AI research briefing preparation.

What a briefing workflow produces

The output is a short document, usually one or two pages, that answers a defined question: what a prospective client does and what has changed for them, how a sector is regulated, or what is known about a policy issue. Every statement carries a citation to the passage it came from, and anything the model cannot support is dropped. The analyst then adds the part no model should write: what it means for this client and this meeting.

Briefings are high-volume work in consultancies, and they are often built in a hurry the night before. For other uses across the sector, see our guide to AI for professional services.

How the briefing pipeline is specified

Inputs

The briefing question and audience; an approved source list (the client’s annual report and media releases, legislation.govt.nz, Stats NZ tables, sector bodies); relevant past work from your own document store.

Systems

Web search restricted to approved domains, a retrieval index over internal documents, a language model that returns statements with passage-level citations, and the firm’s document management system.

Steps

Turn the question into sub-questions; retrieve candidate passages; discard sources older than the cut-off date; draft statements each tied to a passage; run an automatic check that each citation exists and contains the cited text; assemble the briefing with a source list.

Human review

An analyst opens every citation, confirms it supports the statement, deletes or corrects what does not, and writes the implications section. The partner reads the finished briefing, never the raw draft.

Outputs

A one or two page briefing with numbered citations, a source list with access dates, and a note of open questions the sources could not answer.

Measures

Analyst hours per briefing, share of statements corrected or removed at review, and partner feedback on usefulness after the meeting.

Risks

Citations that do not support the claim, out-of-date sources, overseas material presented as New Zealand fact, and confidential content from another client’s matter.

Illustrative example: a partner’s pre-meeting briefing

Illustrative example: a 15-person Wellington economics and policy consultancy prepares about eight briefings a month before first meetings and scoping calls. Assumptions: each briefing draws on 20 to 30 sources; an analyst’s internal cost is NZ$85 an hour. These are synthetic figures, not client results.

In one synthetic briefing of 36 cited statements, the analyst found 30 supported as written, 4 overstated and corrected, and 2 not supported by the cited source and removed. That ratio is why verification time grows. At eight briefings a month, 2.5 hours saved each is 20 hours, or about NZ$1,700 of analyst time, with a better briefing at the end.

Copyright and privacy rules that apply to briefings

Two New Zealand rules shape what goes into the source pile.

Copyright. Section 43 of the Copyright Act 1994 allows fair dealing for research or private study, but a court weighs factors including whether the work could be bought at an ordinary commercial price and the effect on its market, and the section does not authorise more than one copy on any one occasion. Section 42 covers criticism and review with sufficient acknowledgement. Loading a paid sector report into an index everyone queries may exceed its licence; check the terms, and take advice for anything beyond short extracts.

Privacy. Briefings often profile named people: a client’s chief executive, board members, counterparties. Since 1 May 2026, IPP3A requires reasonable steps to tell people when you collect their information from someone other than them, unless an exception applies. The Privacy Commissioner’s guidance lists publicly available information, such as news, public registers and public web pages, as one exception, but not private social accounts. A detail passed on by a mutual contact is indirect collection under the Privacy Act 2020. The simple rule: keep personal detail to public, professional facts the meeting needs.

Judgement that stays with the analyst and partner

• Framing the question. A badly framed question produces a well-cited answer to the wrong thing.
• Deciding which sources count when they disagree, and saying so in the briefing.
• The implications for this client: risks to raise, questions to ask, what not to say.
• Anything that could identify confidential information from another engagement.

How AI briefings go wrong

Fluent prose hides weak support. The common failures are citations that point to a real page which says something narrower, Australian or United States material presented as New Zealand fact, sources that predate a law change, and snippets from a different client’s matter leaking in through a shared index. Retrieval quality is measurable; our note on RAG retrieval evaluation shows how.

• Every statement has a citation the analyst has opened and read
• Sources are dated, and anything older than the cut-off is flagged
• New Zealand facts come from New Zealand sources
• The internal index respects matter-level access, so other clients’ files never appear
• Personal information about named people is public, professional and necessary

Building it and what it costs

The core is a retrieval system over approved sources, which is RAG development with a strict citation format. Model fees per briefing are small next to analyst time; the real costs are source curation, access controls on internal documents and the review screen. Lucent, our agentic investment research platform used by New Zealand financial institutions, is built on the same rule: outputs grounded in citations.

To try the manual version first, see summarising documents with citations and using AI for meeting preparation. Briefings feed directly into proposal preparation, and actions from the meeting itself into meeting actions into project systems.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.