The best AI products put people in the right place. Learn how MedGAN AI decides when AI acts alone, when it suggests, and when a human must decide.
Full automation is rarely the right goal
When companies start with AI, many aim for full automation: let the system handle everything, end to end. It sounds efficient. In practice, it is often the fastest way to lose trust.
AI makes mistakes, and in business those mistakes have consequences: a rejected candidate, a wrong answer to a customer, a bad number in a report. The products that succeed are the ones that put people in the right place, not the ones that remove them. As an AI product manager, deciding where that place is counts among the most important design choices I make.
This is called human-in-the-loop design. Here is what it means, and how we apply it at MedGAN AI.
What human-in-the-loop AI means
Human-in-the-loop (HITL) AI is a design approach where people review, correct, or approve an AI system's work at key points, instead of the AI acting fully on its own. The human is part of the process by design, not an emergency brake added later.
It is not a sign that the AI is weak. It is how you match the level of automation to the level of risk, and how the system keeps getting better: every correction a person makes becomes a signal for improving the product.
Three levels of autonomy
We design every AI feature at one of three levels, based on the cost of a mistake.
AI acts on its own
For low-risk, high-volume tasks where a mistake is cheap and easy to fix. Example: tagging and routing incoming messages. The AI acts, and people review a sample.
AI suggests, a person decides
For tasks where the AI can do most of the work but a person should confirm. Example: drafting a reply or shortlisting candidates. The person approves, edits, or rejects with one click.
A person decides, AI assists
For high-stakes decisions with legal, financial, or personal impact. Example: a final hiring decision. The AI gathers and summarizes information, but the decision belongs to a person.
| Level | When to use it | Example |
|---|---|---|
| AI acts on its own | Low risk, high volume, easy to undo | Routing support tickets |
| AI suggests, person decides | Medium risk, AI does most of the work | Drafting replies, shortlisting |
| Person decides, AI assists | High risk or sensitive impact | Final hiring or credit decisions |
A product can use all three levels at once. What matters is choosing deliberately, task by task.
Five design rules for trustworthy AI
1. Show the reasoning
People trust what they can check. Show why the AI reached a result: the evidence it used, the parts of a document it relied on, or the factors behind a score.
2. Make correction effortless
If overriding the AI takes longer than doing the task by hand, people will stop using it. Editing, approving, or rejecting should take one step.
3. Signal uncertainty
When the AI is unsure, say so and route the case to a person. A system that knows its limits is far more trustworthy than one that is always confident.
4. Hand over with context
When a case moves to a person, pass along everything: the conversation, the data, and what the AI already tried. Nobody should have to start from zero.
5. Learn from every correction
Treat each human edit as feedback. Over time, the cases the AI handles alone should grow, safely.
How this looks in our products
- MedGAN Support answers routine customer questions on its own and hands harder or sensitive cases to the team, with the full conversation attached.
- MedGAN Align reviews applicants and explains its recommendations in plain words, but the hiring decision always stays with a person.
- MedGAN Atlas explains what changed in your data and why, so the people making decisions can check the reasoning before acting.
The same thinking runs through our custom AI development. For more on how autonomous systems work, see our guide to agentic AI.
Start with more oversight, then earn autonomy
Our advice to every client is the same: launch with more human review than you think you need, measure how often people agree with the AI, and widen its autonomy as the evidence grows. Trust is earned in steps, and the numbers tell you when each step is safe. We cover how to track those numbers in how to measure AI product success.
If you are designing an AI product and want to get the balance right, talk to our team.
