Skip to Content

Managed AI

Intelligence you can meter.

Chat, embeddings, knowledge retrieval, OCR, transcription and image generation — six services on one pre-paid balance. No API key to obtain, no GPU to rent, no vector database to run.

Your first $5 is free and needs no card. It is already waiting in the Assistant on your desktop.

One balance

Six services, one wallet.

Every service draws from the same pre-paid credit. Nothing is billed on a separate invoice, and nothing renews without the balance to cover it.

Chat

The Assistant, and any vApp that calls the AI. Billed per token, in and out, at the rate of the model you chose.

Embeddings

Turning your documents into something searchable by meaning. Billed per million tokens indexed.

Knowledge retrieval

The RAG system below: search, reranking and citation over your own corpus.

OCR

Scans and image-only PDFs into text. Billed per page — or run free in the browser instead.

Transcription

Audio and video into text. Billed per minute — or run free in the browser instead.

Image generation

Images from a prompt, inline in the Assistant and available to vApps. Billed per image.

The arithmetic

What does the credit actually buy?

Real rates from the live price book. Pick a balance.

Choose a credit balance to convert
42 chat turns on Sonnet 4.6
250 chat turns on Haiku 4.5
100 pages read by OCR
83 minutes transcribed
1,900 pages added to your knowledge base
16 generated images
833 chat turns on Sonnet 4.6
5,000 chat turns on Haiku 4.5
2,000 pages read by OCR
28 hours transcribed
38,000 pages added to your knowledge base
333 generated images
5,000 chat turns on Sonnet 4.6
30,000 chat turns on Haiku 4.5
12,000 pages read by OCR
167 hours transcribed
230,000 pages added to your knowledge base
2,000 generated images
27,000 chat turns on Sonnet 4.6
162,000 chat turns on Haiku 4.5
65,000 pages read by OCR
900 hours transcribed
1,250,000 pages added to your knowledge base
10,800 generated images

A chat turn is counted as 2,000 tokens in and 400 out. Every other figure is the published per-page, per-minute, per-image or per-million-token rate divided into the balance. Spend the balance on any mix of the six — it is one wallet.

The models

You pick the model. We publish the rate.

Five providers behind one endpoint — Anthropic, OpenAI, Google Gemini, xAI Grok and Ollama Cloud. Switching provider is a setting, not a migration.

Priced before it runs

Every model carries a published rate. A model with no active rate cannot be served at all — the proxy fails closed rather than run up a cost nobody quoted.

Quoted, then metered

The rate is frozen before the call and metered against after it. Editing the price book mid-conversation cannot change what you are charged for a call already in flight.

Cheap and premium, side by side

Haiku 4.5 for volume, Sonnet 4.6 for the everyday, Opus 4.8 when the reasoning has to be right. A vApp can ask for one by name or take the default.

Or bring your own key

Managed is a mode, not a lock-in. Point the same gateway at your own provider account and the platform keeps working — you just pay your provider instead of us.

The Managed AI console listing the available models by provider, each with a switch, and a default-model selector.
Choose which models your organisation may use. The setting applies to every managed call, from every app.

The boundary

Every managed service has a free local one.

Managed is a choice you make per service, and you can unmake it. Nothing here is a one-way door.

Runs on your machine · free

  • OCR — Tesseract, in the browser. The file never leaves the device.
  • Transcription — Whisper, in the browser.
  • Vectors — stored in your own Odoo database. This is the default.
  • Your own API keys — the gateway calls your provider account directly.

Runs on ours · metered

  • OCR — cloud vision, for the scans Tesseract cannot read.
  • Transcription — longer files, faster, no browser tab held open.
  • Vectors — hosted pgvector, when the corpus outgrows the database.
  • Chat and embeddings — our provider accounts, our rates, your balance.

What stays put either way: your records. The proxy carries the text of the request and returns the answer. Your Odoo database is not replicated to us, and the business data a vApp reads stays under the same permissions it always had.

Knowledge

Answers out of your own documents.

Point it at your manuals, contracts, catalogues and drive folders. It indexes them, searches them properly, and shows you where each answer came from.

Two searches, fused

Meaning and words at once: a vector search for what you meant, a Postgres full-text search for the part number you typed. The two rankings are fused by reciprocal rank fusion, so an exact string cannot be lost to a near-synonym.

It rewrites the question first

A follow-up like “and the previous one?” is rewritten into a search that stands on its own. Ask for more recall and it will run several rewordings at once, or search with a hypothetical answer instead of the question.

A reader that re-sorts

Retrieval is fast and approximate. A reranking pass then reads the shortlist against your question and re-orders it, so the top of the list is deliberate rather than merely close.

Tables and images, not just text

PDFs are opened page by page — text, tables as tables, and embedded images. A scanned page is rendered and read by OCR. Word, PowerPoint, Excel, CSV and Markdown are parsed natively; audio and video are transcribed first.

A graph over the corpus

Entities and their relationships are extracted once per document and scored by centrality, which nudges near-ties toward what your corpus treats as important — and groups the whole corpus into themes. On by default; turn it off to save credit.

Permissions per document

Each indexed document is shared deliberately — everyone internal, named people, or named groups. A search only ever returns what the person asking is allowed to read.

Citations back to the source

Answers carry their sources, and a source links to the original document — not a paraphrase of it. When the answer matters, you can go and check it.

Every knob is yours

How many fragments reach the model, how many candidates are retrieved before reranking, chunk size and overlap, temperature — all of it is settings, not a support request.

The Knowledge settings panel: search mode set to hybrid, query rewriting, multi-query and HyDE, AI reranking, small-to-big, and the chunking controls.
The retrieval engine is configuration, not a black box.
The Knowledge ingest panel, listing the accepted formats and the options to keep the original file and to build the knowledge graph.
Drop in documents, paste text, or index Odoo records directly — your own permissions decide which records you can index.

Agent Center

Agents that do the work, at the autonomy you set.

An assistant answers. An agent finishes the job — reads the records, decides, and writes back. The difference that matters is how much it may do without asking, and that is a dial you control.

  1. 0

    Observes

    Reads and reports. Touches nothing. This is where an auditing agent lives, and where every new agent should start.

  2. 1

    Proposes

    Drafts the change and hands it to you. Nothing reaches a record until a person says so.

  3. 2

    Acts with approval

    Queues real actions for review. Approve one and the platform runs it as the person who approved it — under their permissions, in their name, in the audit trail.

  4. 3

    Limited autonomy

    Acts inside the tools and fields it was given, and no further. The ceiling is the agent's own definition, not its judgement.

The Agent Center roster: four agents — Auditor, Collections, Knowledge and Pipeline — each showing its autonomy level and how many tools it may use.
Each agent carries its autonomy level and its tool count on its face.

It runs on your data, with your permissions

An agent reads through the same ORM every app uses, so it can never see a record its user could not. There is no separate copy of your database for the agent to read.

An approval queue, not a log of surprises

Anything above its level waits for a person. You see what it wants to do, on which record, before it happens — not afterwards in an audit report.

It works on a schedule

Agents run on their own from a cron, not only when someone opens a window. The overdue-invoice pass happens on Monday whether or not anybody remembered.

It remembers

Findings and decisions persist between runs, so an agent that flagged something last week does not raise it again as though it were new.

Residency

You choose where the vectors live.

This is a compliance decision, so it is a setting rather than a support ticket — and the trade-offs are written down.

Default

In your database

Vectors are rows in your own Odoo database. Nothing to provision, nothing metered, and they leave when your backup leaves.

Search
Meaning and keywords, fused
Scale
Comfortable to a mid-size corpus
Storage cost
None

High scale, still yours

pgvector, in your Postgres

The same database, with the pgvector extension doing the search. For corpora big enough that the plain path starts to feel it. A guide app walks the setup.

Search
Meaning and keywords, fused
Scale
Large — the search runs in Postgres
Storage cost
None

Zero operations

Hosted by us

Vectors live on our pgvector cluster. Nothing to install, nothing to tune — and the same search you get anywhere else: meaning and keywords, fused by the identical ranking the local modes use. Residency is a compliance decision here, not a retrieval-quality one.

Search
Meaning and keywords, fused
Scale
Large — we run it
Storage cost
$0.15 / 1,000 vectors / month

The knowledge graph works in every mode — it is built from your documents and stays in your database even when the vectors do not.

The Knowledge index view: vector and document counts, embedding dimensions, storage on disk, index capacity used, and the list of knowledge bases.
However you answer the residency question, the index tells you what you are storing and what it costs.

Pricing

Pre-paid, and visible on screen.

A monthly plan grants credit; the credit pays for what you use. You can also skip the plan and simply top up. What you cannot do is spend money you have not put there.

Free trial

Try it

5 of credit, once

No card. It is already on the instance — open the Assistant and start spending it.

Get the desktop first

Plan

Starter

20 per month

$20 of AI credit granted every month.

Start

Plan

Business

499 per month

$650 of AI credit granted every month. That is 30% more credit than the price.

Start

Hosted vector storage is billed separately at $0.15 per thousand vectors per month, and only if you choose hosted residency. Unused credit does not become a refund.

The Managed AI consumption view: spend and call count over thirty days, broken down by chat, embeddings, reranking, storage and indexing.
This is the promise instead of “unlimited”: you can see where the money went, by service, per user and per app — and export it.

It is already on your desktop.

Managed AI is not something to install. Open the Assistant on a Vizum desktop and the first $5 is there — decide about a plan when you have spent it.