The NZ Brush Brain

This working guide is private to NZ Brush and Incredible. Enter the passcode to continue.

That's not it - have another go.

Prepared by Incredible · A working document

Prepared byIncredible
AI Brain · Working session · 12 June 2026

The NZ Brush Brain

The first live run of the "Brain" against real Unleashed data - what we tested together, what held up, what we are tightening, and where it goes next. A working document, updated after every session.

Connect the systems Ask anything Automate the routine Humanise the exception
The shape of it

AI builds the structure; the structure becomes automation

The Brain connects your systems and gives you an open field to ask anything. But the real engine is the loop underneath: every good question the AI answers can graduate into a saved, structured report - fast, near-zero cost, and schedulable. Automate the routine, humanise the exception.

Connected data

One brain across every system

Unleashed is live now, connected through an integration layer we built ourselves - nobody else has this inside Unleashed. Xero is next, behind a hard financial wall. SharePoint and Shopify sit on the radar for later.

Foundation

The AI copilot

Humanise the exception

An open field across all of it. The AI handles the one-offs and the new questions nobody set up a report for - the churn dig, the "what's going on with this customer" - turning unstructured questions into structured queries.

Enabler

Saved reports and automations

Automate the routine

Anything asked regularly graduates from AI to a saved, straight-API report - faster, costing cents not dollars, and schedulable weekly or monthly to managers without anyone prompting a thing.

Routine

Delivery order follows pain and accuracy, not this hierarchy - getting the data dead right comes before everything else.

What we're tackling

Priorities, in order

Ranked by impact against effort, shaped by what Friday's live run taught us. Tap to expand. Nothing locked - we rank these together as the build moves.

1

Get the data dead right

Accuracy before everything - close is not correct
Highest painIn progress

The pain

  • Unleashed pages its data, and the first pass pulled 5 of 29 sales orders - a comparison built on one page is simply wrong.
  • Every sales order carries three dates - order, required, completed - and which one a report uses changes the numbers.
  • NZ timezone handling on date filters needs verifying end to end.

What we're building

  • Every query pulls every page - 12 months means all 12 months.
  • Every result states its source and exactly which date field was used, so you can correct us once, not repeatedly.
  • Default date conventions agreed with Hadleigh, plus NZ time applied to all filters.
"On any sales order report there's three dates - the order date, the required date, the completed date... I've kissed a few of those frogs that you might be kissing through Unleashed data."Hadleigh, on where the gaps come from
2

The report library, then automate it

The reports you already do by hand - plus the churn insight you can't get at all
Highest painBuilds on what exists

The pain

  • Churn insight - which customers, branches or products are trending out against a comparable period - is effectively impossible to pull from Unleashed today.
  • Price review packs are stitched together across spreadsheets and ChatGPT by hand.
  • At-risk customers are spotted by memory and gut feel, not by the data flagging them.

What we're building

  • Customer and sales set: group roll-ups (Fulton Hogan as one, drill into 25 branches), segment reporting, at-risk detection on purchase rhythm, contextualised win-loss shifts, proactive replenishment, ABC customer focus, repeat-pattern recognition for marketing.
  • Pricing set: price review packs across 3-4 sources, cost change triggers, landed cost impact analysis (plastics up 25-35%).
  • Each one validated against live data, then saved as an automation.
"They sound incredibly comprehensive... if it can solve those problems, it's probably powerful enough to do anything else we need it to do."Hadleigh, on the report list
3

Xero, behind a hard wall

Financial data for two or three people only - everyone else stays operational
High painClear path

The pain

  • Operational data is open to everyone in Unleashed by culture - but the financial picture must stay completely separate.
  • Managers can't go into Xero themselves, so financial movement reaches them slowly or not at all.

What we're building

  • An authentication link to you first, confirming we have the right permission scope.
  • Xero connection gated by Xero's own login: everyone sees the option, only the two or three credential-holders can connect, and what each sees mirrors their Xero permissions exactly.
  • Later: write-back - "make that, add that" - so updates flow into Xero without opening it.
"It's probably just to have a dark wall there where you're either in or you're out with Xero... we'll keep that to the two, three people accessing it."Hadleigh, on the financial wall
4

Dashboards on a schedule

Weekly or monthly KPI packs to managers - and pre-visit summaries for reps
High painEasy once reports exist

The pain

  • Key managers don't get organised cross-system metrics; insight depends on someone pulling and consolidating reports.
  • Reps walk into customer visits without a consolidated picture - open orders, contacts, thresholds live across three or four pages.
  • Operational facts (like the freight threshold moving from $300 to $500) live in heads, not in front of newer staff.

What we're building

  • A standing weekly or monthly dashboard to key managers across Unleashed and Xero - the system telling the manager, not the manager digging.
  • One-tap pre-visit customer summaries for reps.
  • Shaped by the example prompts and dashboard concepts Hadleigh is sending through.
"It would be quite good to have some form of standing weekly or monthly dashboard report that goes out to key managers... it's sort of Xero telling the manager that your travel expenses are up 3% month on month."Hadleigh, on scheduled reporting
5

Cost visibility and the right model

Know what each query costs before it becomes a habit
Medium painQuick to build

The pain

  • One heavy multi-branch query cost roughly US$3; the session used about US$7 in total.
  • A few dollars per query, multiplied by 50 people doing it daily, becomes real money fast.

What we're building

  • Per-query cost shown in the interface, with usage logging behind it.
  • "Save this as a report?" prompts so repeated questions stop burning AI dollars.
  • An agreed position on the speed-accuracy-cost matrix - a newer Claude model exists at roughly twice the price, and "always use the most powerful" has a bill attached.
"If we're going to do a complex prompt that's going to do a lot of work... and it'll cost us six or seven bucks, then we can make that decision to go: that's worth it."Hadleigh, on informed spending
+

On the radar, not yet

Parked deliberately, not forgotten
Future

Xero write-back ("make that, add that"), SharePoint and Shopify as additional data sources, and branch-level segmentation - the branch list in the sidebar is global-only today, and you told us you prefer global anyway. Plus whatever surfaces once the wider team starts playing with it.

How we prioritise

Pain against ease of build

Accuracy sits first because everything else inherits it - a churn report built on one page of data is worse than no report. The report library follows immediately because the reports already exist as manual pain, so the value is proven before we write a line.

Pain / impact →
1Data accuracy
2Report library
3Xero wall
4Scheduled dashboards
5Cost visibility
+Future sources
Ease of build →
Data and reports Access and finance Delivery and cost Future / parked
Hadleigh's query wishlist

Hypothetical scenarios

Every query and use case raised across the discovery calls (14 April, 6 May and 28 May) - the test we hold the Brain to. Have a read through, rank what matters most, and where a scenario would benefit from sample or test data, send it through against that item - real data is what turns a hypothetical into a validated report.

Customer and sales analysis

→
"Tell me about Fulton Hogan this year versus last year."Year-on-year comparison is a manual conflagration of reports today - Unleashed snapshots, it never compares.
→
National group roll-ups - all 20+ Fulton Hogan branches (or PGG, Intergroup) in one national view.The dashboard caps at 10 customers per drill-down; the Fulton Hogan supplier agreement reporting is a huge manual export job.
→
Segment and customer-group reporting at the same time - the roading sector as a whole, plus Fulton Hogan as a group.The customer-type field forces a choice of one or the other today.
→
At-risk account detection - flag a customer falling off its normal purchase rhythm.The hibernated-cold-dead behaviour you liked in Prospect CRM; sharpest on the hygiene side, where "you often don't know until it's too late".
→
Contextualised win-loss shifts - Fulton Hogan down 20% may just mean a contract moved to Downer, who will be up.The reporting needs to understand contract movement within the roading sector, not panic at one number.
→
Proactive replenishment prompts - "this customer should have ordered by now, send them a cut-off reminder."Turns purchase rhythm into action, not just a report.
→
ABC and ABCD customer and product tiering - "who are our A customers and A orders right now?"Currently external knowledge living in spreadsheets, strategy documents and the State of the Nation - not in Unleashed.
→
Repeat-pattern recognition for marketing - spot "a problem we solve regularly" and surface it.So the message can be amplified commercially.

Pricing

→
Price review packs combining multiple sources - a new price file plus customer sales history from three or four databases, into one customer-facing report ahead of an increase.The real example: an hour spent in ChatGPT combining spreadsheets.
→
Cost-change triggers - a cost inflates or deflates, a price review fires.Live example: a $60,000 purchase order up 8.5% overall but 5-15% line by line, needing impact analysis on end pricing.
→
Landed cost impact analysis - fluctuating freight and raw materials, and what they mean for sell prices.Plastics up 25-35% right now; customs especially.
→
Queries that understand the full pricing stack - price tiers, quantity breaks, customer special pricing, and spot discounts at sales-order or line level.Half-aware pricing answers are worse than none.
→
Draft the current pricing policy from accumulated context.Something you explicitly said you'd like the AI to write.

Inventory, purchasing and production

→
Global replenishment recommendations - "here's what to buy globally for the company; I don't care where you land it, you need 50 in stock or you'll run out on that supplier's lead time."Advanced Inventory Manager only works branch by branch.
→
Stock-out and overstock flags across the lot - risks, opportunities, overstock, understock."A smart guy to tap you on the shoulder and say I think we're going to run out of banister brushes."
→
Simple cross-tool stock queries - "what's the stock of road brooms?"The original copilot example, and still the everyday baseline.
→
Critical lines analysis - which lines matter most as working capital moves to importations with extended timelines and costs.Currently generating 10-15 email threads a day.
→
Stock-vs-cash equilibrium - never running out, against tying up working capital.Across 5,094 SKUs, roughly 200 suppliers and four warehouses, with a just-in-time / just-in-case strategy that shifts with market conditions.
→
Reorder by supplier and container fill - reorder reporting grouped to fill containers, including consolidating agents.Buying decisions follow the container, not just the SKU.
→
Production planning - what's loaded on the machines, what's next, "these orders weren't due until next week - should we put them in while the machine's loaded?"Replacing Brett's offline spreadsheet ("Brett's brain in a box"), with offline overrides feeding back so the Brain learns.
→
3PL and export decision support - cost to deliver, stock placement and market insight.Behind the Australia digital pioneering plans.

Finance (Xero)

→
Cash flow forecasting.Today: a daily spreadsheet plus a monthly rolling one, fully offline - "any improvement there would be amazing".
→
One question, both systems - "tell me the combined information from both of those servers."Cross-system queries spanning Unleashed and Xero without asking each separately.
→
Ledger-based analysis on the newly renovated ledger codes.Income and segment insight the old codes couldn't give.
→
The wall holds on every financial query - only Hadleigh, Ang and Paula hold Xero licences.Everyone else sees only what's already synced into Unleashed.

Policies, HR and the niggly questions

→
"How do I fix a windscreen chip on my car?"The adoption test, in your words: answering that "would have saved about 17 phone calls in the last week".
→
"How much can I spend on a steak when I'm travelling for dinner on the company credit card?"Plus engaging insurance, reporting injuries, equipment use, and "what's the pricing policy?"
→
Answers drawn from the right sources - employee handbook, health and safety manual, MPI biosecurity compliance, pricing and ICP policies.A curated knowledge layer, not the whole drive.
→
Hard rules versus interpretable guidance, kept distinct."Our policy is you don't kill someone, and our guidance is this is probably how you should avoid killing someone."
→
Covering sales support absences - everything that surfaces "if Paula takes the afternoon off and someone else sits on the desk".Currently a source of mistakes and interpersonal conflict.

Email and communications (future phase)

→
Emails organised by customer rather than by thread, with attachments read and order statuses updated.The Norway distribution example - freight forwarder status PDFs flowing straight into order status.
→
80% draft replies to common inquiries, pulled from the Brain.Humans finish the last 20%.
→
"Can all emails come through here? Is it starting to learn how we act and respond to inquiries?"Your direct ask - the long-game version of the Brain.

Managing the Brain itself

→
Natural-language correction and configuration - "hey dude, you're getting it a bit wrong, what we want you to focus on is XYZ."You asked specifically whether you could just chat to it, rather than typing rules or uploading documents.
→
Version control on Brain rules - publish a change org-wide, or sandbox it first.Tuning the Brain shouldn't mean breaking it for everyone at once.

Cross-cutting constraints - true of every scenario above

Every query respects Unleashed role permissions, mirrored per user by email address (role is the fourth column in the users list; deleted users drop off). Xero stays restricted to the three licence holders - everyone else sees only Xero data already synced into Unleashed. OneDrive access is limited to a curated "discoverable" folder of authorised information, not the full drive ("a lot of noise in that folder... a can of worms"). Crystal Payroll is out of scope - payroll top-line arrives via Xero. Engage Solutions is likely API-limited or without an API; we'll investigate with low expectations. And the Unleashed API allows 500,000 calls per month - usage monitored.

Who connects to whom

The team, and how insight flows

Today, cross-system insight runs through one person exporting, consolidating and reconciling reports by hand. The Brain changes the shape of that. Toggle to see the shift.

Asks / waits for an answer Holds knowledge that feeds the Brain

From question to automation - the Brain loop

01

Ask

Anyone

A plain-language question in the copilot - no report builder, no exports.

02

Plan

The Brain

AI turns the question into a structured query - which system, which fields, which dates.

03

Pull

Unleashed / Xero

Straight API calls - every page, NZ time, source and date field stated.

04

Validate

Hadleigh + Incredible

Checked against live Unleashed reports until exactly right, not just close.

05

Save

The Brain

A good query graduates into a saved report - structured in, structured out, near-zero cost.

06

Schedule

Managers

Routine reports land weekly or monthly without anyone prompting.

The people

Hadleigh

NZ Brush - systems and reporting lead

The validation engine of this build. Knows the frogs - the three dates, the naming quirks, the report nuances - and is feeding the Brain prompts, live reports to validate against, and the users and roles export.

Sales reps

Consumers - on the road

"Give me a summary of this customer before I go in and see them" - open orders, contacts, freight thresholds, all consolidated before the visit.

Branch managers

Consumers - scheduled

Standing weekly or monthly dashboards on key metrics and KPIs across Unleashed and Xero - the system telling them, not them digging.

Finance (2-3 people)

Gated - behind the wall

The only Xero-credentialed users. What each sees in the Brain mirrors their Xero permissions exactly - in or out, no half measures.

Warehouse team

Context - statuses

The Parked, RTG and Placed flow they run is exactly what the Brain now understands when it reports what's live, what's picking and what's not to be touched.

Incredible

Andrés and Matt - build and tuning

Andrés leads the build; together we tune accuracy, wire the integrations and turn the validated queries into automations.

Where the data lives

The systems, and how they fit

Unleashed is home - all operational truth lives there and the whole company can see it. Everything else either feeds the Brain or stays deliberately walled.

An Unleashed brain nobody else has

Unleashed offers no off-the-shelf AI connection, so we built our own integration layer over its API - and it's live now, querying your real data across every dataset in plain language. That layer is the moat this whole build stands on.

Core system · Home · Live

Unleashed

Inventory, sales orders, customers, products and salespeople across every branch. Open to the whole team by culture - visibility is fine, bulk functionality is what eventually wants configuring.

Feeding the Brain now, through our own layer
Finance · Next

Xero

Invoices, bills, expenses - the financial picture, kept completely separate from operations and seen by two or three people only.

Login-gated; authentication link coming to confirm permissions
AI engine

Claude

Powers the copilot. Currently running a deliberately cost-controlled model while we tune accuracy; a newer model exists at roughly twice the price.

Speed-accuracy-cost call to make together
Manual · Legacy

ChatGPT + spreadsheets

How price review packs get stitched together today - new price file plus customer sales history, combined by hand.

Retires into the Brain's pricing reports
Documents · Radar

SharePoint

Document storage raised as a future source.

On the radar, not yet
E-commerce · Radar

Shopify

Online sales channel raised as a future source.

On the radar, not yet
What we still need to know

Open questions

A few answers sharpen the build considerably. We work through these together - some are yours, some are ours to chase.

Data and conventions

?
Which date field should each report default to - completed, order or required?The Waikato gap traced mostly to date choice; agreed defaults stop us relearning frogs you've already kissed.
?
What naming conventions do we lock in for customer matching - Fulton Hogan and friends?Consistent matching is cheaper and more accurate than fuzzy AI search; you're happy gardening data, so let's agree the standard.
?
Are sales quotes available to the Brain yet, or unconfigured at our end?It returned nothing on the call - we're checking which side the gap sits on.

Access and permissions

?
Does the Unleashed users and roles export give us enough to mirror who can do what?The API exposes salespeople, not user permissions - the export is the interim source of truth.
?
Is a duplicated users table acceptable until Unleashed exposes permissions properly?Two copies can drift out of sync; you said the workaround is fine, but we want eyes open on the trade-off.

Cost and models

?
What's the appetite on the speed-accuracy-cost matrix?The newest Claude model is roughly twice the price; "go the most powerful" is a fair instinct with a bill attached - worth a deliberate call.
?
Which queries from the list are BAU automations, and which stay ad hoc AI?Routine on automation costs cents; AI is for genuinely new questions. The split decides where the money goes.

Scale

?
Where exactly is the processing limit on heavy multi-year collations - and do we stagger?The Fulton Hogan run hit a ceiling; we need it characterised before the big comparisons go to the team.
What we're banking on

Working assumptions

These shape the build. The first one is the big lever - and unlike most assumptions, it's been dug into, not guessed.

Unleashed exposes no AI connection and no users-permissions endpoint

Confirmed by digging, not assumed. We've built our own integration layer over the Unleashed API - live and querying real data now. Permissions ride on Hadleigh's users and roles export until Unleashed offers something better; salespeople data (archived flag, territories, email) is not the same thing as user permissions.

Xero access mirrors Xero's own permissions

Login-gated, in-or-out. No separate permission table to maintain on the financial side - if you can't see PAYE in Xero, you can't see it in the Brain.

Confirmed

Date filters run on New Zealand time

We believe Unleashed applies +12 to GMT at your end; we're verifying NZ time handling end to end on every date filter before any report is signed off.

Must verify

Every query pulls every page

Unleashed pages data in 50-record chunks. Partial pulls produced wrong comparisons on the call, so full pagination is now the rule - 12 months means all of it.

Confirmed fix

Heavy queries cost roughly US$3 each today

About US$7 across the session, and barely a blip on the Unleashed API dashboard. Saved automations bring routine cost to near zero - that's the whole loop.

High confidence

Current model is a cost-controlled Claude Sonnet

One generation back, chosen deliberately to hold costs while we tune accuracy. Upgrading is a dial we turn together, not a default.

Confirmed

Global beats branch-level, for now

The branch list in the sidebar is cosmetic today - all queries run global, which is your stated preference. Branch segmentation stays parked until you want it.

Confirmed
Where it goes from here

Next steps, both sides

You said it best: you're the customer and the information providers - we'll demand what we need, you sing out when something doesn't make sense. Here's the demand list.

Incredible

  • Send the full report and query list through, so you can validate the highest-impact ones.
  • Send the Xero authentication link (next week) to confirm permissions, then begin the Xero connection.
  • Make every result state its source and date field; fix full pagination, the 12-months-means-12-months behaviour and NZ time handling.
  • Start converting the regular reports into straight-API automations, with per-query cost visibility in the interface.
  • Characterise the processing limit on heavy collations, and check the sales quotes data source.

NZ Brush

  • Export users and their roles from Unleashed - exactly as it comes out, no amending.
  • Send example report prompts and dashboard concepts across sales, finance and KPI metrics.
  • For the highest-impact items on the list, pull the matching live Unleashed report so we can validate against it.
  • Keep flagging the frogs - date fields, naming quirks, anything that doesn't line up.

The loop, until next Friday

  • The homework: each time a report takes too long, or a question can't be answered, jot it down - "would the Brain have done this?" Those examples decide what we automate first.
  • Validation beats vibes: "pretty close" doesn't ship - every report gets checked against your live data before anyone relies on it.
  • Next call: same time next week - Friday 19 June, 8.00am NZT. We'll have Xero moving and the accuracy fixes in by then.