Skip to content

AI for critical systems. Under control.

Reelease understands the code of your critical systems and turns it into use cases your AI agents run only as your rules allow. Every change is authorized by a person and recorded. Your data stays where you decide and never trains a model.

We email you the link to open on your computer.
Read the transcript

The safe way to put AI to work on critical systems and in regulated sectors.

Here, a mistake is a transfer that never goes out or a missed deadline.

Laura asks for what she needs and the agent proposes a plan.

Reelease checks the rules and the permissions…

…and signs the plan: if anyone alters it, it doesn’t run.

The lookups go ahead on their own.

The change waits for someone else: whoever starts the task can’t approve it.

María approves and the transfer goes out, on the same core as always.

Resolved within the shift, with nothing rewritten.

Every task leaves its run record: who asked, what was done and who approved it.

And you can stop an agent instantly.

On the systems you already run. No migration, no rewrite. And it keeps working when you replace them.

And how does the agent know what to do?

First, Reelease has read your system, exactly as it is.

It reads the code and the documentation and builds a map your team reviews.

And it leaves everything documented, with its sources, as PDF or Word.

Your team chooses which use cases each agent can execute.

Only what your team publishes can run. Everything else stays documented, for reference.

Every agent starts with no permissions: an admin turns on only what it needs.

Your AI, where you decide.

It runs where you decide, with the AI model you choose.

Your data is looked up, never copied. Nothing is used to train models.

For sectors that can’t improvise: the same control, with concrete cases in every sector.

Reelease, the control layer between AI and your systems.

AI for critical systems. Under control.

On a critical system, an agent’s mistake is a transfer that never goes out or a missed deadline.

It does what you sign off on. Nothing more.

The agent only runs the steps you published, and each one declares whether it reads or changes data. Your rules check it first, and the people you choose approve the exact values.

See how human approval works

On the systems you already run

  • IBM Z mainframe
  • Oracle
  • IBM i
  • Natural and Adabas
  • Desktop applications
  • .NET Framework
  • Java and the JVM
  • Web, scripts and APIs
  • Databases
  • Microsoft 365
See every integration

From your system to an agent at work

  1. 01

    Understand

    Reelease reads the code and documentation, builds a map of the system and draws out its use cases for your team to review. Every fact points to its source in your code.

    Bank of Z map

    • 308files read
    • 43COBOL programs
    • 9CICS transactions
    Use case in Reelease

    Executable use case

    Transfer funds between two accounts

    Reads
    What the screen asks for: source account, destination account and amount.ReadLines 487–488, 494
    Rule
    The source and destination accounts can’t be the same.Lines 455–460
    Changes
    Both balances: XFRFUN debits the source, credits the destination and logs both entries in PROCTRAN.ChangeLines 496–502
    Approves
    Someone else. Two-Key Approval applies by default: whoever starts the task can’t approve its change.
    Ends
    With your system’s own reply (“Transfer successfully applied.” or why it was refused) and its run record.Lines 456–457, 636–639
  2. 02

    Publish

    You choose which executable use cases each agent has: what it can do on your systems. Everything else stays documented, for reference.

  3. 03

    Execute

    Your team asks for the task in the chat, and the agent does it in an isolated folder, in the cloud or on the operator’s computer, only within what its rules allow. In Ask mode it looks things up and drafts without touching your systems. Everyone works with their own agents and work accounts.

    Held transfer TRX-RET-9002
    Before it starts: permissions, computer and accounts
    1. Read: done
    2. Reason: done
    3. Proof: waiting for approval
    4. Retry: pending
    Approval requiredChange

    Step 3 · Record the source-of-funds proof (RTDC)

    Someone else approves: Andrés Molina

  4. 04

    Control

    Every run leaves its run record: who started it, who approved each change, the fingerprint of the approved values and how your system responded. And you can stop an agent at any time.

    Run record
    1. Lucía Prieto launches the task
    2. Lucía Prieto reviews and approves the plan
    3. Andrés Molina approves change RTDC
    4. Andrés Molina approves change RTRY
    5. Run completed
    Kept even if the agent is deletedExport record

Autonomy, within your limits

You decide what each agent can do and who supervises it. Every guarantee can be checked: try changing the sample plan.

Sealed Plan

If the plan or its values change after approval, it doesn’t run.

Plan · Held transfer TRX-RET-9002
  1. 1Read transfer TRX-RET-9002GTRXRead
  2. 2Check why it is heldGRETRead
  3. 3Record the source-of-funds proofRTDCChange

    Values of the change in step 3

    transferIdTRX-RET-9002
    documentTypeSOURCE_OF_FUNDS
  4. 4Retry the transferRTRYChange
Plan fingerprint
73f640c5ed65 4431…matches
Step 3 values fingerprint
e23737b7f0fe 21a2…matches

SealedMatches what was approved: it runs as is.

Try tampering with it:

The fingerprint is computed in your browser with SHA-256, the same algorithm as the sample run record. Nothing leaves your device.

Check it yourself

Two-Key Approval

By default, whoever starts the task can’t approve its changes.

The four-eyes principle, or maker-checker, by default. The approval is bound to the exact values: if they change, it no longer counts. An agent can make the rule stricter, never looser.

Starts with no permissions

A new agent can’t touch any system.

  • Read emailOff
  • Call your systemsOff
  • Use the operator’s computerOff

An admin turns on each capability, one at a time.

Run record

Every run leaves its record.

Who started it, which plan version was approved, which rule applied, who approved each change and which account talked to each system. It carries the fingerprints of the plan and of each change’s values, which you can check in your browser. It’s kept even if you delete the agent.

Stop at once

Suspend it and everything stops.

Banking operationsActive · 2 tasks running
  • Check hold TRX-RET-9014Running
  • Read transfer TRX-RET-9021Running
  • Access to its connectionsAllowed

Whatever is running is canceled and the agent loses access. Only an admin can resume it.

In the European Union

Service in Madrid, database in Frankfurt.

AI models run in the EU too. If you need it, in your cloud or on your premises.

MadridFrankfurt

Designed to help you comply

Each agent’s documentation, ready for your risk assessment.

  • EU AI ActUsing AI systemsArts. 12, 14 and 26
  • DORABanking and insuranceArts. 9 and 28 to 30
  • GDPRPersonal dataArts. 5, 28 and 32
  • NIS2Essential and important entitiesArt. 21

The assessment is yours; the controls and the documentation, ours.

See compliance

For your committee:Security fact sheetData processing agreementSubprocessors

Credentials are also encrypted with AES-256-GCM, each operator uses their own accounts, and every run is checked before it starts.

See data and privacy

Cases you’ll recognize

One task from each sector: what the agent does, who approves and what you get.

Held transfer

  • CICS
  • email
  • tickets
Situation

Fraud controls hold a transfer and the customer complains.

Step by step
  1. Agent: Looks up customer, transfer and balance in CICS

  2. Agent: Reviews why it was held and prepares the retry

  3. Person: Someone else on the team approves the retry

  4. Agent: Runs the retry and emails the customer

Control: By default, whoever starts the task doesn’t approve the core change: someone else does.

What you gain

The transfer goes out within the shift, and the customer hears by email before complaining again.

Fewer complaints end up at the Bank of Spain, and your team doesn’t chase the same case twice. If the supervisor asks, the run record answers.

Why it matters

8,338

complaints about accounts, transfers and direct debits reached the Bank of Spain in 2024, mostly over fraudulent transfers and fees.

Source: Bank of Spain, 2024 Complaints Report
More banking cases

Compared with what you’re already evaluating

Migrating, automating screens, an AI chat or process agents. Reelease understands your system from its code and works on the core under your control, in Europe.

  • Understands your system from its code

    ReeleaseYes: No hallucinations, line by line

    The alternatives

    Migrate

    Partly: Only at the start

    RPA

    No: From the screen

    AI chat

    No: What you tell it

    Process agents

    Partly: The described process

  • Works on your core

    ReeleaseYes: CICS, PL/SQL and APIs

    The alternatives

    Migrate

    No: Replaces it

    RPA

    Partly: Through the screen, fragile

    AI chat

    No: Doesn’t touch it

    Process agents

    Partly: Mostly back-office

  • Runs a deterministic plan

    ReeleaseYes: Sealed Plan, step by step

    The alternatives

    Migrate

    Partly: When the project ends

    RPA

    Partly: Fixed script

    AI chat

    No: Only answers

    Process agents

    Partly: Configured process

  • A person approves every change

    ReeleaseYes: Two-Key Approval, exact values

    The alternatives

    Migrate

    No: The project is approved

    RPA

    Partly: If planned for

    AI chat

    No: By hand, afterwards

    Process agents

    Partly: At fixed points

  • Proof of every run

    ReeleaseYes: Full run record

    The alternatives

    Migrate

    No: Project documentation

    RPA

    Partly: Bot logs

    AI chat

    No: A chat thread

    Process agents

    Partly: Step log

What you’ll ask before you start

Where the AI runs, who approves, what happens if it gets something wrong, how you start and how you leave: the short answer, with the detail one click away.

Where does the AI run, and where does my data stay?

Wherever you choose: on the Reelease service in the European Union (services in Madrid, database in Frankfurt), in your cloud or on your premises, even with no internet access. Your databases are queried at run time, not copied.

See where your AI runs
Which AI model does it use? Can I bring my own?

By default, the models Reelease offers, inside the European Union. Your own provider or an open-weight model on your hardware works too. You choose per system and per agent.

See the models
Is any model trained on my data?

No. What the agent reads is used for that task only, and no AI provider trains its models on it.

See what happens to your data
Who approves what the agent does?

Every change is authorized by whoever you choose, per agent and per step: by default, before it runs and after seeing the exact values. Two-Key Approval applies: whoever starts the task can’t approve its changes. If an admin pre-authorizes a fixed-value step, they record the reason and it goes into the run record.

See how human approval works
What if the agent gets it wrong?

If the model misreads a request, the mistake stays in the proposal. What it proposes goes through your rules before it touches anything: every step declares whether it reads or changes data, and only what is in your use cases goes into the plan. A sealed plan runs exactly as approved, with its steps and its values; if it changes after it’s sealed, it stops. Each change goes out only with its approval. And you can stop the agent at any moment.

See how every run is controlled
Do I have to migrate or rewrite anything?

No. Agents work on what you already run. If you replace a system tomorrow, you switch the connection and everything else stays the same.

See how you replace a system
How do I start without putting production at risk?

Every agent starts with no permissions, and Ask mode only reads. You grant access step by step, and every change goes through the approval you define. Free, Pilot and Pro have published prices; Enterprise is by contract.

See plans and pricing
What if I leave Reelease one day?

Your systems and your code stay as they were: there’s nothing to undo. On paid plans you export your organization in open formats (systems, agents, knowledge, use cases, policies and run records), plus the map and documents in PDF or Word. When the service ends, you have 90 days to take it all.

See the data processing agreement

Let’s talk about your first use case

Tell us which systems you run and what you want to solve. We’ll show you Reelease on one of your cases and work out together where to start.

You choose the cookies

We use technical storage to remember your choice and, only if you accept, analytics cookies (Google Analytics) to understand how the site is used. You can change your choice at any time. Learn more in our cookie policy.