Contact
HomeBuilding

AI Review Panel. The AI suggests, the analyst decides. A review flow from a regulatory data tool, rebuilt as a pattern.

Direction 2026/14 — Sustainability Ratings Providers

Northbridge Markets Authority · Published 3 Nov 2026

1.This Direction is made by the Northbridge Markets Authority under the powers conferred by section 12 of the Sustainability Ratings Order 2026.

2.It sets out the application periods during which firms must apply to the Authority for permission to provide sustainability ratings. Firms that do not apply within their window may not continue to provide ratings after 30 June 2027.

3.The first application window opens on 1 February 2027 and closes on 30 April 2027. It applies to firms providing ratings to more than 50 professional clients.

4.This Direction comes into force on 1 January 2027. Comments on the accompanying guidance may be sent to the Authority by 15 December 2026; the Direction itself is not subject to consultation.

Live demo. Open a field, hover its evidence, then accept, dismiss or apply all pending.

Try itai-patterns.app/patterns/review/ai-review-panelGitHubgithub.com/carlosmarch/ai-patterns

The problem

Datamaran's regulatory team reads new regulations so clients don't have to. Every regulation that comes in gets classified, dated, scoped and tagged before it reaches the platform.

Most of that was typing. About 55 fields per regulation, entered by hand: issuing body, instrument type, binding force, effective date, consultation deadline, affected companies, topics. The backlog had grown to around 350–400 regulations nobody had processed yet, and every one of them was a regulation clients couldn't search or get alerts for.

A model could fill most of those fields. The question was what happens between the model filling them and a client seeing them.

My role

Senior Product Designer: I designed the UX for Submaran's AI Prefill and Reviewer flows, working with the regulatory team, product and engineering. During the rollout I sorted tester feedback into 35 prioritised items, ran a structured feedback session, and wrote the Slack updates so testers knew what had changed and why.

The rule: the AI suggests, the analyst decides

The goal was to change the team's role from data processors to data validators. That set one rule for the interface: a suggestion is never data until a person says so.

The analyst reads the regulation on one side and reviews the AI's values in a panel on the other. A value is saved only when the analyst accepts it, edits it or dismisses it. That takes a few clicks per field, no more.

The faster option was to auto-publish: let the model write straight to the platform and fix errors later. I chose against it. Later means after a client has read them, and AI accuracy varies a lot by field. A date written in the text is easy to extract. A topic, or whether a document counts as a consultation, is a judgment.

When the analyst is happy with what's left, Apply all pending accepts every pending field at once. It's still the analyst's decision: the button does nothing until they press it, and anything they dismissed stays dismissed. A segmented status bar and the filter tabs show how much is left, so the fields that need judgment can be checked first.

What a reviewer needs to decide

Checking a value is quick if you can see why the model chose it. So every field carries three things:

  1. The value, short and normalised: one date format, enums from the field's allowed list.
  2. The reasoning, one to three sentences. When the model has doubts, it says so plainly ("It may not count as a consultation deadline") instead of hedging every value.
  3. The evidence, as verbatim quotes from the regulation. Hover one and it lights up in the source; click it and the source scrolls to it.

The evidence has to be a verbatim substring of the source. If a quote can't be found in the text, it's dropped. Highlighting the wrong passage is worse than highlighting nothing, because it looks like proof.

Reusing the pattern

Later, the team needed four new lifecycle fields: Consultation Close Date, Transposition Deadline, Current Status and Applicability Date. I designed them with the regulatory team and engineering using the same apply and dismiss pattern, so analysts didn't have to learn a second way of reviewing AI output.

The outcomes

  • AI-prefilled regulations went from 331 a month in July to 821 in October. That's 99.5% of everything ingested. The review flow is what made that volume safe to publish.
  • Across 623 published regulations, analysts applied the AI value unchanged 79–95% of the time on nine core fields, and reviewers corrected only 1–5% of those values.
  • On topics, reviewers still correct about a quarter of regulations. That's the number that justifies the human step. Accuracy wasn't the same everywhere, and the design didn't pretend it was.

Faster processing means fresher data on the platform: better search relevance and earlier alerts for clients.

What I'd do differently

We tracked adoption and accuracy well. We didn't measure how long a review takes, before and after. That's the number that best shows the change from processing data to checking it, and it's the one I can't give.

Next time I'd set a time-per-regulation baseline before launch.

From product to pattern

The panel lives in an internal tool. The judgment behind it doesn't depend on regulations, so I wrote it up as the AI Review Panel in AI Patterns, for any product where people check values a model pulled from a document: contracts, invoices, filings, insurance forms.

Writing it as a pattern meant saying what the panel is not for:

  • Values with no single source passage, like summaries, sentiment or scores. There's nothing to quote, so show the reasoning in a trace instead.
  • One or two fields. An inline suggestion next to the input is lighter than a panel.
  • Fields people edit freely and check later. Without an accountable review step, the per-field ceremony is overhead.

About this demo

The panel at the top works the way it does in Submaran. It's a port of the AI Patterns component, rebuilt on this site's own buttons, icons and tokens. The regulation is fictional, so no client data is shown: a Direction from a made-up markets authority, with eight fields a model might extract from it. Two start accepted and one dismissed, so the status bar and tabs have something to show.

The full UX spec (anatomy, behaviour, content rules, accessibility) is on the pattern page.