Understanding and reasoning
Interpreting flexible requests, generating language, or proposing tool use when connected to the application's services.
The question we welcome
Why not just put Claude or ChatGPT in the application?
Sometimes, that's exactly the right choice. AI can interpret requests, generate responses, and invoke tools when appropriately integrated. Andamento serves a different need: developer-controlled execution of structured voice interactions.
Intelligence versus conversational execution
AI can decide what an answer means. An app still needs an experience that works.
An AI model may interpret “Actually, skip clip 11.” The surrounding application must decide how to capture that change, confirm the intended action, and respond if the user interrupts. Andamento provides developer-controlled machinery for that structured conversational interaction.
Running the conversational process
Coordinating prompts, listening, flow steps, interruption behavior, recovery, and callbacks that let the application respond.
This is not an exclusive division: some integrated voice-agent platforms supply overlapping functionality, and an application can assemble its own runtime. AI is optional in an Andamento-based flow.
Responsibilities, not rival personalities
Three roles in one possible application architecture.
You can use these together, or build a voice application with Andamento and no generative AI at all.
Optional · intelligence
AI model
Interprets free-form language, reasons about an open-ended request, generates text, or proposes a tool action.
Example: interpreting “Actually, skip clip 11.”Conversational runtime
SpeechTiles Andamento
Coordinates prompts, listening, dialog flow, recognition responses, interruption behavior, confirmations, and application callbacks.
Example: running the configured confirmation step before proceeding.Domain-specific product
Your application
Owns the project or patient records, business rules, service integrations, permissions, user interface, and actual work.
Example: applying a grade or saving an inspection result.Important: This is a conceptual separation of responsibilities, not a claim that every application must route data through these three components in this order.
An honest comparison
What is different—and what already overlaps?
| Need | AI model / agent stack | Andamento |
|---|---|---|
| Open-ended reasoning | Core strength of an LLM | Not its purpose |
| Understanding user intent | Can interpret flexible language | Can work with recognized responses and application-provided interpretation |
| Structured multi-step flow | Possible with orchestration code or agent frameworks | Core focus: developer-defined conversation execution |
| Speech and turn coordination | Available in some voice agent platforms; otherwise custom work | Framework capabilities for prompts, listening, interruptions, and recovery |
| Business action | Requires application tools and authorization | Requires host-app integration and authorization |
| Safety and compliance | Require system-level design and testing | Require system-level design and testing |
Modern voice agent platforms may already provide turn-taking, tool execution, and interruption support. Andamento should be evaluated against those real alternatives, not against a language model in isolation.
The difference in practice
Consider a single instruction versus a guided workflow.
One action
“Show tomorrow's appointment.”
An AI integration with an appropriate calendar tool may be all the application needs. There may be little reason to add Andamento.
A structured interaction
“Let's complete your daily check-in.”
The app must walk through a controlled sequence, capture answers, handle unclear input, confirm critical details, and hand data to the host application. That's a better candidate for Andamento.
The healthcare example describes a conceptual interface, not clinical advice or a ready-made healthcare product.
A real buying question
Who handles what in a guided voice workflow?
Imagine a technician must complete a required inspection while keeping both hands occupied. An AI model could help interpret an unusual note; it doesn't replace the need to coordinate the inspection dialog and connect it to the company's system.
The application defines the task
Which checks are required, which answers are acceptable, and where results must be stored.
Andamento runs the interaction
Ask the next question, listen, apply configured interruption and retry behavior, confirm as defined, and notify the app.
AI contributes only if needed
Optionally interpret an open-ended note or assist with language, under the application's chosen integration.
Why buy it? To avoid engineering all of step 02 yourself when the available SDK behaviors meet the requirements of your application—not because an AI model is incapable of speech or conversation.
A practical distinction
Andamento's conversation engine runs on the device.
The local runtime executes developer-configured conversational behavior. An AI model is optional; the application can connect one when interpreting or generating language is valuable.
Local runtime
Andamento
Conversation sequencing, prompt and listening coordination, and configured response handling happen inside the Apple application.
Service selection
Speech recognition
Can use on-device or cloud recognition depending on the application's configuration and available platform services.
App choice
AI and business services
Reasoning, external tools, and remote data integrations are chosen and governed by the application developer.
A separate licensing-permissions mechanism may involve network communication; “on-device” does not mean that every possible configuration is fully offline.
Questions we're not avoiding
Frequently asked questions
Can Claude or ChatGPT manage an entire conversation without Andamento?
Yes. A model and the right surrounding services can produce an end-to-end conversational experience. The comparison is between buying an existing runtime and building or assembling a suitable one—not between “AI can” and “AI cannot.”
Does Andamento include a generative AI model?
No. Andamento is a conversational execution framework, not a general-purpose generative model. An application may add AI where useful.
Will Andamento replace my application?
No. The host application still owns its workflows, data, visual UI, business logic, external systems, and final actions.
Does it guarantee an AI will never make a mistake?
No. Application developers must design validation, authorization, user confirmations, logging, and appropriate safeguards for their domain. No SDK alone guarantees correctness or regulatory compliance.
Is this useful without AI?
Yes. Many structured voice dialogs can use speech recognition, prompts, and developer-defined flows without needing generative reasoning.
Does Andamento need to call a SpeechTiles cloud service for each turn?
No. Its conversational execution runs on-device. Depending on the application's speech configuration and AI choices, other services may make network requests, and licensing permissions are a separate consideration. See Performance & Privacy.
NEXT STEPS
Explore the SDK. Test your own workflow.
Review the public documentation, register for developer access, or tell us about the conversation your application needs.