Hybrid AI workflow
The controllable AI workflow: who decides what, and when.
In a hybrid organisation, AI colleagues do the preparatory work and your people make the decisions. That only works if the workflow enforces it: fixed moments at which a person decides, and a trail of what was decided.
ForDirectors, operations managers and whoever is responsible for quality and compliance.
What makes a workflow controllable
Control is in the design, not in good intentions.
Six requirements, and for each one how we set it up in LEO. If you use another system, they are good questions to ask.
- 01
A person starts the work
Whoever chooses the assignment and the documents sets the boundaries of the work.
How Prudai does it
A workflow only starts once the user confirms the intake card. LEO can pre-fill the card, but never chooses the case documents itself.
- 02
An approval point after the plan
The plan is the cheapest moment to change course. A mistake in the assignment is fixed here with one sentence.
How Prudai does it
You choose approve, revise (with an explanation) or stop. In most workflows the work waits for that decision; in a few fact-finding workflows it continues after a set period.
- 03
The reviewer knows what to check
Pressing a button is not control. An approval point comes with an instruction.
How Prudai does it
Every approval point in the LEO workflow templates has a written instruction for the reviewer. If your own workflow lacks one, LEO proposes one at the start; if that fails, the reviewer gets a general instruction.
- 04
Challenge before review
A language model readily agrees with itself. A counterpart role looks for what the others missed.
How Prudai does it
In many workflows a counterpart role reviews the work from the other side, such as the tax inspector or the opposing party. It is an AI role with an instruction, not a guarantee that every counterargument is found.
- 05
Every decision is recorded
Afterwards it must be on record who decided what, and when.
How Prudai does it
The decision is recorded with who, when and which round before the work continues. If recording fails, the work does not continue.
- 06
Nobody forgets an approval point
An approval point that sits unanswered holds up the whole process.
How Prudai does it
When a decision is waiting, the owner gets a notification in the app and, depending on notification settings, by email. A reminder follows after 24 hours, then one per waiting day.
Design
Where do you put the approval points?
Let a person decide where the direction of the work is set, where it becomes irreversible, and where it affects a person. In between, the AI may keep working.
At the start
Which assignment, which documents, for which side? Whoever chooses the documents decides what the AI gets to see.
After the plan
Is the assignment right, is anything missing, is the scope well chosen? Changing course is still cheap here.
Before anything leaves the building
Advice, a letter to an opposing party, a decision about a client or employee.
Too many approval points is a risk too: whoever presses approve all day stops really looking. Choose few moments, and make them count.
Set it up yourself
Your own way of working, with the same safeguards.
In LEO you copy a template into your own workflow and adapt it in the workflow editor: steps, decision points and human checks. The copy only gets the sources included in your subscription.
The same principle runs through IRMA. The registers work entirely without AI. AI suggestions are off by default for each organisation; if you want to use them, we switch them on for your organisation. Even then, nothing enters the register until an employee approves it, and every approved change is in the audit log.
Accountability
Who decided? That is recorded.
At an approval point, who decided is recorded. That is also the question that comes back in liability for AI: who had control over the work, and who could have seen the mistake?
For high-risk AI systems, the European AI Act sets requirements for human oversight. Following the amendment by Regulation (EU) 2026/1744, they apply to Annex III systems from 2 December 2027, and later to products under Annex I. An organisation that uses such a system then has its own duties, including human oversight (Article 26). For most knowledge processes, an approval point is good practice, not a legal duty. In addition, Article 22 of the GDPR gives people the right not to be subject to a decision based solely on automated processing that produces legal effects or similarly significantly affects them. A genuine review by a person means a decision is not based solely on automated processing.
Frequently asked questions.
What is a controllable AI workflow?
A defined process in which AI colleagues carry out steps and a person decides at named moments. In most workflows the work does not continue at those moments without a decision, and every human decision is recorded.
What is the difference between human in the loop and human on the loop?
With human in the loop, the process waits for a person’s judgement. With human on the loop, a person watches and can step in, but the process keeps running. Which one fits depends on what is at stake.
Can I change course without starting over?
Yes. At the approval point you choose revise, with an explanation. By default an approval point comes round at most three times; in the last round you choose approve or stop.
What happens if nobody responds?
The owner gets a notification in the app and, depending on notification settings, by email. A reminder follows after 24 hours, then one per waiting day. In most workflows the work waits until a decision is made.
Is an approval point a legal requirement?
For high-risk AI systems, human oversight duties are coming, for Annex III from 2 December 2027. For most knowledge processes, an approval point is good practice. Article 22 of the GDPR covers decisions based solely on automated processing that produce legal effects or similarly significantly affect someone.
Does this also apply to an ordinary chat with an AI colleague?
No, this is about workflows. In an ordinary chat you review the answers yourself before you act on them.
Which decision in your process do you always want to make yourself?
Start there. We then set up the rest of the work so that it is ready for you at that moment.
Schedule a conversation