Human-in-the-loop

Human-in-the-loop is the principle that a person stays in control of what an AI system does, at the points that matter. The term became widely known in machine learning, where people correct a model's outputs to train it. It now extends to AI agents, which act with real autonomy: the human no longer just trains the system, but frames what it does. In automated and agentic processes, the agent researches, analyses, acts and, above all, proposes. At the critical steps, a person validates the outcome, overrides or takes over, and every step goes on record.

Why human-in-the-loop matters now‍

The loop has moved. In machine learning, the human sat in the training loop, improving the model. With agents that read, draft, compare and act, the human sits in the decision loop, governing what the agent does. The question that decides whether agents reach production is not what they can do. It is who answers for what they did. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, and names inadequate risk controls among the three causes. Regulators ask the same question: for high-risk systems, the EU AI Act requires that a person can understand, intervene and override. And the people who would use these agents ask it too. In the Ofelia Summit 2026 barometer, 62% of respondents want a human to validate every agent action. Human-in-the-loop is the answer to all three. It is not a brake on AI. It is the condition for putting AI on work that commits the company: a discount, a contract, an invoice, an access.

Where to put the human in the process‍

Not every step needs a person. Put the human where the decision carries weight. In our view, three places stand out. Approvals: a discount above a threshold, a contract with non-standard terms, a purchase above a set amount. The agent prepares the file, the manager decides. Exceptions: the case that does not match the rule, the missing document, the customer nobody recognises. The agent flags it, a person takes over. Engaging actions: sending an email to a customer, issuing an invoice, changing a CRM stage, creating a user account. These actions commit the company. The agent drafts, a human confirms. Everywhere else, let the process run. A human who validates every single step becomes the bottleneck the process exists to remove. The goal is a few well-placed decisions, not constant supervision.

Human-in-the-loop by design: the agent proposes, the human decides, the record keeps both‍

Agents can now do far more than draft. They research an account across systems, analyse a margin against comparable deals, compare a contract with a playbook, fill a record from several sources, create a document or a page, and act on their own within set limits. Two things never belong to the agent: the guardrails it runs within, and the outcome it produces. People set the first. A person owns the second.

Take a quote-to-cash process. Agents can prepare most of it: the quote, the account history, the standard contract, the handoff brief for Finance and Customer Success. The critical moments stay with a person, and most of them are exceptions. A discount above policy: the agent analyses the margin, the account history and comparable deals, and recommends. The right person decides, by threshold: under 20%, the sales director; between 20 and 40%, the COO; above 40%, the CEO. A contract with redlines: the client returns seven. The agent checks each one against Legal's playbook, applies the five that match it and isolates the two that do not. A lawyer decides those two, not the whole contract, and approves what goes back to the client. Everywhere else, the process runs.

Then comes the part most teams skip: the record. For every step, it keeps who or what acted, on which input, under which rule, what the agent recommended and what the person decided, with a name and a timestamp. Six months later, the CFO or the auditor does not just see that a 28% discount went out. They see why, who approved it, and on what basis. The same record shows which recommendations people accept every time, and where an agent can earn more room.

That is human-in-the-loop by design: the process model fixes, before the first run, which steps are human, which are system tasks, which are agentic, the guardrails each agentic step runs within, and what the person sees when a decision reaches them. The person spends time on the decision, not on the preparation. Speed goes up. Control stays where it belongs.

Human-in-the-loop and the EU AI Act‍

Regulation now names the principle. For high-risk AI systems, the EU AI Act sets two obligations that matter here: human oversight and record-keeping. Human oversight means a person can understand what the system does, intervene and override it. Record-keeping means the system logs its activity so that anyone can trace decisions afterwards. The NIST AI Risk Management Framework points the same way: accountability needs defined human roles and a record of what the system did. Human-in-the-loop by design answers both at once. Defined decision points give the oversight. The trace of each proposal and each decision gives the record. Not every business process falls under the high-risk category, and this page gives no legal advice. Still, the direction is clear. A CIO who can show where humans decide, and prove it with a record, has an easier conversation with auditors, regulators and the board. Building it in from day one costs far less than adding it later.

From validating every action to governed autonomy‍

Human-in-the-loop is not a fixed setting. It is a dial. The autonomy literature gives the positions a name: in the loop, the person approves before the action; on the loop, the system acts within rules and the person supervises and can take over; out of the loop, the system acts alone. Most organisations start at the cautious end: a person validates every action the agent proposes. That is the right place to begin. It builds trust, it surfaces the cases the rules did not anticipate, and it produces a record of decisions. After a few weeks, the pattern becomes visible. Some steps get approved every time without a change. Those are candidates to move on the loop. Other steps keep producing edits or rejections. Those stay in the loop. In our view, engaging actions never go out of the loop. The move happens step by step, never for the whole process at once. The end state is governed autonomy: agents act where the rules are clear, people decide where judgement matters, and the trace shows both.

Ofelia and human-in-the-loop‍

Across the Ofelia suite, the principle is the same: agents work inside a frame people define, people decide, and every move lands in the audit trail. Where an agent acts depends on the product.

In Ofelia Agentic, Ofelia acts as an agent in two places. When someone asks a question in Slack or Microsoft Teams, Ofelia searches your approved documentation, answers and guides the next step, without leaving the conversation. The person stays in charge of what happens next. When a process owner builds a workflow, Ofelia proposes the model from existing documentation: the steps, the owners, the thresholds. The process owner reviews it, edits it in a few clicks, tests it in sandbox and publishes it. No developer, no IT ticket, and nothing goes live without that validation.

Once published, the workflow runs in Ofelia Workflow, and it is fully deterministic. It fixes the steps, their order, the owners and the thresholds. No agent decides the sequence. Human tasks go to the right person: in the revenue lifecycle, the discount above policy routes by threshold to the sales director, the COO or the CEO, and the routing survives the next reorganisation. Automated system tasks move the record, the document, the notification. Some tasks can be entrusted to an agent: preparing the quote, checking credit and payment terms against policy, comparing redlines with Legal's playbook. That is where human-in-the-loop takes its full meaning. The agent hands its result to a person, who validates it before the workflow moves on. In the quote-to-cash example above, Legal handles two decisions, not seven. When the signature lands, the workflow offers each owner to start what comes next. Nothing fires on its own. The provisioning ticket, the invoice, the kickoff: one click each, by the person who owns it.

In Bonita BPM, the same principle holds where the process needs its own application. Human tasks and approvals are modelled in BPMN 2.0 and assigned by role. The AI Agent Orchestrator allows to delegates a step to an agent inside a process that keeps control: the agent acts within the model, the process decides what happens next, a person validates where the model says so. Fifteen years of running processes for banks, ministries, insurers and retailers taught us where that validation belongs.

AI speed, human control: the agent proposes, the human decides, and the record shows both.

See Ofelia handle your workflows with HITL

Sources
European Union, AI Act, Article 14 (human oversight) and Article 12 (record-keeping). NIST, AI Risk Management Framework 1.0. Gartner, press release June 2025, over 40% of agentic AI projects will be cancelled by end of 2027. Ofelia Summit 2026 barometer (directional, internal).