep saying "let me check on that"?

This is a filler phrase the AI uses to avoid awkward silence while it processes your customer's request. Behind the scenes, the AI is thinking, looking up information, and querying your portal or backend system — and the filler phrase signals to the caller that a response is coming.

We understand this can feel repetitive. This is a known feedback item, and the engineering team is actively working on:

• Reducing the frequency of filler phrases

• Making fillers sound more natural and varied

• Optimizing processing speed to reduce the need for fillers altogether

Why does the AI sometimes pause or lag during conversations?

Noticeable pauses can occur when the AI is searching through large menus, syncing data, or looking up information such as business hours. The engineering team is actively working on reducing these delays.

In the meantime, the system uses conversational fillers like "one moment" or "hold on, let me check" to manage customer expectations during processing. If you're experiencing excessive pauses, contact support — the issue may be related to your menu size or configuration.

What should I expect during the first week after going live?

The first week is essentially a soft landing period. Here's what to expect:

• Call auditing — The support team will review incoming calls to catch any issues early.

• Minor adjustments — You may notice small inaccuracies or unexpected responses. These are normal and will be corrected through knowledge base and FAQ updates.

• Collaborative refinement — The team will reach out to discuss any problematic calls and work with you to determine the best corrections.

• Improving accuracy — Response quality improves rapidly as the system is fine-tuned to your specific business.

Does my setup affect AI performance?

Yes, significantly. Incomplete or incorrect configuration is one of the most common causes of poor AI performance. Key setup factors include:

• KYC (Know Your Customer) configuration — Incomplete KYC setup can result in missing functionality and degraded performance.

• Knowledge base entries — A well-populated knowledge base allows the AI to answer more questions independently instead of transferring calls.

• Greeting message length — Short, simple greetings lead to better customer engagement. Long or complex greetings — especially those that explicitly offer a transfer option — can cause up to 90% of callers to request a human agent immediately.

• Language settings — Verify that only your intended languages are enabled. Unexpected accents or language behavior may result from incorrect language opt-in settings.

Why do so many callers ask to transfer to a human?

If your greeting message explicitly offers callers the option to transfer to a human agent, data shows that approximately 90% of callers will request the transfer immediately — effectively bypassing the AI entirely.

Best practice: Keep your greeting as brief as possible and avoid offering the transfer option upfront. The AI can still transfer callers when needed, but not leading with that option dramatically improves AI engagement rates.

Can the AI handle specific types of inquiries?

The AI excels at:

• General inquiries — Business hours, location, menu questions, pricing, and service information

• Order taking — Guiding customers through placing orders

• Reservation and booking — Providing booking links and reservation information

• Party and event inquiries — Sharing rates and basic details (though confirming specific date/time availability may require a transfer to a specialist)

• Natural upselling — Suggesting add-ons when contextually appropriate during conversation

Current limitations include:

• Confirming real-time availability for specific dates/times without staff input

• Pronouncing website URLs naturally (the AI may spell them out letter-by-letter)

• Avoiding duplicate booking link sends when customers ask follow-up questions about the same topic

How does Maple handle AI quality issues when they're reported?

When you report a call quality problem, here's the process:

1. Call identification — The QA team identifies the specific calls with issues

2. Recording review — Call recordings are listened to and analyzed to determine what went wrong

3. Engineering escalation — When AI behavior anomalies are found, call recordings and issue summaries are sent to the engineering team for review

4. Resolution and testing — Fixes are developed, tested through QA, and deployed

If you experience an issue, providing specific call details (date, time, phone number) helps the team investigate more quickly.

Is the AI voice customizable? What if it sounds robotic?

Maple offers multiple voice options, and the platform is continuously improving voice quality. If the AI sounds choppy, robotic, or has unnatural pauses:

• Try a different voice — Alternative voice options are available and may sound more natural

• Check for recent updates — Voice quality updates are deployed regularly

• Report the issue — Support can switch your voice setting and apply the latest improvements

If you change your voice setting and it doesn't seem to persist, try logging out and logging back in to clear any caching issues.

How is AI performance measured and tracked?

Maple tracks several performance indicators:

• Failure rate — Percentage of calls where the AI cannot complete the customer's request

• Transfer rate — How often calls are escalated to human staff

• Reservation and order metrics — Volume of successful reservations and orders processed through the AI

• Call quality audits — Regular review of call recordings for accuracy and customer experience

These metrics are reviewed regularly to identify trends and drive improvements across the platform.