You drive. Yig drafts. — A HITL design principle

Most products that put “AI” on the homepage promise to replace something — a workflow, a role, a deliverable. The pitch lands well in a board deck. It rarely lands well in a finance team that has signed an audit opinion.

Yig takes the opposite stance.

You drive. Yig drafts.

The agent prepares the manual half of close, forecast, and budget. Your team reviews and ships. Nothing leaves the review gate without a human on the other side.

This is not a marketing line. It is a design principle that constrains every component of the system.

The three-step pattern

Every Yig workflow follows the same shape:

1. Yig drafts. The agent reads the data already in your stack — TB, GL, IC pairs, prior cycle templates, source documents — and produces a draft. The draft is structured (table, schedule, narrative paragraph, journal entry) so it can be reviewed quickly.
2. You review. The draft is presented inside the surface where your team already works: Slack, Excel, Teams, a Word doc, a CLI session. Reviewers comment, edit, accept, or reject — line by line. Every action is logged with reviewer, timestamp, and diff.
3. You ship. Approved outputs land where they were always going to land — your close folder, your board deck, your audit binder. Yig is not the final destination of any artefact. It is a contributor to your existing destination.

Why we reject “AI that automates”

There are three reasons we will not build “press button → close ships.”

• Audit reality. Auditors do not sign off on outputs that no human reviewed. SOX 404, ISA 315, and every internal control framework we have read presume a human gate. Removing the gate does not save time; it relocates the time to the audit response.
• Edge cases are the work. In a clean month, automation looks magical. In a real month, the work is the edge cases — the one-time settlement, the FX assumption that just went stale, the entity that re-organised. An automation that hides edge cases produces wrong answers faster.
• Mastery transfer. The reviewer who sees what Yig drafted, why, and where it hesitated, learns something each cycle. The reviewer who only sees the final number learns nothing. Over a year, the team that reviewed gets faster. The team that automated gets dependent.

The yellow-flag pattern

When Yig encounters a judgement call it cannot resolve from the data alone — a one-time item, a forward-looking assumption, a policy choice — it does not guess. It marks the section, states what it does not know, and waits.

Yig flagged this line because the variance is larger than the line’s 12-month volatility. Human judgement needed: is this a one-time item?

The reviewer answers. Yig regenerates the surrounding paragraph with the human input woven in. The output reflects the reviewer.

This is the moment that defines the product. Show the AI saying “I do not know, you decide” and the trust posture is immediately legible. We would rather be slightly slower and unmistakably honest than fast and impossible to defend.

What HITL looks like at each surface

Surface	Draft	Review gate	Ship
Slack	Inline message + linked draft note	Reply-thread comments per row	Post to workbook
Excel add-in	Side-pane draft	Per-row accept / reject, yellow flags	Stage to sheet
Word add-in	Inline paragraphs, yellow blocks	Click each block, edit, resolve	Final draft
CLI	Diff view	approve / reject with comment	Stage to file

Different chrome, same pattern. The pattern is the product.