The controlled agent operating system for business operations.

ABN runs inside the customer environment, maps approved systems, learns operational workflows and creates controlled AI agents that prepare evidence-backed work for human approval.

Customer data stays in the customer environment. Agents operate within permissions. Every important output is linked to evidence.

Book a call →See how it works →

Local-first by design

Raw customer operational data is designed to stay inside the customer-controlled node in normal use.

Fast value, slow certainty

First process map and early findings can begin during onboarding; autonomy increases only as evidence and confidence improve.

Logged agent runs

Each controlled run records phases, evidence references, confidence and approval state for review.

01

Three layers.
One conscience.

The ABN Node consists of three tightly interwoven layers. Observer sees. Process Graph understands. Agent Engine acts. Between them sits a traceable conscience —Culture Rules — that reviews every output before it reaches you.

01OBSERVER

Observer Layer.

Reads your systems in real time via Nango. Data Minimizer whitelists fields; PII Guardian tokenises values. No raw values leave the Trust Layer.

Connections200+
Polling5 min
StorageWatermark only
WritesRead-only
FIG. 01.A — READING FRAMES
02PROCESS GRAPH

Process Graph Engine.

Builds a living process graph via pm4py. Every variant, every deviation, every baseline. The graph is your process map around the clock.

Algorithmpm4py
VersioningAuto
Pattern libraryGrowing
LLM inputTokens only
FIG. 02.A — GRAPH WALK
03AGENT ENGINE

Agent Engine.

Generates agents from five archetypes. The OPERA loop runs five phases per run. Confidence Gates decide whether the agent runs, flags or waits.

Archetypes5
Phases / run5
LLM phases1 of 5
Avg run time2.4 s
FIG. 03.A — FIVE ARCHETYPES
02

The agent knows
what it doesn't know.

Every agent follows the same disciplined phases. Four are deterministic — same input, same output, always — and before delivery a Verify gate proves each value against its source. A single phase — Reason — uses an LLM, and then with statistics only, never raw values.

FIG. OPERA — PENTAGON LOOPDRG.OP-001 · v2.0
O
Observe — build a reliable picture.Fetches events, measures freshness, computes data confidence.
deterministic
P
Plan — patterns first, LLM as a last resort.When a plan matches the Pattern Library with high enough confidence, the agent skips the LLM entirely.
deterministic
E
Execute — safe steps retry, writes go to a person.A read or analysis step retries on failure (up to 2) with a per-step timeout; an external write is never retried automatically.
deterministic
R
Reason — interpret the results.The LLM sees tokens, not values. Returns a conclusion + confidence.
LLM · no data
A
Act — Culture Rules, then verify before delivery.Impact quantified in EUR; a Verify gate proves every value against its source (RAL) and blocks unproven output before the signed report is delivered and logged.
deterministic
03

Built for every company
with operational work.

Start with one workflow. Expand into an agent team.

ABN is not limited to the industries shown in the menu. The platform adapts to repeatable work that depends on systems, approvals, documents, exceptions and evidence.

Browse by industryBrowse by process
04

Trust by architecture.
Not by policy.

We don't ask you to trust a promise. The Node is designed so your raw operational data stays inside your environment in normal use — and ABN Platform AB does not receive a copy of it.

We built ABN to run on your own server, where the work happens locally. We trade in intelligence, not in your data — the Node is designed so your raw operational data stays inside your environment, and ABN Platform AB never receives a copy of it.
JACOB LINDGREN · Founder · ABN Platform AB
Subprocessor register →DPA →SOC 2 Type 1 readiness →
FIG. ND — BOUNDARY