Automation and AI with control

Automation
you stay in control of.

BD Automate AI is a platform where teams build their own workflows, connect systems and use AI. Every step is a checkpoint, high-risk actions wait for a person, and for every run you see what happened, what went wrong and what it cost.

Version 1.0 works with REST APIs, data tables, email and webhooks. Integrations with Microsoft 365, Exact and AFAS, among others, are planned. Before use with real data, an independent security test comes first.See the roadmap

What BD Automate AI does for you

01 · Build

Build a workflow the way you draw it.

Drag steps onto the canvas and connect them: triggers, conditions, AI, email, approvals and more. The builder checks while you work, and you test a draft step by step before it goes live.

  • 22 step types, from webhook to AI agent
  • Publishing is only possible after a successful test
  • Every version is kept and can be restored

02 · Execution

Every step is a checkpoint.

If a step fails, a restart continues from that step. Steps that already succeeded are not repeated: a sent email does not go out again. Per step you set how often a temporary error is retried.

  • Retries with increasing waits
  • An error path for when it still goes wrong
  • An event ID starts only one run

03 · AI

AI you can check.

Models classify, summarise and extract data from text. External content is marked as untrusted and strictly separated from the instructions, and structured output is checked against a schema before a next step works with it.

  • Anthropic or an OpenAI-compatible API
  • Prompt library with versions, test sets and evaluations
  • Costs per call recorded and capped

04 · People and agents

High risk? Then a person decides.

Deleting, bulk email and other high-risk actions only go live when they sit behind an approval, and the requester never approves their own. AI agents work with a fixed list of tools and hard limits.

  • If the action changes after approval, it does not go ahead
  • Limits on steps, run time, calls and costs
  • An agent cannot carry out forbidden actions

05 · Integrations

Connect what you already use.

Any public HTTPS API, the built-in data tables, email through your own SMTP server and webhooks in and out. Other systems start workflows through a signed webhook or through the REST API with their own key.

  • Credentials encrypted and never visible
  • Webhooks only with a valid signature by default
  • Microsoft 365, Exact, AFAS and more are planned

06 · Security

Security in every layer.

Organisations are strictly separated, every action is checked on the server and outgoing API requests cannot reach internal addresses. In logs, secrets are always masked, and recognisable personal data such as email addresses, IBANs and BSNs by default.

  • Seven roles with permissions per action
  • Emergency switch: offline immediately, even without the interface
  • Encrypted backups with verified recovery

Modules

From canvas to control.

The most important components, each with a sample screen. The data comes from the BD Automate AI demo environment and is made up; the AI steps there run on a labelled test model.

A view of every automation

The dashboard shows runs, errors, the queue and AI costs, based on real production runs. Filter by period, workflow, team, integration, AI model or cost centre.

  • Key figures link through to the underlying records
  • Integrations with problems visible right away
  • Workflows with a striking number of errors listed separately

Sample screen with made-up data (the app itself is in Dutch).

Roadmap

Honest about where BD Automate AI stands today.

Version 1.0 has been built and tested. This is what exists now, what does not exist yet, and what is needed before you work with real data.

Ready and tested (version 1.0)

  • Workflow builder with 22 step types, versions and publishing after a test
  • Engine with checkpoints, retries, error paths and dead-letter queue
  • Triggers: manual, webhook, schedule, REST API and changes in data tables or documents
  • AI models, prompt library, assistants and agents with limits
  • Approvals, document review, costs, limits and audit log
  • 138 unit and integration tests and 21 browser tests

Not built yet, and needed before real data

  • Not yet: integrations with Microsoft 365, Google Workspace, Exact, AFAS and eleven others
  • Not yet: OCR for scans, and incoming email as a trigger
  • Not yet: invitations and notifications by email, and single sign-on
  • Before real data: a test with your own AI provider and keys
  • Before real data: a security test, a load test and a DPIA
  • Before real data: backups in a second location and monitoring

In the app, planned integrations are labelled ‘Planned’ and cannot be connected; the test model and the test CRM are called ‘Simulation’ there.

Made for the Netherlands

Dutch organisations as the starting point.

Not a translated foreign platform, but Dutch terms, notation and attention to privacy from the very first step.

Fully in Dutch

Screens, error messages and explanations in clear Dutch, also for people who are not developers.

Dutch numbers and dates

Conditions and transformations understand decimal commas and Dutch dates, and calculate without rounding errors.

Privacy before the model

If desired, email addresses, IBANs and BSNs (with eleven test) are masked before text goes to an AI model.

Documents with checks

AI extracts data from invoices and orders, the code checks IBAN, dates and subtotal plus VAT, and a person approves.

Schedules in Dutch time

Schedule workflows with a timetable and a time zone, such as every working day at 07:00 in Europe/Amsterdam.

Twelve templates

From preparing invoice data to following up customer activity, for customer service, finance, sales, HR, logistics and reporting, among others.

Security

Automation without a back door.

An automation platform holds the keys to many systems. That is why every outgoing connection, every AI call and every action runs through the same controlled paths.

Two-factor authentication

A code from the authenticator app that only works once and ten one-time recovery codes. Mandatory for platform administration, such as the emergency switch and recovery.

Credentials under lock

Encrypted with AES-256-GCM, with a key outside the database, and only usable for hosts allowed in advance. You never see the value, only a name and a hint.

Signed webhooks

Incoming webhooks require an HMAC signature with timestamp by default. Resending an intercepted request does not start a second run.

Segregation of duties

Whoever requests an approval cannot approve it themselves; the database enforces that. The approval records the exact action with a check value.

Masked logs

Passwords and tokens always, and email addresses, IBANs and BSNs by default, are masked in step input, output and the audit log. Integration tests check this.

Emergency switch

Offline with password and 2FA code: user sessions ended, webhooks and API on 503 and the queue paused. Also via the command line or with NOODSTOP=aan.

Verified backups

Encrypted with AES-256-GCM and verified immediately. Recovery only in emergency mode, all-or-nothing, with a safety backup first.

Independently reviewed

A separate reviewer, without knowledge of the design choices, found 17 findings, 2 of them high risk. All 17 have been resolved and covered by tests. An external pentest is still to come.

Who it is for

For everyone who makes processes smarter.

From whoever builds the workflow to whoever approves it and keeps an eye on the costs.

Process improvers and developers

Build, test and get published.

  • Visual builder with test dialog
  • Debugger per step
  • Templates as a starting point

IT and integration management

Integrations without loose scripts.

  • Credentials and webhooks in one place
  • Host policy per organisation
  • API keys with scopes and expiry date

Team leaders and approvers

Deciding where it matters.

  • Seeing exactly what you approve
  • Asking for more information with an explanation
  • Reviewing documents before forwarding

Controllers and management

A view of every run and every euro.

  • Costs per team and cost centre
  • Limits that warn or block
  • Audit log with CSV export

Approach

This is how we look at it together.

Step by step, with your own processes as the yardstick.

  1. Step 1: Introduction

    We discuss which processes take manual work today and show in the demo environment how they run as a workflow.

  2. Step 2: Try it out

    With fictitious data you build and test a workflow yourself, with the labelled test model instead of a real AI service.

  3. Step 3: Setup

    We tailor roles, teams, limits, allowed hosts and your AI provider to the organisation.

  4. Step 4: Further development

    Feedback goes straight to the developer. New integrations follow in tested versions.

Questions

Frequently asked questions

Is your question not listed? Ask it during the introduction.

Can we already use BD Automate AI?

Version 1.0 works: building, testing, publishing and running workflows, with AI, approvals and the integrations that exist. Before you work with real (personal) data, an independent security test, a DPIA, a load test, a test with your own AI provider, backups in a second location and monitoring are needed. During the introduction we discuss what that means for you.

Which AI models can we use?

Models from Anthropic and through OpenAI-compatible APIs. Your own model, for example with Ollama or vLLM, only works if it is reachable through a public HTTPS address with an API key: the security blocks models on the internal network. The test environment used a labelled test model; the connection with real providers follows their APIs, but has not yet been tested with real keys.

Does the AI make decisions on its own?

Only within what you allow. AI output is never trustworthy by itself: structured output is checked against a schema, an agent only uses tools from its list, and high-risk actions wait for a person.

Is there an integration with Microsoft 365, Exact or AFAS?

Not ready-made yet. Those three and twelve other integrations are planned and labelled ‘Planned’ in the app. Many of these services can already be reached through the general REST API connector; integrations with OAuth will follow.

What does AI usage cost?

That depends on your provider and model. Per call the tokens are recorded (from the provider, or estimated if it does not report them) and the costs are calculated with the prices you set per model. Every amount states where it comes from. Limits warn, block or ask for approval before the call.

What happens when a step fails?

Every successful step is a checkpoint. On a temporary error, a safely repeatable step tries again if you set that per step, waiting longer each time. If that does not work, the error path follows if you have drawn one; otherwise the run fails, and after a temporary error it goes into the dead-letter queue. A restart continues from the failed step; steps that already succeeded, such as sent emails, are not repeated.

Can the platform read documents and scans?

PDFs with a text layer and text files, yes. No OCR service has been connected for scans and photos yet; they honestly get the status ‘No text’. A person always reviews extracted data before it is forwarded.

What happens in case of a breach or outage?

A platform administrator takes the platform offline with password and 2FA code: user sessions are ended, webhooks and API return 503 and the queue pauses, so no emails or external actions go out after the running steps. Encrypted backups can be restored, with a safety backup first.

Introduction

Curious about BD Automate AI?

In the demo environment we show how a workflow is built, tested and approved, and look together at which of your processes qualify.