In this article
What does human in the loop mean?
IBM defines human in the loop as a process in which a person actively takes part in the operation, supervision or decision-making of an automated system, to support accuracy and safety. [1] Its page does not separate it from other models of oversight.
Databricks does. In its description, human in the loop means a person validates, corrects or approves AI output before action. Human on the loop means a person monitors and steps in by exception. Human over the loop means a person sets policy and audits, and is not involved in individual decisions. [2] We found no standards body or textbook that sets out the three terms, so treat these as one vendor's usage.
What does it look like in AI outreach?
In outreach the question is when the person can act. A person who reads the message before it sends is in the loop. A person who watches a dashboard and can pause a campaign is on the loop. A person who writes the rules and reads a monthly report is over the loop. A person who sees the message only after it has gone is doing review. That gives you a log of what went out and no control over it.
| Model | When the person acts | What they see before it sends | Who is accountable |
|---|---|---|---|
| In the loop | Before the message sends | The message, and the research behind it | The person who approved it |
| On the loop | While campaigns run, by exception | A dashboard or alert, not each message | The person monitoring, and the tool |
| Over the loop | Afterwards, by audit | The rules they set and a report | The person who set the policy |
| Review after sending | After delivery | Nothing: a log of what went | No one: nothing could be stopped |
The model you need depends on what an error costs. A wrong email to a buyer cannot be recalled, which is why we put the person before the send.
Why keep a person at all?
People over-trust automated output. A systematic review of 74 studies chosen from 13,821 papers found automation bias to be a fairly robust and generic effect across research fields, and in a small meta-analysis within it, wrong decision-support advice raised the risk of a wrong decision by about 26%. [5] The studies are mostly clinical, not sales, and the review also found that training, accountability and design choices reduce the effect.
Agents also fail often on business tasks. On CRMArena-Pro, a benchmark of sales, service and quote tasks in a synthetic Salesforce setup, leading agents reached only about 58% success on single-turn tasks and about 35% on multi-turn tasks. [6] That measures task success and says nothing about how often a person would correct the output.
More approval is not always better. Anthropic reports that experienced users of its coding tool run fully auto-approve in over 40% of sessions, against about 20% for new users, and says it does not read that as people giving up oversight. [7] Its guidance for agents notes that too many prompts get tuned out and too few risk misreading intent. [8] Put the approval where a mistake is expensive.
Does the law require it?
For high-risk AI systems in the EU, yes. Article 14 of the AI Act requires them to be designed so people can oversee them while in use: understand the system's limits, stay aware of automation bias, interpret and override outputs, and stop the system. [3] The page shows the Annex III systems applying from 2 December 2027.
Annex III lists eight areas: biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration and border control, and justice. [4] None names sales or outreach, and the nearest, recruitment under employment, is about hiring. The page does not say sales is excluded, so that is our reading, and it is not legal advice. Ask a lawyer about your own case.
Where Mach puts the person
In a Mach Playbook, every step that does something or decides something needs a person's approval before the Playbook can run. A wait or an end step has nothing to read, so it needs none. Until every step is approved, Run refuses and says how many steps still need you.
Changing the offering, the customer profiles, the personas or the channels switched on drops every approval and keeps the drawing, so the person reads again what changed.
A suggested answer to a reply sends only when the seller presses Approve. One press sends once. The send goes through the one door every sales email uses, which checks for a stop on the person or their company, an address that has bounced or complained, the warm-up cap and the sending domain, and records the touch.
Before a person reads a first message, rules can refuse it. For one firm in October, Mach was asked for 167 first messages, kept 62 and refused 105 under the writing rules. The reasons are in why Mach refuses some first messages.
How to test any AI sales tool
Ask at which step a person can block a message from sending. If the answer is 'after it sends', that oversight is review.
Time the review of ten real drafts before you commit to a list. Then track how many drafts are approved and how many rejected. A high rejection rate points at the research or the template. A very low one may mean nobody is reading carefully.
Common questions
What is human in the loop AI?
A process in which a person takes part in an automated system's operation, supervision or decision-making. [1] In practice it means a person approves, corrects or decides before the AI acts. [2]
What is the difference between human in the loop and human on the loop?
In the loop, a person approves or corrects output before action. On the loop, a person monitors and steps in by exception. Over the loop, a person sets policy and audits. [2] These are one vendor's definitions; we found no standards body that sets them out.
Is human in the loop required by law?
For high-risk AI systems under the EU AI Act, Article 14 requires human oversight. [3] Annex III names eight high-risk areas and none is sales or outreach [4], but the page does not say sales is excluded. That is not legal advice.
Does having a person in the loop make AI safe?
Not by itself. People over-trust automated advice, and a review of 74 studies found the effect across research fields. [5] Training, accountability and design reduce it, so design the approval step so the person has something to check.
What does human in the loop look like in AI outreach?
A person approves each message before it sends, and can stop it. In a Mach Playbook, every step that does or decides something needs approval before the Playbook can run, and a suggested reply sends only when the seller presses Approve.
Sources and further reading
- What Is Human In The Loop (HITL)?IBMBack to text
- What is Human-in-the-Loop (HITL)?DatabricksBack to text
- Article 14: Human OversightEU Artificial Intelligence Act (Future of Life Institute)Back to text
- Annex III: High-Risk AI Systems Referred to in Article 6(2)EU Artificial Intelligence Act (Future of Life Institute)Back to text
- Automation bias: a systematic review of frequency, effect mediators, and mitigatorsJournal of the American Medical Informatics AssociationBack to text
- CRMArena-Pro: Holistic Assessment of LLM Agents Across Diverse Business Scenarios and InteractionsarXivBack to text
- Measuring agent autonomyAnthropicBack to text
- Trustworthy agents in practiceAnthropicBack to text
