Enterprise · Human-in-the-loop

AI agents that work with people, not instead of them

Name: S.C.A.L.A.
Price: 9.90 EUR
Author: S.C.A.L.A.

Where quality, cost or safety carry responsibility, the question is not whether to automate but how much autonomy each step has earned. S.C.A.L.A. designs agentic processes in which the AI prepares or executes and a named person validates — and autonomy grows only when measurements justify it.

How to read this page: what already exists in the S.C.A.L.A. assistant SARA is marked “In SARA”. What we design for client projects is marked “Pilot · to be validated”. No client results are reported here.

Two ways to fail

Doing everything by hand does not scale. Letting a machine do everything does not hold when one error becomes a rejected production batch, a wrong quote or a message the customer should never have received.

Technical documents arrive heterogeneous and incomplete, and the expertise sits in the exceptions.
Handoffs between design, estimating and production lose information every time someone retypes it.
Accountability must stay with a person who can say yes, correct, or stop.

Three levels of autonomy

A synthesis of our own, built on public work on levels of automation and human oversight (see references). Each process step is assigned one level, and the level can change.

1 / 3

Assistive

The AI prepares, the person decides everything

Used for irreversible or costly outputs, for the first period of any new process, and while the AI’s quality is still unmeasured.

The AI does: Drafts, extracts, proposes — and states where each item comes from.
The person does: Reviews and signs every output before anything is executed.
Controls: Full review, side-by-side view of source and proposal, no execution without sign-off.
Measured by: Correction rate per field, review time.
Setting name in SARA: OSSERVA

2 / 3

Supervised at checkpoints

The AI executes within rules, the person approves at control points

Used when rules are explicit but judgment is still needed, and errors can be caught before production.

The AI does: Executes inside defined rules and routes to a person what is below a confidence threshold, missing, or outside the rules.
The person does: Approves at checkpoints and handles the routed items. Signs 100% of production files at the start, then a sample.
Controls: Routing by confidence per field, mandatory-field checks, exception queue with an owner.
Measured by: Correction rate, review time, share routed to review, errors reaching production.
Setting name in SARA: SEMI-AUTO

3 / 3

Autonomous with audit

The AI executes, the person checks by exception and by sample

Used for serial, repetitive cases with certain rules, stable correction history and errors that are cheap or caught downstream.

The AI does: Executes end to end and flags exceptions.
The person does: Resolves exceptions and audits a sample, whatever the confidence.
Controls: Fixed-percentage sampling, limits and stop switch, step down to a lower level if measures worsen.
Measured by: Errors reaching production in the audit sample, exception rate, drift over time.
Setting name in SARA: FULL-AUTO

The level belongs to a step, not to a system

Reading a drawing, choosing a component by certain rule and signing a file for production are different steps with different risk. Each starts at the level its risk deserves, moves up only on measured evidence — corrections, review time, errors reaching production — and goes back down when those get worse.

Design principle · Pilot · to be validated

The loop: the AI proposes, a person validates, corrections are reused

Every correction is information. The point is that it is kept, traceable, and available to the next proposal.

The AI proposes
With the reason or the source of each item, so the reviewer can judge quickly.
A person validates
Accepts, corrects or rejects. The decision is recorded with who and when.
Corrections are reused
A correction becomes knowledge the system can retrieve next time.

In SARA

Proposed actions can wait in an approval queue; approvals record who approved and when.
A negative rating with a correction adds a knowledge entry that later answers can retrieve.

Pilot · to be validated

For technical-document projects, corrections would become new rules or reference cases. We confirm in Discovery that this works for your process before promising it.

This is a knowledge base that grows, not retraining of the underlying models.

Where this applies

Areas of work where we are exploring or designing human-in-the-loop processes. These are areas, not references.

Area · pilot · to be validated
Made-to-measure furniture
Turning project documents into production documents, with a person signing every production file at first and a sample later.
Area · pilot · to be validated
Hospitality
Serial configurations — many identical units — where the AI checks the series and a person handles the exceptions.
Area · pilot · to be validated
Contract
Preparing quotes and estimates from project documents, with approval above agreed thresholds.

No client results are published on this page.

Controls and audit

In SARA

Autonomy level per account: off, observe, semi-auto, full-auto.
In semi-auto, low-risk actions run and high-risk ones go to an approval queue that expires if left unanswered.
Approver and time are recorded on every approval.
Guardrails correct or replace answers that break defined rules.

Pilot · to be validated

Confidence thresholds per field, not one global value.
Checks for missing mandatory information.
Fixed-percentage sampling, independent of confidence.
Exception queue with an owner and a deadline.
Correction rate, review time and errors reaching production, reported per step.
Stop switch and automatic step down to a lower level.

Security and data

Application data is hosted on EU infrastructure, encrypted at rest and in transit, with role-based access. The person who signs and the person who audits are named by the client; we provide the traceability.

Which AI models process which documents, and where, is decided per project and written into the contract.

Read about security →

How we start

Discovery
A defined, paid piece of work: we map the process, measure the baseline, set the starting level for each step and write the criteria for moving up.
Pilot by milestones
Short cycles with a demonstration on your real cases and a go or no-go at each milestone.
Operations
The process runs at the agreed levels, measured, with a named owner on your side.

Public references

The three levels are our synthesis. These are the public sources behind the ideas (consulted 23 October 2026).

EU AI Act, Article 14 — Human oversight
Oversight proportionate to risk and autonomy; awareness of automation bias; ability to override and to stop. Applies to high-risk systems.
Parasuraman, Sheridan, Wickens (2000), levels of human interaction with automation
IEEE Transactions on Systems, Man, and Cybernetics — Part A, 30(3). Automation level is chosen per function.
NIST AI Risk Management Framework 1.0 (January 2023)
Clear roles and responsibilities for people who use, interact with or manage AI systems.
Amazon Augmented AI: human review activation conditions
Review by confidence range, by missing key and by fixed-percentage sampling.
Anthropic, Building effective agents (19 December 2024)
Human checkpoints and stopping conditions for agents.

Have a process where a person must stay in control?

Tell us which one. In a Discovery we map it, measure where it stands and propose the level for each step.

Request a Discovery →

Or write to [email protected].

AI agents that work with people, not instead of them

Two ways to fail

Three levels of autonomy

Assistive

Supervised at checkpoints

Autonomous with audit

The level belongs to a step, not to a system

The loop: the AI proposes, a person validates, corrections are reused

The AI proposes

A person validates

Corrections are reused

Where this applies

Made-to-measure furniture

Hospitality

Contract

Controls and audit

Security and data

How we start

Discovery

Pilot by milestones

Operations

Public references

Have a process where a person must stay in control?

AI agents that work with people, not instead of them

Two ways to fail

Three levels of autonomy

Assistive

Supervised at checkpoints

Autonomous with audit

The level belongs to a step, not to a system

The loop: the AI proposes, a person validates, corrections are reused

The AI proposes

A person validates

Corrections are reused

Where this applies

Made-to-measure furniture

Hospitality

Contract

Controls and audit

Security and data

How we start

Discovery

Pilot by milestones

Operations

Public references

Have a process where a person must stay in control?