The questions we hear before people get started

Belowdecks is an agentic operating system: AI agents move your company’s processes forward on your own server, from prospecting to after-sales service. Each agent gets a specific mission, tools and a cost cap. It starts working when something happens: an email arrives, someone fills in a form on your website, a file lands in a folder, a scheduled time comes, or one of your applications sends a webhook.

Actions that bind your business are set up to wait for your operator’s go-ahead, on a confirmation card or through an approval. Every run ends with a summary, its cost and a verdict: Delivered, Partial or Blocked.

Four actors are involved, and this website always names them the same way.

The end customer is someone outside your company: a visitor to your website, a customer, a supplier or a candidate. They write to you through your usual channels, such as your website, email or a form. In a conversation, they can confirm the card that sums up their own request, if your website assistant is set up that way, and that ends their part.

The operator is the person in your company who runs Belowdecks. They receive the agents’ questions, approve their work and decide what goes out. They do it in their own space: an email with signed links, the instance’s administration or the team space. Several people on your team can share this role.

The agents are autonomous AI agents working on your dedicated instance. They start on an event, prepare, compare, draft and record, stop to ask the operator a question, then finish with a verdict.

The Belowdecks team installs your instance and sets up the agents, functions and connections with you, then monitors everything. It works behind the scenes and never decides for you.

An example with fictional people: Marc, a property manager at Immeubles Rive-Sud, writes on the Entretien Boréal website. The front-desk agent asks him a question, then Marc confirms the summary of his request. The quoting agent prepares the quote from the price list and past quotes. Julie, the operator in sales at Entretien Boréal, receives it in her own space, approves it, reads the email that goes with it and decides to send it. Sending goes through a function or an email connection set up with you at installation.

Before you start, we install your dedicated instance, switch on agent confinement and check that it holds. We then set up the agents, functions, collections and connections to your software with you, and each agent goes through a rehearsal on your data before it is switched on.

Once the instance is running, we monitor it. Error reports reach us automatically, we apply updates and we review the instance’s health check, which flags a backup that is too old, for example. We also act on the insights drawn from the runs, such as a failure that keeps coming back or a time limit exceeded, and we follow the questions inbox to spot the ones left waiting. During the first month, a thirty-minute meeting each week goes over these points with your operator.

Decisions stay with your operator: answering the agents’ questions, confirming cards and choosing what goes out. Any change that affects your process is decided with you.

A chat assistant answers when you write to it, and everything after that is up to you: you copy its answer, send it and update your files. Belowdecks agents start on an event, without anyone opening a conversation, and carry out the steps of a process using tools: they read a document, compare it with a collection, prepare a draft or update a record.

When they are missing information, they ask your operator and wait for the answer. Your operator can also steer them in a conversation, in plain language, and an agent can answer your customers on your website.

An automation tool runs fixed steps that you map out in advance, which suits work that is entirely predictable. A Belowdecks agent reads free-form text, sorts a request, compares two documents and asks a question when a case falls outside the usual.

Belowdecks also keeps code for the predictable parts. Functions are named steps, tested on examples before you approve them, that run without a model and so without tokens. After a successful run, an agent can save its mission as functions, and later runs go through them; if a function fails, the agent takes over again.

We present them as five cycles: from first contact to cash, from order to cash, from purchase need to supplier payment, from hiring to an employee’s departure, and from bookkeeping entry to financial statements. After-sales service, internal requests and supplier follow-ups fit as well.

You start with one cycle, often with a single trigger and one or two agents, then extend it. The diagnostic is where we choose that first cycle and decide what stays in your operator’s hands.

Anything binding on your business, such as recording an order, changing a customer record or writing to your accounting software, is set up during installation to wait for your operator: a confirmation card in a conversation, or an approval before the agent starts. When the card calls a function or launches an agent, it shows a pre-filled form the operator can correct before confirming. When an end customer writes on your website, they can confirm the summary of their own request if the assistant is set up that way; any other card they raise goes to your operator, through a signed link sent by email. In the cycles we set up, decisions about people, such as hiring or dismissal, are made by a person in your company.

Other limits come from the software. Once confinement is switched on and verified at setup, an agent sees only its working folder. It can store a secret but never read it back. Each run has a cost cap, and a master switch stops any new run from starting.

It stops and puts its question to your operator in the questions inbox, often with answer options. The operator answers in the instance, from an email without signing in, or through the API, and the agent resumes its run with that answer.

A notice can also go out by email to the operator, or by webhook to one of your tools, when a question is waiting, when a run fails or when an approval is needed.

Your data is stored on your dedicated instance: a server, a database and accounts reserved for your business. No data crosses from one client to another.

Four flows leave the instance: the text an agent sends to the provider of its AI engine (Anthropic, on your own account); email notices and reports, sent through an email delivery service; technical error reports, which reach the Belowdecks team through an error-tracking service; and whatever you choose to send to the services you connect. The Security page details each safeguard.

Yes. We deliver and maintain your instance, and our team has access to it to keep it running and to answer your support requests. That access is governed by the agreement you sign with us. Answers to the agents and confirmations stay with your operator.

Anthropic bills tokens directly to your own account, opened in your name, with a spending limit you set, so you do need an Anthropic account. We do not resell tokens. Belowdecks can also run the Codex (OpenAI) and pi engines, and our offer is built on your Anthropic account.

Inside Belowdecks, every agent has a cost cap per run. Before you enable a scheduled agent, you see a forecast of its monthly cost, and the monthly report shows what each agent cost. Functions handle some tasks without a model, so they use no tokens.

You keep your software. Belowdecks connects to what you already use: your mailbox, your website (one tag to paste), schedules, RSS feeds, a folder where your tools drop files, and your other applications if they can send a webhook or accept API calls.

Belowdecks has no ready-made connector for every product, so at the diagnostic we work out with you how to connect yours: through their API or through an MCP server that opens their functions to your agents.

It depends on the cycle you choose and on your software, and we set it with you at the diagnostic, a two-hour workshop. After the setup, each agent goes through its rehearsal on your data, followed by a settling-in month with a thirty-minute check-in every week.

Agents work in the language you tell them to, and your operator can write to them in French or English. The assistant on your website answers your customers in their language when its instructions say so, and a draft reply can follow the sender’s language.

That assistant’s greeting and privacy notice can be written in French.

Yes. You can export your projects (agents, skills, functions, triggers and collections) at any time. The contents of your collections are added if you ask for them, up to a cap per collection, and the export says when it was truncated. An export never contains the values of your secrets. Your Anthropic account is opened in your name, and you keep it after you leave.

Belowdecks lets you say precisely where your data is and who can access it: it is stored on an instance reserved for your business, the Belowdecks team’s access is governed by your agreement, and the website assistant shows the privacy notice you write. An end customer’s session on your website sees only its own conversation and expires after 12 hours.

The AI engine’s provider may process the text your agents send it outside Quebec. Compliance with Law 25 remains your business’s responsibility, and we bring it up at the diagnostic.