octabits

AI integration

Eight questions to answer before your AI goes live

Introducing AI, or buying an AI gateway for it, settles how people reach a model. What often stays open: which data may go there? Who decides? And what can be shown later? We help you find the answers and build them in — with open-source libraries that run in your own environment.

An agent prepares a proposal. A person reviews it, and their decision is recorded.

The MIT-licensed libraries run in your application on your existing Postgres database; their source is public from November 2026. We do not host the kit or process your data. You can also run the model in your own environment.

Diese Seite auf Deutsch

  1. Classification
  2. AI proposal
  3. Human review
  4. Ledger

The difference

Detecting is not
classifying.

A PII filter finds names and account numbers. It cannot determine whether the information is confidential. The oversight kit checks each piece of content’s classification before every model call: public, internal or confidential. The classification determines what the model may receive. A detector running in your environment can raise the confidentiality level, but cannot lower it.

Measure your filterThe eight questions

AI assistant

Summarise for management: In a power cut the data centre lasts only 40 minutes. The contact is Jana Berger[NAME].

  • PII filter1 name redacted · nothing else found
  • ClassificationconfidentialLocal detection: vulnerability · can only raise confidentiality

✕ Confidential content is out

✓ Stays in your environment

Record

  1. 09:12 · public · EU provider
  2. 09:40 · internal · 1 name redacted · EU provider
  3. 10:02 · internal → confidential · vulnerability ×1 · self-hosted model

A digest is stored, not the content.

Confidential content goes only to a model running in your environment. The record keeps the classification and the kind of finding, not the text.

An example with invented data. We test detection accuracy using your data.

The questions

Eight questions, in the order an AI request travels

Each answer gets a traffic light: green means enforced in code and provable, amber means settled on paper only, red means nobody knows. Under each question: what the oversight kit does today.

Our answers at a glance

6of 8 enforced by the kit2we settle with you

  1. 01

    Who decides how confidential a piece of information is?

    Our answerEnforced by the kit

    There are three fixed levels: public, internal and confidential. Before every model call, the system checks what the model may see. Confidential content stays with a model in your environment.

  2. 02

    If someone types confidential data into an AI chat, does the system notice — and how reliably, in German?

    Our answerWe settle it with you

    Yes, once a detector is in place. It reads the text before each model call and can only make its classification stricter, never looser. If it fails or does not answer in time, the call counts as confidential. It runs in your environment: a pattern check for IBANs, tax IDs and phone numbers, and a language model you run yourself for names, health data or passwords. How reliable it is in German depends on the model, so we measure it with you, on your data (step 1 below).

  3. 03

    Is it then blocked, redacted or rerouted — enforced in code, or by policy only?

    Our answerEnforced by the kit

    The code enforces these rules. Internal content is redacted before it is sent outside your environment. Confidential content stays with a model you run yourself. If a model call bypasses the check, the build fails.

  4. 04

    Where do the models run, under which law — and can you switch provider?

    Our answerEnforced by the kit

    The model integration defaults to a provider based in the EU and also supports models you run yourself.

  5. 05

    What may the AI decide on its own?

    Our answerEnforced by the kit

    Nothing directly. The model returns a proposal and does not write to the database. A person accepts, edits or rejects it, and every change can be reversed.

  6. 06

    Can you prove a single AI decision six months from now?

    Our answerEnforced by the kit

    Yes. A ledger records each applied change: who decided it and on whose behalf. Every entry is chained to the one before, so later edits show. The kit generates the evidence for a period, with a readable cover sheet.

  7. 07

    How do you know the results are good — and would you notice if they got worse?

    Our answerEnforced by the kit

    The evidence shows what reviewers changed or rejected, and how often per field.

  8. 08

    Who is responsible for each AI application — and is there a complete list?

    Our answerWe settle it with you

    That is organisational work, not software: we draw up the register of your AI applications with you in the check.

How we work together

From the smallest step to a running application

Each step stands on its own, and its result is yours. You can start with the measurement or with the check.

  1. Step 1

    Measuring your PII filter

    We test your AI gateway’s filter on real, anonymised inputs from your organisation, beside a model in your environment. The result shows what each finds, what each falsely flags and what both miss.

    Fixed price, in the written proposal

  2. Step 2

    AI integration check

    We assess your organisation against the eight questions, compile a register of your AI applications and document the recommended actions.

    Fixed price, in the written proposal

  3. Step 3

    In your existing system

    One AI workflow in your system: reviewable, logged and reversible.

    8–10 weeks

  4. Step 4

    A running application

    We build a complete application for your workflow using the oversight kit. Approval uses the kit to manage software assessments and approvals.

    Your application →Approval →

Step 3 · The engagement

Implementation in your system

We implement one workflow in which a model prepares a proposal for human review, such as a vendor assessment, a record update or a document awaiting signature. Using the oversight kit, we add a review screen, a hash-chained change log with revert support, and checks on the content sent to each model call. The workflow runs in your environment on your Postgres database. You own the application code.

You get
A working workflow in your environment, tests and documentation of which data the model may access. We measure the time per case and how much reviewers change each proposal to help you assess the results.
Duration
8–10 weeks
Pricing
Fixed price, agreed after a first call; optional ongoing support

How it works

How the kit works: from proposal to recorded decision

01The cycle, for every AI request↻ and again with the next request

01

Classification

Content is classified as public, internal or confidential before each model call. Confidential content stays with a model in your environment. Internal content is redacted before it is sent outside. Content without a classification counts as confidential, and a detector can make a classification stricter, never looser. The same check applies to results your system hands to an outside agent over MCP. The record of each call keeps a digest of the content sent, rather than the content itself.

02

AI proposal

The model returns a typed proposal for existing records. It includes the proposed changes, the current values, the proposer and the person or organisation they are acting for. It does not write directly to the database.

03

Human review

A designated reviewer checks the proposal in your application. They can edit it, accept individual changes or reject it. If the underlying record has changed, the reviewer sees that before applying the proposal.

04

Ledger

A hash-chained ledger records each applied change, who made it and on whose behalf. Reversing a change adds a new entry and preserves the earlier records. The latest hash is regularly secured outside the database, so even a fully rewritten chain shows up at the next check.

Who it is for

Software vendors and regulated operators

Software vendors whose AI takes part in regulated decisions

If your product uses AI in recruitment, education, credit, insurance or access to services, your customers ask how people supervise its decisions and what the system records.

The obligations for high-risk systems listed in Annex III of the AI Act, among them automatic record-keeping (Art. 12) and human oversight (Art. 14), apply from 2 December 2027. That date comes from the 2026 amendment.

Human review and a tamper-evident change log integrated into your own product.

Regulated operators whose officers sign off in person

In critical infrastructure and in the financial sector, new software is assessed before it goes live. The security and data protection officers gather supplier evidence and each prepare an assessment before approving its use.

§ 30 BSIG counts supply-chain security and secure procurement among the minimum measures, and under § 38 management has to approve and oversee them. DORA requires a register of ICT third-party providers. The GDPR requires processing agreements, a record of processing activities and, where the risk is high, an impact assessment.

A model gathers the evidence and links each statement to its source. The responsible person reviews and signs the result. The kit runs in your environment. You can run the model there too.

Your legal adviser can assess whether your system is high-risk and which obligations the kit can help you meet. We provide technical implementation, not legal advice.

What exists today

The 9 packages of the kit

Each package has its own documentation explaining how to use it and where its limits are. We can build the application around these packages for your workflow.

All packages
  1. AI review and oversight9 packages

Sub-processors

What stays on your list of sub-processors

An AI application is often put together from one hosted service per job. Each of them receives your data, needs a processing agreement and is one more entry your data protection officer has to assess. On our packages these jobs run inside your application and your own database. What remains is where the application runs and which model it uses.

A common setup12sub-processorsOn our packages2sub-processors
HostingHosted serviceYour hosting provider
AI modelHosted serviceModel vendor, in the EU by default
Sign-inHosted serviceIn your application
Background jobsHosted serviceIn your database
WorkflowsHosted serviceIn your database
Live updatesHosted serviceIn your database
NotificationsHosted serviceIn your database
Document search for AIHosted serviceIn your database
Encryption keysHosted serviceIn your database
Approval of AI actionsHosted serviceIn your database
AI tool access (MCP)Hosted serviceIn your application
MonitoringHosted serviceSelf-hosted

Mail is left out of both columns: our packages send through your own mail server by default, and a sending service adds one entry on either side. With a model you run yourself, the second entry goes as well.

Not yet

What the kit does not do yet

  • Not yet

    Documentation of the system as a whole.

    The kit generates the evidence for each action and model call. During implementation, we work with you to document the system, its purpose and its limits.

  • Not yet

    Storage of the full content of model calls for a defined retention period.

    The kit currently stores only a digest. Requirements to retain the full content need to be assessed for each installation; this feature is not yet implemented.

  • Not yet

    A ready-made product for your workflow.

    Approval is one for software approval; we implement other workflows with the kit in your own codebase.

Discuss your workflow with us

Tell us what the model should prepare, who reviews the result and how the process works today. We can then discuss which workflow to implement first.

hello@octabits.ioContact