The confident-answer machine

The reason chat-style AI tools invent facts, cite made-up cases, and fabricate quotes isn't that they're broken. It's that they were shaped, deliberately, to always give you something. Modern language models are trained on human feedback, and one of the strongest signals in that feedback is completion. A response that says "I'm not sure, could you tell me X?" gets scored lower than one that guesses confidently and reads well. So the models learned to guess confidently and read well.

For consumer chat use cases this is mostly harmless. If you ask a chatbot for a recipe and it invents an implausible cooking step, you'll notice when the meal tastes wrong and move on. For professional documents, it's catastrophic. A consulting report with three made-up client names, a proposal that promises a feature that doesn't exist, a SOW that references the wrong regulatory framework: these aren't detectable at taste-time. They ship, and the consequences arrive later, in front of the client.

The generic response, which every AI vendor pitches now, is "just add better guardrails" or "run a RAG pipeline". Those help a little. They don't change the underlying incentive: the model is still optimising to complete, so when the retrieved sources are ambiguous or thin it still fills the gap with plausible fabrication instead of flagging the gap.

What a clarifications loop actually is

A clarifications loop is a workflow choice, not a model choice. The document generation system is designed so that, at every step where the AI is uncertain, the right behaviour is to surface the uncertainty as a question rather than to guess. The question goes to the human author. The answer becomes context for the rest of the draft, including subsequent sections.

The unit that matters is the clarification card: a small, in-line object that says "I noticed I don't have a confident answer for X, and X affects this section and the ones after it. Here's what I'd ask you if we were sitting together." The author accepts (with an answer), dismisses (marks it as intentional), or defers (comes back to it later). Accepted clarifications become persistent notes attached to the document, so the AI carries the answer forward across regenerations.

This is the opposite of a chatbot's "let me try to be helpful" reflex. The system is being explicitly told that saying "I don't know" is a better answer than inventing. The training doesn't need to change; the prompt and workflow around the model do the work.

Four things a clarifications loop needs to work

Not every "add a question mark" implementation counts as a clarifications loop. The pattern works when four conditions hold.

1. The AI knows what “confident” means for this document

The template needs to encode what a good answer looks like for each section, so the AI can tell when it has one and when it doesn't. A generic model doesn't know that a SOW deliverable needs a format, a recipient, a date, and an acceptance test. A template that enumerates those criteria does. When one is missing from the source material, the AI has a specific, non-ambiguous thing to ask about.

2. Uncertainty budgets are bounded per section

Every section produces at most a small number of clarifications (typically two or three). Without a budget, an over-cautious system asks fifteen questions per section and the author gives up. With a budget, the AI has to triage: which uncertainties actually matter downstream? The clarifications you get are the ones that would change what gets written in later sections, not the fine-grained stylistic ones.

3. Answers propagate forward, not sideways

When the author answers a clarification about the client's preferred term for "engagement", every subsequent section that uses that term picks up the answer automatically. If the same clarification would have been re-asked in section seven, the loop is broken. This is where most in-editor AI plugins fail: they have per-request context but no persistent notes layer that survives across sections.

4. Dismissing is different from ignoring

If the author saw a clarification and deliberately chose not to answer it, the AI needs to remember that decision. Otherwise the same question comes back on the next regeneration and the author feels harassed. A dismissed clarification is a signal: "you were right to ask, but the answer is intentional silence." Distinct from an unanswered clarification, which is a signal: "come back to this."

When AI should ask, and when it should just proceed

The instinct when you first meet a clarifications loop is to want the AI to ask about everything. In practice that produces worse output, because the author's attention gets diluted across noise. Good clarifications loops are opinionated about when to ask.

A useful heuristic: ask when the answer would change the subsequent sections of the document. Don't ask when the answer would only affect a phrasing choice inside the current section. So "is the September 12 workshop the discovery date or the kickoff?" is a good clarification, because getting that wrong cascades through Scope, Schedule, and Deliverables. Whereas "should we use active or passive voice in this paragraph?" is not, because it's reversible in one edit pass.

The other useful axis: ask when the source material is ambiguous, not when it's missing. If the source says the budget is "in the region of AUD 90k", that's ambiguous and worth clarifying. If the source doesn't mention budget at all, the AI should proceed without a budget claim in the current section rather than fabricate one. Different failure modes, different responses.

What the loop looks like in the SkyDraft workflow

Every section that SkyDraft generates ships with up to a few clarification cards attached beneath it. The card includes the section it originated from, the question, and a suggested answer format. The author can answer in place, dismiss, or defer. Answered clarifications become pinned tiles in the document's notes panel, alongside required-information answers, so they persist across every subsequent generation.

When you regenerate a later section, the AI sees the pinned clarifications as part of its context. It knows that Acme prefers "engagement" over "project" without being reminded. It knows the discovery date is September 12 and the kickoff is September 19. The scaffold of confirmed knowledge grows as the document does.

This is also why we build against a section-by-section generation model rather than a one-shot draft. Clarifications only work if there's a human review gate to surface them at. A one-shot 20-page draft that ends with "and by the way, here are 47 things I wasn't sure about" is worse than the fabrication it's trying to prevent. The gate happens every section, which is where the questions stay small enough to actually answer.

Why this matters for enterprise AI

Enterprise buyers are past the "will AI hallucinate?" question. Everyone knows the answer is yes. The next question is "what do you do about it?" and most vendors answer with post-hoc detection (a second model checking the first), retrieval augmentation, or temperature tuning. All useful, all incremental. None of them address the incentive problem: the model is still being asked to complete rather than to flag.

A workflow-level clarifications loop changes the incentive. The AI is scored not on "how complete does this section read" but on "how well did this section carry forward the answers to previous clarifications and surface any new ones." Under that scoring, the behaviour shifts. Fabrication rates go down not because the model got smarter, but because the surrounding system stopped rewarding fabrication and started rewarding admission.

This is a practical, boring, workflow-engineering answer to a philosophically flashy question. That's usually a good sign: the interesting problems in AI product design are the ones where the fix is orchestration rather than a bigger model. If your team is choosing an AI drafting tool for professional services, the clarifications behaviour is one of the easier things to test in a pilot: give it an intentionally ambiguous source document and see whether the output flags the ambiguity or invents around it.

For related patterns in why generic AI struggles with structured documents, see Proposal writing with AI: 6 patterns that work, 4 that don't and Statement of Work AI drafting: what's actually different from a proposal. For the constraints SkyDraft has committed to (including the specific one about never inventing content when the source is silent), the protocol page spells them out.

Signals to watch for when evaluating tools

If you're shortlisting AI drafting tools for professional services work, the presence of a clarifications behaviour is easy to bluff and hard to actually deliver. A few concrete questions cut through the marketing.

Can you show me a clarification that was asked, answered, and referenced in a later section? If the demo can only produce clarifications inside a single section, the persistent-notes layer is missing.
What's the budget? If every section asks fifteen questions, the tool is unusable. If every section asks zero, the tool isn't triaging and will fabricate quietly.
What happens when I dismiss? The same question returning on the next regeneration is a sign the loop is cosmetic rather than structural.
Can I answer a clarification without regenerating the section? If the answer changes future sections but not the current one, that's the intended behaviour and the mark of a tool that's thought about this.

These aren't exotic questions. They're the questions you'd ask if you were designing the system yourself. Every vendor that has built a real clarifications loop will have crisp answers. Every vendor that hasn't will pivot to talking about their base model or their RAG pipeline.

Where to start

If your team is running an AI drafting tool today, the simplest test of whether it has a real clarifications loop is this: draft a section from an intentionally incomplete source. Something that omits a key fact (budget, deadline, deliverable format). If the output includes a specific claim about that fact, you have a fabrication. If it flags the missing fact and asks, you have a clarifications loop.

SkyDraft pilot workspaces are open. We set up the first template with you on a real document, in a working session. The clarifications loop is on by default; you'll see it working within the first section.