An assistant in the application manager

A little under two weeks ago, in Transpareo in the agentic age, we described how an assistant works with Transpareo from outside: you connect your own assistant, and it reads requirements, checks passports and publishes them.

The same assistant now lives directly in the application manager as well. There is nothing to install and nothing to connect. An administrator stores an AI provider’s key once, and from then on anyone they allow opens the assistant from the button in the top bar.

What it does for you

The assistant appears as a small window over the page, which you can move around and pull larger. It stays open while you carry on working in the application manager.

It knows where you are. Open it on a product and it relates your questions to that product, with suggestions to start from that fit the page. Ask on a draft “What is still missing before I can publish this passport?” and it asks the platform what a passport for this product requires, has the draft validated and names the missing details. Pages it mentions come as links, so you are there in one click.

It reads what you hand it. Drag a supplier’s spreadsheet into the window as an XLSX, CSV, ODS or JSON file. It lands in your workspace and is deleted after 24 hours. The assistant reads it section by section, proposes which column belongs to which property and passes it to the importer. It reads a data sheet as a PDF or a photo of a rating plate too; for that, your browser sends those files to the provider along with the request.

It knows the help. How the importer maps columns, or what a granularity is, it looks up in the help articles and answers with what they say.

Past conversations are in the window’s history, searchable and holding the last 30.

It acts as you

For every person who uses the assistant, Transpareo creates an access of its own, which appears in the list of API accesses as Assistant of followed by the person’s name, and which can be switched off there at any time.

That access is limited to the person’s rights on every single call. Someone who may not publish passports gets an assistant that publishes none. When an administrator withdraws a right, the assistant lacks it on the next call.

The assistant can never do more than the person using it.

Whatever the assistant changes goes through the same interface as every other change and lands in the same audit log. When the platform refuses something, it receives the same error message a person sees at the command line: which field is missing and what to do. With that it usually corrects itself.

Sensitive steps wait for you

Before any step that cannot be quietly undone, the assistant stops and shows you a card: which operation, with which details, exactly as it will be sent. You approve or deny, with a reason if you like, and the assistant reads that reason and acts on it.

Whatever cannot be quietly undone happens only once a person has approved it.

By default this covers publishing and withdrawing passports and products, creating whole series at once, and anything that cannot be reversed. Where an operation demands an explicit confirmation, such as voiding a passport, the window fills it in after you approve; the assistant never types it itself.

If you want to approve every change yourself, switch it in the window: the clock icon in the window’s header opens past conversations, and below them sits the setting Approvals. The default is Risky writes only; with Every write the assistant puts every change in front of you. The choice applies to the browser you make it in. There is no setting without approvals at all.

By default the assistant does not ask you to approve an import, because an import carries three safeguards of its own:

• If it is unclear where a column belongs, the assistant asks you before anything is written.
• Before writing, a dry run checks every row. If a single row fails, none is written.
• A finished import can be taken back as a whole.

With Every write, an import gets a card too. It names the columns that will become new property types.

An assistant also reads text from suppliers, and that text can carry an instruction someone wants to slip past it. So the card shows the actual details that will be sent, and the assistant’s answers appear as plain text in which nothing is executed. This is the part that took the most of our attention while building it, and the one we consider the most important.

Your provider, your contract

Which AI provider answers is your decision. The list holds Anthropic, which uses its own protocol, and a dozen or so providers offering the OpenAI chat interface, among them Mistral, IONOS and OVHcloud. The European ones come first, and each shows the country where the computing runs. You can also enter an address of your own, and with some setup that includes a model running in your own building.

The key is stored encrypted and used from your browser. The request goes from your machine straight to the provider; for the assistant, our servers talk to no AI provider at all. What the provider gets to see is what the assistant reads to answer your question: product data, help text, attached files. What happens to it there is governed by your own contract with them.

Running the assistant entirely in the browser is a decision for your confidentiality. What you write and what the model answers can be read by you and your AI provider alone. The text of your conversations never reaches our servers, so we can neither store it, analyse it nor hand it to anyone. What does reach us are the calls through which the assistant reads or changes data in your workspace, as with any other access, and attached spreadsheets for the importer.

Only you and your AI provider can read your conversations with the assistant.

This has one consequence we want to name openly: past conversations are kept only in the browser they were held in, and another device starts with an empty history. A history across all devices would have meant storing your conversations with us.

What it costs

Transpareo charges nothing extra for the assistant; it is included in every plan. You pay the provider for the computing, billed per token.

To keep the cost in view, the window lists for every answer how many tokens it used: read, written and taken from the cache. A click on the small circle at the end of the answer shows these figures, together with the tools the assistant called for it.

For that you need an account on the provider’s developer platform, with billing set up there. A chat subscription, the kind many people use day to day, is a separate product: it includes no API access, and its login does not work here. The provider form says so right where you enter the key.

How to set it up

1. Add a provider. Under AI providers, choose the provider, enter the key and pick the model from the list the provider reports for your key.
2. Grant the permission. Administrators always have the assistant. Everyone else receives the permission Use the assistant panel through their role.
3. Ask. The button in the top bar opens the window.

If you would rather work with an assistant outside the application manager, or automate workflows with scripts, the connection and the command line from the article on the agentic age remain available. Both routes reach the same tools.

Most of the work on a product passport lies in putting the product data in order, as we described in The DPP forces digitisation. The assistant now helps with exactly that work, on the page where it happens.