Right, let's talk about a headache. If you're helping a customer allocate Copilot Credits for M365 Copilot Cowork, you have to size it all up first — and doing that by hand is a fiddly, time-consuming job. Here's the problem, and a much easier way to solve it. 

Start with the Cowork Estimator 

Here's the good news: you don't have to start from scratch. Microsoft has published the Customer Cowork Estimator — a handy tool that turns personas, prompt complexity and expected usage into an estimated credit number. It's the perfect place to begin, and it does the core maths for you. 

To get the most out of it, you just need to feed it good inputs — and that's where a little groundwork comes in. For each part of the business, you'll want to: 

• Work out who's who. There are four types of user — corporate knowledge workers, management and senior leaders, customer-facing folks, and technical staff. 

• Count how many of each. Quick for a small team, a bit more involved for a big one. 

• Pin down what they'd genuinely use Cowork for. The real multi-step workflows that hop across apps and actually do things — not just a quick chat or a summary. 

• Judge how heavy each workflow is. Light, Medium or Heavy — since each level uses a different number of credits per run. 

• Estimate how often it runs. Daily adds up to a lot of runs a month; weekly is far fewer. 

The estimator handles the sums beautifully once those inputs are in. Gathering the inputs themselves — persona by persona, workflow by workflow, across a few thousand people — is simply the part that takes time. And that's exactly where this agent lends a hand: it builds on the estimator by automating the groundwork that feeds it. 

How this agent helps 

 This agent takes on that groundwork for you. Instead of sizing everyone from the outside, the admin simply switches the agent on for every user. Each person then sizes their own needs — and the agent does the clever bits for them, ready to drop into the estimator. Here's what it does behind the scenes: 

• Spots the right persona. It works out which of the four personas each user fits. 

• Finds the real workflows. It looks at the top Cowork scenarios the user would actually run. 

• Keeps it honest. It checks each one is a genuine Cowork job — several steps, more than one app, real actions and a bit of decision-making — not something a Scheduled Prompt or plain Copilot Chat could do just as well. 

• Grounds it in evidence. It reads the signals from recent work — emails, meetings, documents and Teams chats — so the estimate is based on what people actually do, not thin air. 

• Does the sums. It maps each workflow's complexity to credits per run and totals it up — giving you numbers that line up neatly with the estimator. 

• Shows the value. It gives a view of pay-as-you-go versus a pre-purchase plan, and a sense of the return on the spend. 

• Stays transparent. It states its assumptions, flags how confident it is, and sticks to permitted data and the usual privacy and compliance rules. 

And here's the kind of report it hands back — persona, the top workflows, complexity, credits, cost and a clear recommendation, all in one place: 

Why this approach works better 

So why hand it to the users? A few good reasons: 

• More accurate. The numbers come from each person's real workload, not a top-down guess. 

• Consistent by design. Everyone follows the same method — same personas, same complexity bands, same maths — so the results line up and roll into one clean figure. 

• No over-buying. Because it weeds out the workflows that don't really need Cowork, your customer only pays for credits they'll genuinely use. 

• A proper business case. You get cost clarity and a feel for the return, so you're handing over more than just a number. 

• It scales. Ten users or ten thousand — the effort on your side stays much the same. 

• It saves you hours. You swap manual sizing for gathering and rolling up. Your time goes on advising, not tallying. 

Not ready to roll it out? Other ways in 

If an admin isn't quite ready to switch the agent on for everyone, that's OK— there's no need to. This agent is an M365 Copilot agent, and it has free access to Work IQ — the same engine that will eventually power Cowork. Because that access is free, the agent can read how people really work and take the guesswork right out of sizing. And there's more than one way to get at it: 

1. Build it with Agent Builder. Ready-made instructions, a description and a starter prompt are all sitting in the GitHub repo. Anyone with access to M365 Copilot can pop them into Agent Builder and stand the agent up in minutes. 

1. Prefer not to build an agent? Just use the prompt. There's a comprehensive prompt you can drop straight into the M365 Copilot Chat experience. Same sizing, same evidence — no agent to create. 

Ready to give it a go? 

Whichever route suits you best, getting started is dead simple. If you're rolling the agent out to your users: 

1. Have a quick word with your customer's admin about switching the agent on for their users. 

1. Let each person size their own Cowork needs. 

1. Gather it all up and roll it into a single estimate. 

Prefer to keep it in your own hands? Build the agent from the GitHub repo with Agent Builder, or drop the comprehensive prompt into M365 Copilot Chat — you'll get to the same place with even less setup. 

THE BOTTOM LINE 

That's the heavy lifting done for you. You'll save yourself hours, spare yourself the guesswork, and hand your customer a credit plan they can genuinely trust — with the numbers and the business case sitting right behind it. Give it a spin on your next engagement and see how much quicker it gets you there. 

Resources 

Everything you need is in one place: 

• Customer Cowork Estimator — https://aka.ms/CustomerCoworkEstimator 

• GitHub repo — agent instructions, description, starter prompt and the full Copilot Chat prompt: Cowork Investment Assessment 

Posted by Ben Weiss, Senior Developer Relations Engineer, Android Developer Relations

Welcome back to the blog post series "Build intelligent Android apps" where we take a basic Android app and transform it into a personalized, intelligent, and agentic experience. In our previous post, we explored how to leverage Firebase AI Logic to build cloud-hosted and hybrid AI features.

In this article, we'll show you how we designed and integrated these capabilities into our travel planning app, JetPacker, using Android AppFunctions. We'll explore the rationale behind our feature choices, discuss the specialized tooling we used to accelerate development, and dive into the code that makes it all work.

Designing AI-ready features: making choices that matter for your users

To select which features to provide to the intelligence system, we looked for tasks where a voice or text command is objectively faster than tapping through screens. In this side-by-side screen recording you can see this contrast perfectly: on the left, a user tapping through multiple screens to log an expense; on the right, the same task completed instantly in the background via a privileged agent.

Our first choice was expense tracking. Logging a coffee expense during a trip usually takes quite a few taps—unlocking the phone, opening the app, finding the active trip, navigating to the expenses tab, tapping the add button, taking a picture of the receipt, and checking the result. By providing the addExpense and getExpenses features as AppFunctions, the system agent handles the heavy lifting. When the user says, "Add a five-dollar coffee expense to my Paris trip," the agent automatically searches for the correct trip ID in the background and inserts the expense, skipping the manual UI flow entirely.

We also prioritized itinerary management. Finding what activity is next on a busy trip itinerary usually requires scrolling through a dense timeline view. By providing getItinerary and addItineraryEvent to the system, the user can simply ask, "What am I doing next in Paris?" and get an immediate answer.

Finally, we focused on hands-free note capturing. Typing out reminders or notes while walking down a busy street is difficult and unsafe. Exposing a voice note capability allows the user to say, "The flight was amazing, I saw a beautiful sunset and managed to sleep well," and the privileged agent automatically transcribes and saves it directly into the travel database  using the addVoiceNote AppFunction.

Android MCP powered by AppFunctions

This entire experience is built on Android MCP. Under this design, the app acts as a local MCP server. Rather than remote APIs, you provide your app features directly to the on-device intelligence system.

Android AppFunctions is the API that brings this concept to life. It reads annotated Kotlin functions and compiles them into type-safe, sandboxed tool definitions that the privileged agent can discover and invoke locally on the device.

Diagram highlighting our apps, the android platform, and system agents coordinate AppFunctions.

Under the Android MCP model, your app acts as a local MCP server that exposes structured tools, while the Android platform serves as the central tool registry. On the MCP client side, agent apps are registered with the intelligence system after being granted system-privileged permissions to access the registry.

When a user interacts with a registered agent, its LLM determines if the request can be handled by an AppFunction, queries the platform's metadata, and executes the appropriate registered functions in the background. This local MCP client-server design gives you full control: you choose exactly which features are accessible to the agent, keeping the rest of your app's data private.

How we accelerated development with Android skills

To streamline the integration process, we leveraged the AppFunctions development skill. The AppFunctions development skill is a complete development companion. It guided us through the entire lifecycle: mapping Kotlin data classes to serialize parameters, generating the necessary Service entry points, refining our KDoc documentation to ensure the LLM understands parameter boundaries, and setting up automated testing using ADB.

Providing app features to the intelligence system

Enough with the theory, let's dive into the implementation.

Configuration and dependency setup

We begin by adding the AppFunctions dependencies. One for the API and one for the Kotlin Symbol Processing compiler.

implementation("androidx.appfunctions:appfunctions:1.0.0-alpha10")
ksp("androidx.appfunctions:appfunctions-compiler:1.0.0-alpha10")

Modeling custom data types

Any custom object exchanged with the agent must be annotated with @AppFunctionSerializable. In our TripSerializable.kt file, we define our trip data model:

@AppFunctionSerializable(isDescribedByKDoc = true)
data class TripSerializable(
    /** The trip's unique identifier. */
    val id: String,
    /** The trip's title. */
    val title: String,
    /** The trip's destination location. */
    val location: String,
    /** The trip's start date in milliseconds. */
    val startDate: Long,
    /** The trip's end date in milliseconds. */
    val endDate: Long,
    /** A list of participants. */
    val participants: List<String>,
)

Providing features using the @AppFunction annotation

Next, the skill wrote the Kotlin functions that perform the database queries and annotate them with @AppFunction. We can view this in searchTrip:

/**
 * Looks for trips based on optional filters like id, title (name), location, and dates.
 *
 * @param id The unique identifier of the trip.
 * @param title The title or name of the trip.
 * @param location The destination location.
 * @param startDate The minimum start date in milliseconds.
 * @param endDate The maximum end date in milliseconds.
 * @return A list of trips matching the filters.
 */
@AppFunction(isDescribedByKDoc = true)
suspend fun searchTrip(
    id: String? = null,
    title: String? = null,
    location: String? = null,
    startDate: Long? = null,
    endDate: Long? = null
): List<TripSerializable> {
    return withContext(Dispatchers.IO) {
    // implementation
}

Since AppFunctions run on the UI thread by default, we use withContext(Dispatchers.IO) to switch to a background dispatcher. Additionally, we refine our KDoc to use clear, imperative verbs and specify parameter constraints. This documentation compiles directly into the tool's schema, which the privileged agent uses to resolve parameters and handle runtime errors.

The service entry point and Hilt integration

To register these features with the intelligence system, we create an abstract base class that extends AppFunctionService. We annotate it with @AppFunctionServiceEntryPoint:

@RequiresApi(36)
@AndroidEntryPoint
@AppFunctionServiceEntryPoint(
    serviceName = "JetPackerAppFunctionService",
    appFunctionXmlFileName = "jetpacker_app_function_service"
)
abstract class BaseJetPackerAppFunctionService : AppFunctionService() {
    @Inject internal lateinit var tripDao: TripDao
    // DAOs and database references are injected here...
}

During compilation, KSP generates the final concrete service subclass, JetPackerAppFunctionService, as declared with the serviceName parameter. We also register app_metadata.xml in the app's manifest. This file provides global operational rules for JetPacker's declared AppFunctions.

Testing and verifying your AppFunctions

Once implemented, you should verify that your AppFunctions are registered and working correctly.

Running devices or emulators with Android 17 or newer, you can use ADB commands from your terminal to list and invoke your functions. Running adb shell cmd app_function list-app-functions displays all registered functions for your package. You can then execute a specific function and test its database integration by running adb shell cmd app_function execute-app-function while passing a raw JSON parameters string.

Instead of these ADB commands, you can also use the AppFunctions Testing Agent to inspect your configuration, list and execute AppFunctions, and even see how your AppFunctions behave in a real conversational flow.

Wrapping it up

When thinking about app features that can be contributed to the intelligence system using AppFunctions requires a slight shift in how we think about code and documentation. AppFunctions enable you to use this new interaction model for apps, which allows using an agent to access app features..

First, the AppFunctions development skill is an essential lifecycle tool, helping you discover features, implement and refine AppFunctions for your apps. Second, KDoc comments are a compiled API asset; clear parameter descriptions directly impact the execution accuracy of the system agent. Finally, Android MCP provides local-first execution allowing apps to safely collaborate with AI agents.

Contributing app features through AppFunctions makes your application ready for the intelligence system. Let us know how you are adapting your apps for the agentic era!

Learn more

Check out the other parts of this blog post series:
Part 1: Introduction of the app and a high-level overview.
Part 2: On-device intelligence. Deep-dive into ML Kit’s GenAI APIs and Gemini Nano to build privacy-first features like itinerary summarization, receipt parsing, and local audio processing.
Part 3: Hybrid and cloud reasoning. Explore how to use Firebase AI Logic to ground LLM answers in real-world data like Google Maps and web context.
Part 4 (this post!): System integration. Integrating with the Android intelligence system using AppFunctions.
Part 5 (coming soon): In-app agentic workflows. Extend the app with an end-to-end booking assistant powered by A2UI and ADK.

Interested in more on Android Development? Follow Android Developers on YouTube or LinkedIn!

All code snippets in this blog post follow the following copyright notice:

Copyright 2026 Google LLC.
SPDX-License-Identifier: Apache-2.0

Welcome back to the blog post series "Build intelligent Android apps" where we take a basic Android app and transform it into a personalized, intelligent, and agentic experience. In our previous post we explored how to build intelligent on-device features using Gemini Nano through ML Kit's Prompt API.

In this post, we will look at how you can leverage Firebase AI Logic to build cloud-hosted and hybrid AI features: 

• Grounding answers in real-world context
• Routing requests dynamically between cloud and local execution using hybrid inference
• Translating content with custom routing systems

Sometimes a use case requires AI models with greater world knowledge, a much larger context window, or the ability to handle complex queries. In those scenarios, we can leverage cloud models. 

Other times, you want the best of both worlds: using hybrid inference to run on-device when available to lower costs, while falling back to the cloud to ensure compatibility for all devices.

Cloud and hybrid features in Jetpacker: Museum assistant with web grounding, hybrid restaurant review drafting, and  support chat featuring custom-routed live translation.

Let’s look at how we implemented three cloud and hybrid features in Jetpacker:

• a museum assistant with web grounding
• hybrid restaurant review drafting
• hotel support chat featuring custom-routed live translation.

Use LLM grounding for up-to-date informationMuseum assistant chatbot with LLM grounding

The Museum assistant is an interactive chatbot designed to help users plan their museum visits. It provides visitors with up-to-date details regarding specific exhibits, current opening hours, ticket pricing, and more.

Museum assistant is a chatbot that answers questions, such as 

When building AI features, getting the model to answer with fresh, accurate, and specific real-world information is a common challenge. While cloud models possess massive amounts of world knowledge, they might not know about seasonal exhibits or the current day’s opening hours. 

Grounding data is added to the context window to enable the model

To bridge this gap, we can use grounding techniques to add extra context to the model’s context window. The Firebase AI Logic SDK supports three types of grounding:

• URL grounding: Grounding responses using content from a specific webpage (e.g. current ticket prices or museum rules).
• Google Search grounding: Letting the model query the real-time Google search index for up-to-date details.
• Maps grounding: Using Google Maps location data.

In Jetpacker, we dynamically construct the available tools based on enabled feature flags and initialize the generative model using the Firebase AI SDK:

// implementation("com.google.firebase:firebase-ai-logic")

private var toolList = mutableListOf<Tool>()

init {
    if (ENABLE_SEARCH_GROUNDING) {
        toolList.add(Tool.googleSearch())
    }
    if (ENABLE_URL_GROUNDING) {
        toolList.add(Tool.urlContext())
    }
}

private val generativeModel = Firebase.ai(backend = GenerativeBackend.googleAI())
    .generativeModel(
        modelName = "gemini-3-flash",
        systemInstruction = content {
            text("You are a helpful museum assistant answering questions about a museum. Use plain text.")
        },
        tools = toolList
    )

When the user queries the assistant, if URL grounding is enabled, we append the specific museum resource URLs directly into the prompt:

val groundingText = if (FeatureFlags.ENABLE_URL_GROUNDING) {
    "\n If the following message above is about the rules and terms to visit Le Louvre, " +
    "if needed answer this urls ${urlList.joinToString()}"
} else {
    ""
}

val prompt = "$text $groundingText"

var response = chat.sendMessage(prompt)

Hybrid inference: On-device review generation with Maps deep link

Not every AI task requires a cloud-based model, and not every device is online. To help developers balance latency, cost, and offline availability, we recently introduced the Firebase API for Hybrid Inference.

In Jetpacker, the restaurant review feature lets users review select topics and automatically drafts a review. To enable this for all users, we prioritize local execution with Gemini Nano, and fall back to cloud models on devices that don’t support Gemini Nano. 

// implementation("com.google.firebase:firebase-ai-logic")
// implementation("com.google.firebase:firebase-ai-ondevice:16.0.0-beta03")


// Initialize the model with hybrid routing configuration
val reviewModel = Firebase.ai.generativeModel(
    modelName = "gemini-3.1-flash-lite",
    onDeviceConfig = OnDeviceConfig(
        inferenceMode = InferenceMode.PREFER_ON_DEVICE
    )
)

The Hybrid Inference API supports four distinct routing modes:

• PREFER_ON_DEVICE: Prioritizes local execution and falls back to cloud if Gemini Nano is unavailable.
• PREFER_IN_CLOUD: Prioritizes cloud execution and falls back to on-device if the device goes offline.
• ONLY_ON_DEVICE: Restricts execution strictly to the device.
• ONLY_IN_CLOUD: Restricts execution strictly to the cloud.

Once the review is generated, we copy it to the clipboard and use an intent to open Google Maps directly to the restaurant's review page, providing a seamless user experience:

private fun copyAndOpenMapsReview(context: Context, reviewText: String, placeId: String) {
    val clipboard = context.getSystemService(Context.CLIPBOARD_SERVICE) as ClipboardManager
    val clip = ClipData.newPlainText("User Review", reviewText)
    clipboard.setPrimaryClip(clip)

    val uri = Uri.parse("https://search.google.com/local/writereview/mobile?placeid=$placeId")
    val intent = Intent(Intent.ACTION_VIEW, uri).apply {
        setPackage("com.google.android.apps.maps")
    }
    context.startActivity(intent)
}

Custom hybrid routing: Hotel support chat translation with simulated personas

The hotel support chat was built to let users finalize logistics and check on hotel details. This feature uses system instructions to configure a localized receptionist assistant. By passing specific information—such as the preferred language and hotel information—in the instructions, we can set up a conversational persona representing a specific hotel.

private val generativeModel = Firebase.ai(backend = GenerativeBackend.googleAI())
    .generativeModel(
        systemInstruction = content {
            text("""
              You are a helpful hotel receptionist at $hotelName only speaking $language. 
              Answer politely in $language. The bar closes at 10pm and breakfast is from 7am to 10am.
              There's someone at the desk 24/7. You can retrieve your luggage from the storage room 
              at the back of the lobby at any time.
              """)
        },
        modelName = "gemini-3-flash-preview"
    )

Because receptionist responses are in the hotel's local language (for example, French for Hotel Le Meurice in Paris), we need to translate messages to the user’s preferred language. 

Hotel support chat messages are automatically translated to the user’s preferred language 

While hybrid models can configure simple routing preferences, complex scenarios require custom routing logic. In Jetpacker, we implement a custom routing stack that takes into account:

• Language identification: Using the on-device ML Kit Language Identification API, we can detect the incoming message language.
• On-device translation (Gemini Nano): ML Kit’s Prompt API lets us translate common language pairs directly on the device, saving bandwidth and cloud cost.
• Cloud translation (Gemini 3 Flash): For more complex languages, we use Gemini Flash 3 to get a higher quality translation.

// implementation("com.google.android.gms:play-services-mlkit-language-id:17.0.0") 

// ML Kit for Language Identification (powered by Google Play Services)
private val languageIdentifier = LanguageIdentification.getClient()

// On-device translator model (prefer Gemini Nano) for translating common language pairs
private val hybridTranslationModel = Firebase.ai(backend = GenerativeBackend.googleAI())
    .generativeModel(
        modelName = "gemini-3-flash",
        onDeviceConfig = OnDeviceConfig(mode = InferenceMode.PREFER_ON_DEVICE)
    )

// Cloud translator model for more complex language pairs
private val cloudTranslationModel = Firebase.ai(backend = GenerativeBackend.googleAI())
    .generativeModel(
        modelName = "gemini-3-flash"
    )

When a message needs to be translated, we identify the source language and apply our custom routing logic, executing either on-device or cloud translation:

fun translateMessage(message: SupportChatMessage) {
    viewModelScope.launch {
        // 1. Detect language using ML Kit Language Identification
        val sourceLang = try {
            Tasks.await(languageIdentifier.identifyLanguage(message.text))
        } catch (e: Exception) {
            "Undefined"
        }

        // 2. Custom routing: we've verified the translation quality for English and Korean with Gemini Nano, and will translate message on-device for those two languages
        val routeToCloud = sourceLang != "en" && sourceLang != "kr"

        val prompt = "Translate the following text to $selectedLanguage. Just return the translated sentence: ${message.text}."

        val (translatedText, routePrefix) = if (routeToCloud) {
            val result = cloudTranslationModel.generateContent(prompt)
            result.text to "[Cloud]"
        } else {
            val result = hybridTranslationModel.generateContent(prompt)
            result.text to "[On-Device]"
        }

        if (translatedText != null) {
            _translations.update { current ->
                current + (message.id to "$routePrefix: $translatedText")
            }
        }
    }
}

In this example, the custom routing logic only takes into consideration the translation’s source and target language. However, based on your app’s use case, you can expand the routing logic to include other factors such as the on-device model version, network connectivity, battery status, and more.

Securing the AI Pipelines: Firebase App Check

Lastly, using AI in the cloud opens up possibilities of API key abuse or unauthorized billing. To secure API calls, we integrated Firebase App Check using both Play Integrity (production) and the local Debug Provider (for local development or emulators).

In the JetPackerApplication.kt file, we install the debug provider at startup and trigger anonymous authentication to establish a secure user session:

//  implementation("com.google.firebase:firebase-appcheck-playintegrity") 
//  implementation("com.google.firebase:firebase-appcheck-debug")  
//  implementation("com.google.firebase:firebase-auth") 

override fun onCreate() {
    super.onCreate()
    Firebase.initialize(context = this)
    Firebase.appCheck.installAppCheckProviderFactory(
        DebugAppCheckProviderFactory.getInstance()
    )
    Firebase.auth.signInAnonymously()
}

When building locally on an emulator, App Check prints a local token secret to logcat:

Enter this debug secret into the allow list in the Firebase Console: a8c2dd4c-xxxx-xxxx-xxxx-ef6c114ba27e

Once registered in the Firebase console, local requests are fully verified and authenticated by App Check, protecting our backend while letting us test the app locally.

Conclusion

By combining cloud model capabilities (grounding, system instructions) with on-device capabilities (hybrid routing, translation, security app checks), we created a travel app that is smart, secure, and available offline.

Check out the full source code for Jetpacker on GitHub, and explore the Firebase documentation to get started:

Firebase AI Logic Documentation
Firebase Hybrid Inference API

Learn more

Check out the other parts of this blog post series:

Part 1: Introduction of the app and a high-level overview.
Part 2: On-device intelligence. Deep-dive into ML Kit’s GenAI APIs and Gemini Nano to build privacy-first features like itinerary summarization, receipt parsing, and local audio processing.
Part 3 (this post!): Hybrid and cloud reasoning. Explore how to use Firebase AI Logic to ground LLM answers in real-world data like Google Maps and web context.
Part 4: System integration. Integrating with the Android intelligence system using AppFunctions. 
Part 5 (coming soon): In-app agentic workflows. Extend the app with an end-to-end booking assistant powered by A2UI and ADK.

Interested in more on Android Development? Follow Android Developers on YouTube or LinkedIn!

All code snippets in this blog post follow the following copyright notice:

Copyright 2026 Google LLC.
SPDX-License-Identifier: Apache-2.0