Configure detection

Detection is where you tell Airentect what sensitive means for your organisation. A generic classifier knows what a credit card number looks like; it does not know your project codenames, your regulators, or which data subjects you handle. The detection profile closes that gap.

Detection profiles

A detection profile carries the context the classifier uses on every prompt:

• Country. Sets the jurisdiction and pre-fills a baseline of that country's regulations, so an Australian organisation starts with Australian obligations rather than a blank list.
• Industry. Sector and sub-sector. Choosing a sector layers its specific regulations on top of the country baseline. Anything you have already added by hand is kept.
• Regulations. The obligations that apply to you. Start from the pre-filled baseline and add or remove as needed.
• Data subjects. Whose data you handle, such as customers, employees, or patients.
• Sensitive topics. Subject areas that are confidential in your context, such as an unannounced product or an ongoing dispute.
• Internal terms. Your own vocabulary: project codenames, internal system names, deal names. These are the highest-signal entries because no general-purpose classifier could know them.

The profile sharpens detection rather than broadening it. Context tells the classifier what is confidential to you; it does not flag general or public questions that merely mention those topics. An employee asking a public question about your industry is not an incident. An employee pasting a document that names your codename project is.

Guided setup

You do not need to build the profile field by field. The Guided setup button on the profile page opens a wizard that walks through five quick questions covering region, industry, regulations and data subjects, sensitive material, and a final review. It takes about two minutes, and every answer pre-fills the sensible defaults for your jurisdiction and sector so you are editing a reasonable starting point instead of typing from scratch.

The wizard is safe to re-run. Changing an answer merges new defaults into your existing lists rather than overwriting what you have added.

Test detection

The Test tab runs a real classification against your current configuration, using the same gateway your endpoints will use. Paste a prompt, run it, and see exactly what the classifier finds: the risk categories, the detected entities, and how your profile influenced the result.

Use it before rollout with a handful of realistic cases:

• A prompt containing one of your internal terms, which should be flagged.
• A prompt with an obvious secret such as a password or an API key, which should be flagged.
• A harmless public question that mentions your industry, which should pass.

Warning: Test with realistic but non-sensitive stand-ins. The Test tab performs a genuine classification, so treat what you paste there the same way you would treat a real prompt.

If a test result surprises you, adjust the profile and run it again. Iterating here is cheap; discovering a gap after a thousand endpoints are enrolled is not. Once detection behaves the way you expect, move on to policies, which decide what actually happens when something is flagged.