> ## Documentation Index > Fetch the complete documentation index at: https://docs.invoca.com/llms.txt > Use this file to discover all available pages before exploring further. # Signal AI Studio: FAQs > Answers to common questions about Signal AI Studio terminology, training rounds, labels, and how it compares to Signal AI Legacy. ## AI Signals overview ### Signal AI Studio terminology * **Signal AI Studio** = Invoca's no-code solution that lets you create your own AI Signals (called "AI (Studio)" in the platform) trained on your business's calls. * **Signal AI Legacy** = the original product launched in 2017 for creating AI Signals. Signals created with this product are called "AI (Legacy)" in the platform. * **Signal AI** = Invoca's brand name for all AI functionality in the platform, including Signal AI Studio, Signal AI Discovery, AI summaries, transcripts, redaction, Agent Voice ID, keyword spotting, and more. * **Signal AI Silver & Signal AI Gold** = the plans you can purchase to add certain AI products to your account. Signal AI Studio requires the Signal AI Gold Add-On. ### How is Signal AI Studio different from Signal AI Legacy? Signal AI Legacy has provided the ability to create custom AI Signals for predictive business outcomes since 2019. It's mature and supports training in the UI and via manual CSV uploads, but it runs on older machine learning technology, typically requires a large number of human-scored calls to reach acceptable accuracy, and only allows creating, training, and deploying one Signal at a time. Signal AI Studio is the newer self-serve, guided solution for creating custom AI Signals, offering quicker time-to-value because it's designed to build a robust AI model with 50-80% fewer labeled calls. The self-guided UI lets you train several Signals at once, with a clear view of predicted accuracy and training progress in real time. Signals can be activated (deployed) individually once they reach an acceptable level of accuracy and prediction consistency. Both products predict the outcome of a conversation based on examples of trues and falses, and both work best when example calls are labeled consistently, with a strong point of view on what constitutes a true or false call outcome. ## Signal AI Legacy **Can I still create and train Signal AI Legacy signals?** Yes — Legacy signals aren't being deprecated and can still be created (or existing ones maintained), but only to solve for certain signal types and scenarios. **When should I consider switching out a Signal AI Legacy for a Signal AI Studio signal?** If a Legacy signal is performing well, there's no need to switch it out — just know Legacy signals are scheduled to be deprecated in about 12 months. If you do want to switch a Legacy signal out, keep in mind this may affect your conversion rates depending on the accuracy of the new signal versus the prior one. ## Signal AI Studio **How do I start training signals in Signal AI Studio?** First, you need the Signal AI Gold package. If this is your first time training a Signal AI Studio signal, you'll be asked to complete a Signal Definition Form, which helps develop a quality objective definition and ensures the successful deployment of your new signal. Once submitted, Analytics Services reviews it and provides feedback to your CSM to confirm the Signal is a quality candidate for Signal AI Studio. See the Signal Definition Form overview video in the Invoca Academy to learn more. **What is the level of effort you should expect to complete training a signal using Signal AI Studio?** On average, expect a 6-10 hour commitment to define, train, and validate your signals. **What is a label?** Each AI Label represents a specific intent, conversion, or category you want to train the AI to identify (for example, "Scheduled Appointment" or "Appointment Lead"). A Label is part of an AI Model, with a max of 9 Labels. Labels are only created in Round 0 and can't be changed or removed in later training rounds. After successful training, a Label can be used to activate an AI Signal. Labels can be trained and activated separately as individual Signals, and don't count against Signal allocation limits until activated as a Signal. **What are good and bad examples of a label definition?** A good label definition should be simple yet detailed, ideally answering only one question for ease of training and to ensure a high-quality, accurate signal. For example, a label representing a lead intent signal might be "service inquiry." A good definition can be "a caller who expresses interest in a service by requesting information about the service, engaging in a detailed discussion about their needs and asking for a quote." A bad definition would be "when someone calls and asks a bunch of questions about our service." **How many labels can I train at once?** You can create and train up to 9 labels together at a time during initial label creation, in Round 0. Creating more than one label at a time speeds up the overall training process by letting you use the same call to train multiple labels at once. If you start with multiple labels, you won't be able to remove them at any point during training — you'll need to start over, or mark the label "Not Sure" to avoid affecting the AI Model. **Why should I train multiple labels together?** Training multiple labels together increases efficiency and saves time by letting you use the same call to train multiple labels at once — for example, appointment booked, appointment rescheduled, appointment canceled. Training multiple related signals simultaneously also results in a more complex, inter-related AI Model that more accurately reflects real-world scenarios. **What are examples of labels that I should train together (multi-labeling)?** Labels trained together don't have to be closely related, but training is easier if they're in the same "conversation neighborhood," for example: Appointment booked, Appointment rescheduled, Appointment canceled, New patient. Or: Sales inquiry, New customer, Sale complete, Promotion offered. **How does training work in Signal AI Studio?** Training splits into two phases: Labeling Calls (Round 0) and Verifying Labels (Round 1 and beyond). In the Labeling Calls phase, you use thematic search to enter terms that surface relevant calls for each label, determine which labels are True, False, or Not Sure for each call, and continue labeling until you reach at least 20 true and 20 false examples for one Label, and 10 true and 10 false for additional Labels, before progressing to the next round (clicking Train AI Model). In the Verifying Labels phase, calls are surfaced automatically based on your prior search criteria from the Labeling Calls phase, AI starts making predictions of True, False, or Not Sure, and for each Label, you confirm or update AI predictions based on your analysis of the transcript. After reaching the prescribed labeled/verified call count in each round, click **Train AI Model** to process labels. This starts an 8-12 hour process that applies your training inputs to improve the AI's accuracy round over round. Verification continues until a label reaches suggested predicted accuracy for activation (at least Round 4, 80% or better prediction accuracy, and a prediction range under 10%). **What does labeling calls mean?** Labeling calls means marking each as **True** (the call fits your label definition), **False** (it doesn't), or **Not Sure** (you're not clear whether it fits). **How is the initial Labeling Calls phase different from the Verify Labels phase?** In Labeling Calls, you manually search for relevant calls using thematic search and review each call to mark True, False, or Not Sure. In Verifying Labels, calls are surfaced automatically and the AI makes predictions for each Label — you only verify whether each prediction was correct. **What's a good example of a topic to search for?** If you want to train the AI to determine whether an appointment was booked, rescheduled, or canceled, you might use phrases such as "I'd like to schedule a tour," "Can I book a time for," "I want to book an appointment," "I'd like to reschedule my appointment," or "I need to cancel my appointment." It's a good idea to listen to 25-50 calls that fit your desired Signal definition scenarios to understand the types of phrases used. **What happens after each round of labeling?** Each completed round kicks off a training round to process your labeling and improve AI prediction accuracy — this may take up to 12 hours. If your Invoca notification settings are on, you'll be notified via app notification and email once training completes. **How do I know if the label is going to be good, or if I'm training it right?** Once you reach Round 3, predicted accuracy displays for each label, with help text guiding your progress and suggesting when a label is ready for activation. Use this criteria to decide whether to activate: minimum accuracy range should be at least 80% or higher, the accuracy range should have a difference of 10% or less between the two numbers (for example, 85-95%), and correcting AI predictions should occur less than 10% of the time during verification rounds. If accuracy dips between training rounds, the model may still be learning from new examples, or may have gotten confused by inconsistent labeling — continue with 1-2 additional rounds to see if accuracy improves and stabilizes. If it doesn't, the label's definition may be unclear or too broad — consult your CSM to discuss next steps. **Is it okay to see so many incorrect predictions during verification training?** Yes — it's normal to see incorrect predictions during verification, since the AI is still learning and refining round over round. Label consistently to avoid confusing the AI. **How is predicted accuracy calculated for each label?** Accuracy is calculated by evaluating how often the AI model correctly predicts calls based on its training. The accuracy score displays in the Studio UI, showing performance round over round. The accuracy metric incorporates an F1 score, providing a more comprehensive picture than accuracy alone by including both precision and recall. **How long does it take to complete each round of training? How many calls am I expected to review to train a label?** Each round requires about 1 hour of training. Most common signals (for example, lead, conversion) can be successfully trained with 250-350 labeled calls over 5-7 rounds on average (about 40 calls per round). The total number of rounds varies by use case. It's estimated to take 6-10 hours to define, train, deploy, and validate one or more signals over the course of 2 weeks. **Do labels count as signals?** No — you can create and train as many labels as you like without them counting against your network's Signal limits. Once ready, you can deploy a Signal that uses those labels. Labels are "free," so you can experiment without worry, activating them as signals once you're happy with the predicted accuracy. **What if I already activated a label as a signal but then did more training with it? Do I need to delete and remake the signal?** No. Additional training rounds may refine label accuracy and can be manually updated for a deployed signal using that label — navigate to the Manage Signals page, click the Signal, and select a different round number/accuracy. We recommend regularly maintaining activated Signals with additional training on an as-needed basis (for example, seasonal changes). **Who should I reach out to if I'm having issues with Studio AI?** Create a support case with details about the issue. The support team reviews the case and forwards it to the correct team based on the reported issue. For any other questions, reach out to your CSM or account team. ## Additional resources **Where can I find additional Signal AI Studio resources?** A great resource for learning more about Invoca's Signal AI technology is the Signals Learning Path in the Invoca Academy. ## Where to go next * [Signal AI Studio Overview](/s/article/signal-ai-studio-overview) * [Creating AI Signals in Signal AI Studio](/s/article/creating-ai-signals-signal-ai-studio) * [Training AI Signals in Signal AI Studio](/s/article/training-ai-signals-signal-ai-studio)