Flint Group Built by NBH

Vision Document

The leading residential mortgage brokerage in Australia.

How an AI agent, built on the systems Flint already runs, takes the admin layer off loan writing, so your team writes more loans and every client feels looked after.

Prepared for Ben Robinson & Redom Syed
Prepared by NBH  ·  Trav White, Micah Howard
Scope Loan writing  ·  Documents received to application approved

Read this as a vision, not a delivery. Everything here is a demonstration of what is possible on your existing technology. Nothing in this document is built or live for Flint yet. Where a figure is an estimate rather than a measured number, it is labelled as illustrative.

02 / 08

A brokerage at scale, carrying an admin layer it should not have to.

Flint is one of Australia's most awarded residential brokerages: Best Residential Broker in Australia four years running, Broker of the Year three. The loan book is deep. The problem is not demand. It is the weight of manual admin wrapped around every loan.

~800
live loan applications in process right now, across the application pipeline
2,150
settled loans sitting in the post-settlement back-book, largely un-worked
~30
people across App Docs and Loan Processing, largely offshore
~$20k/mo
cost of that processing layer, roughly $240k a year

Where the hours actually go

The real load is not the loan count. It is the correspondence and the checklist work stacked on every deal. Reading lender and client emails, drafting replies, chasing documents, ticking off tasks. Measured across the live deals in your own HubSpot, a single loan carries between 60 and 110 emails and 20 to 64 tasks depending on its stage. Settlement prep is the heaviest.

Multiply that across roughly 800 live deals and you are looking at well over 50,000 emails and 20,000 tasks of in-flight manual admin at any one time. That is what the Loan Processing and Data Entry layer does all day. Every piece of it lives in HubSpot, and every piece of it is reachable through HubSpot's API.

The judgment that wins loans is human. The admin around it does not have to be.

03 / 08

What you saw in the demo.

Two surfaces, one workflow. A Flint-branded portal where the client uploads documents and the AI reads them live, and the broker's view inside HubSpot where everything outstanding, and the servicing numbers, appear without anyone keying them.

What the client sees

A premium document portal, Flint branded

The client opens a clean, warm, Flint-branded portal. Encrypted, Australian hosted. No printing, no email ping-pong, no re-sending the same payslip three times.

  • A clear checklist of what is needed and what is already in
  • Drag a document in and it is recognised on sight
  • Fields stream in with confidence meters as the AI reads
  • The checklist ticks itself; the client sees real progress
What the broker sees

The deal in HubSpot, already assembled

On the deal record, a Flint card surfaces the state of the file. The processor would normally build this by hand. Here it is simply present and current.

  • Document readiness, 9 of 12 received, with what is still outstanding
  • Applicant snapshot, including self-employed structures, company, trust, PAYG
  • The servicing add-back and total adjusted income, surfaced not spreadsheeted
  • Lender back-channel status and a settlement countdown

The live moment

The part worth pausing on: when a company financials PDF was dropped into the portal, it was a genuine live read, not a canned animation. The document was recognised, the fields streamed in, and the serviceability add-back, the kind your processors normally extract by hand from messy accountant-prepared statements, was pulled out and surfaced in seconds.

That is the whole thesis in one gesture. The document is read once, accurately, and the result flows into the systems your team already works in. Nobody re-types it.

04 / 08

Four outcomes, in the order that matters.

The point of all this is not the technology. It is what it does for the client first, and the business right after. Here is the order we would build toward, and the systems each outcome touches.

01
Customer experience

Clients who feel genuinely looked after

This comes first on purpose. A premium client portal replaces the document scramble: clear, branded, transparent, the same warm feel a Flint client expects in person. Approvals come back faster because the file moves the moment a document lands. Clients are kept proactively informed instead of having to chase, because the agent drafts the update the instant a lender status changes. The relationship your brokers are known for is protected and extended, not replaced.

Built on Client portal HubSpot comms AFG back-channel
02
Operational efficiency

The admin layer, automated

The Loan Processing and Data Entry work is the target. Documents are read on arrival. Lender emails from the AFG back-channel are triaged, summarised, and turned into stage moves and client updates. Checklist tasks close themselves as their conditions are met. The post-settlement back-book, today largely un-serviced, gets worked systematically. The aim, in Redom's framing, is to automate the bulk and retain your best people for client-facing roles, not to thin the team for its own sake. Section 5 maps this plainly: what comes out of the stack, what could come out over time, and the spine that stays and gets stronger.

Built on HubSpot Middle Broker Engine / AFG Google Suite
03
Revenue

More loans, from the same people

When the admin moat clears, credit analysts spend their time on the judgment work that actually writes loans rather than chasing documents. The 2,150-loan back-book becomes a reactivation and repricing engine instead of a dormant list. Faster, cleaner settlements mean more volume through the same pipeline, and a connected view of each client opens genuine cross-sell. Growth stops being gated by headcount.

Built on HubSpot pipelines Back-book data CoreLogic / Cotality Quickli
04
Speed

An agent working the systems in real time

Underneath the first three is one capability: an agent that works across your systems continuously, not in a nightly batch. A lender email arrives and the file moves within minutes. A document lands and the checklist updates while the client is still on the page. Speed is not a separate feature; it is what makes the other three outcomes felt by the client and the broker rather than reported in a monthly summary.

Built on HubSpot API Event triggers Aircall
05 / 08

What comes out of your stack.

The reason to do this is felt by the client first. When the manual admin layer comes out of the loan-writing process, the file moves the moment a document lands, and every client feels looked after instead of chased. So the question is not just what the agent adds. It is what comes out of the stack you run today. Three honest groups: what comes out, what could come out over time, and the spine that stays and gets stronger.

01 Comes out

The manual processing layer is the explicit goal. This is the work the agent is built to take off the loan, the roles Flint wants to replace, the keying and chasing that the client never sees but always feels.

Loan Processing & Data EntryComes out

The App Docs and LPS team, roughly 30 people and about $20k a month, around $240k a year, largely offshore. The layer the agent is built to replace.

The agent takes over the processing this team does by hand, so the people who stay move to client-facing and judgment work.
Manual keying into Broker EngineComes out

The re-typing left over after the existing Middle-to-Broker-Engine feed. The feed already pushes some data across; people still key the rest into the application by hand.

The agent targets exactly that residual keying, extending the feed that already runs rather than rebuilding it from scratch.
Document chasing & readingComes out

Chasing the same payslip three times, then reading messy accountant statements by hand to pull a servicing add-back. Slow, error-prone, and felt by the client as friction.

The agent reads each document on arrival, extracts the figures with every field traceable to its source line, and ticks the checklist itself.
The email & task adminComes out

The 60 to 110 emails and 20 to 64 tasks of manual admin wrapped around a single loan. Triaging lender mail, drafting replies, ticking conditions off one at a time.

The agent triages the lender back-channel, drafts the client update, and closes each task as its condition is met, so the file moves without a person pushing it.
02 Could come out, Flint's call

These are options Flint chooses, not things ripped out. The current position is to keep them for now and decide after seeing the demo work. Nothing here moves without a deliberate decision.

Middle borrower intakeCould come out

Middle (an MA Financial product) is where the file starts today. The Flint-branded document portal shown in the demo could take its place. The plan is to keep Middle for now and decide once the portal is in front of clients.

Flint's call. If the portal earns it, intake moves there; if not, the agent reads from Middle through a purpose-built feed instead.
Manual serviceability runsCould come out

Serviceability is keyed into Quickli by hand today. A sanctioned one-way path exists to create a scenario from the CRM; reading the result back is the gated unlock.

Could move to an agent-run Quickli check once that read path is open. A commercial gate, not an engineering one.
Lender back-channel chasingCould come out

The manual emails to AFG and ANZ to ask where an application sits. Structured, parseable notifications already arrive from the AFG online back-channel.

Could become an agent-tracked status: the back-channel is read on arrival, the deal stage moves, and the client is told before they ask.
03 Stays, and gets stronger

The spine of the stack stays. The point is not a smaller toolset for its own sake. It is a leaner stack where the human work moves up to judgment, not keying.

HubSpotOrchestration layer, stays

Your CRM, pipelines, stages, comms and tasks. Roughly 95 staff and 54,000 contacts, on a full, mature API. This stays the agent's home base, and every other system feeds in and out of it.

Gets stronger. The agent reads and writes deals, moves stages, drafts emails and closes tasks here directly, so HubSpot becomes the one current view of every loan.
Broker Engine / AFGStays

The lender system of record and lodgement path. AFG owns Broker Engine; the partner API is the gated long-pole. This stays as the application's home, with the agent feeding it.

Gets stronger. Instead of people keying into it, the agent feeds clean, read-once data in, extending the Middle-to-Broker-Engine feed already running.
Credit analyst judgmentStays human

Credit assessment, lodgement strategy and the call on a tricky file. The judgment that actually wins loans, and the reason a Flint client picks Flint.

Gets stronger. With the admin moat cleared, your best people spend their time on judgment, not chasing documents. The agent never makes the credit call.
The net effect

A leaner stack, where the human work moves up to judgment, not keying.

A document arrives; the agent reads it once and pushes the structured data toward Broker Engine. A lender status lands in the AFG back-channel; the agent updates the HubSpot deal, drafts the client update, and closes the task it satisfied. The manual processing layer comes out. The judgment-heavy steps stay human. What is left is a smaller stack, doing more, with your people aimed at the work that writes loans.

06 / 08

How it sequences, and why this order.

The instinct is to start with the hardest integration, Broker Engine data entry. The research says invert it. The highest-value work is also the most buildable, and the gated unlocks have commercial lead times, not just engineering ones, so those conversations start now in parallel.

Phase 1
Buildable now No external unlock needed

HubSpot-native: the admin around the loan

Lender-email triage and summarisation, auto-completion of checklist tasks, deal-stage moves driven by the AFG back-channel emails, client-comms drafting, and the Sales-to-Application handoff. This is where most of the Loan Processing hours actually live, and it needs nothing more than the HubSpot access we already hold. We also build the document system into HubSpot and summarise interactions to assist broker handover notes.

HubSpot API AFG back-channel emails Document reading
Phase 2
Commercially gated Start the conversations now

The gated unlocks, chased in parallel

The AFG partner API is the single biggest unlock for true data entry into Broker Engine. CoreLogic / Cotality licensing powers post-settlement equity work on the back-book. Plaud, for in-person call capture, is waitlist-gated. None of these are blocked on our engineering; they are blocked on commercial agreements and partner timelines, which is exactly why they start on day one rather than after Phase 1 ships.

AFG partner API CoreLogic licensing Plaud beta access
Phase 3
Fallback only Where no API materialises

Browser-automation fallbacks, used sparingly

For any surface where no API arrives in time, Middle document retrieval and reading Quickli serviceability results back, the agent can fall back to working the browser the way a person would. This is the last resort, not the plan: it is more brittle and carries terms-of-service considerations, so we use it only where the cleaner paths above do not land, and we replace it as soon as they do.

Middle retrieval Quickli read-back
07 / 08

Built with the guardrails a brokerage needs.

An agent touching client files and lender comms in a regulated industry has to be safe by design. These are not afterthoughts bolted on; they are how the thing is built from the start.

Control

Human kill-switch

Every automated path can be paused instantly, by stage, by client, or entirely. Nothing runs that a person cannot stop in one action.

Data

Onshore and encrypted

Client data stays Australian-hosted and encrypted in transit and at rest. The portal your clients see is built to that standard from day one.

Compliance

Judgment stays human

Credit assessment, lodgement and strategy remain with your people. The agent clears the admin around those decisions; it does not make them.

Trust

Transparent and traceable

Every extracted field is traceable to its source line. Every drafted client email is reviewable before it sends. Nothing happens in a black box.

What the agent will never do

  • Make a credit decision, lodge an application, or set a client's strategy on its own.
  • Send a client communication that a Flint broker has not had the chance to review and approve.
  • Move client data off Australian soil or store it outside the controls Flint signs off on.
  • Fire HubSpot automations blindly: existing stage-triggered workflows get audited and reconciled before go-live, so no client is double-messaged.
The ambition

From the most awarded brokerage to the leading one.

Flint already wins the awards. The opportunity in front of you is to win the operating model: a brokerage where the admin layer is automated, every client gets a premium experience, the back-book works for you, and your best people spend their time writing loans and looking after clients. That is what turns a great brokerage into the leading one. We have shown you it is possible. Building it for real is the next conversation.

To build it for real, here is what we need from Flint

The process Loom from Luka and Will
A walkthrough of the full loan-writing process and one real application, end to end, so we map the agent to how Flint actually works rather than how we assume it does.
Luka · Will
An intro to AFG
A warm intro to your AFG BDM so we can resume the partner-API conversation already started in the earlier integration effort. AFG moves slowly, so this is the long pole; it starts now.
Via Ben
A scoping session with Middle
Middle's team has already said they are willing to build the API capability for Flint's workflow. We book that session and scope the Middle-to-HubSpot feed so document intake stops being manual.
Via Flint
Prepared by NBH
For Ben Robinson & Redom Syed · Flint Group
Sign-off: Redom