What Human-in-the-Loop (HITL) Means
HITL design means building AI features where humans remain meaningfully in control of the outcome — not just technically able to override, but genuinely empowered to do so.
The spectrum of AI autonomy
AI features exist on a spectrum of autonomy:
Fully automated: AI acts without human review. (e.g., spam filter silently deletes email)
Human-on-the-loop: AI acts, human can review after the fact. (e.g., fraud detection flags a transaction, human reviews log)
Human-in-the-loop: AI suggests, human decides. (e.g., BoostDraft surfaces an inconsistency, lawyer decides whether to fix it)
AI-assisted: Human acts, AI provides context. (e.g., search autocomplete, definition lookup on click)
BoostDraft sits mostly in human-in-the-loop and AI-assisted territory. This is intentional for legal documents: a lawyer must always make the final call. The AI is a highly capable assistant, not a decision-maker.
HITL design is about preserving the user's liberties — their ability to move freely, correct the AI, and stay in control. An AI with no override path is like a stone with no liberties: it gets captured.
Why 'technically possible to override' isn't enough
Many AI features claim to be 'human-in-the-loop' because there's technically a dismiss button. But real HITL requires:
Legibility: The user understands what the AI did and why. Not just 'AI flagged this' but 'this cross-reference points to a deleted section'.
Low friction override: Dismissing or correcting AI output should be easier than accepting it. If the dismiss action is buried, users rubber-stamp AI decisions.
Calibrated confidence: The AI should communicate uncertainty. 'High confidence inconsistency' vs. 'possible inconsistency — review recommended' are very different signals.
No automation bias: Design against the tendency to accept AI output without scrutiny just because it's convenient. In legal docs, this can be catastrophic.
BoostDraft's HITL design principles
BoostDraft's features are designed around three core HITL principles:
1. Show, don't replace — the AI highlights and annotates, but never replaces text without explicit user action. The original text is always visible.
2. Evidence-first — every suggestion includes the evidence. 'Defined term inconsistency: Section 1.1 defines 'Company' as Acme Inc., Section 8.3 uses 'the Company' to refer to a different entity.' You see the evidence before you decide.
3. Easy exit — every AI suggestion has a clear one-click dismiss that's at least as prominent as the accept action.