Pitch Guide — Job Spine, Estimating First

Prerequisites

Settle before the first proposal

Open Question Q2 — corpus ownership. Who owns the Cost Memory corpus if this later sells to the client's competitors. It will come up in the room, and an improvised answer loses the deal.

Open Question Q6 — retention-regime scope. Get a written opinion before any compliance claim about retentions appears in client-facing material.

Architecture, decision records and phase gates live in the companion document, job-spine-concept-blueprint.md v0.1. This guide covers the commercial approach only.

What you are actually selling

Not software. In the first meeting you are selling one number the owner does not currently know: what their last two years of jobs actually cost, by trade stage, against what they quoted.

That number is the whole pitch, for three reasons. It is theirs, so it cannot be argued with. It is almost always uncomfortable, so it creates urgency without you manufacturing any. And producing it requires building exactly the data spine the product needs — so the first sale and the first build step are the same work.

Everything else — the agents, the variation workflow, the claims automation — follows from owning that corpus. Do not lead with it. An owner who has just seen that villa renovations lose 4% while bathrooms make 19% will ask you what to do about it. Let them ask.

The sequence, in one line

Mine the history they already have → show them the money they're losing → use that corpus to price the next job → keep the loop closed so it compounds.

What you are not selling

Say these out loud early. Each one buys credibility that a feature list costs you.

Not a takeoff tool. Buildxact already does that, does it well, and costs NZ$199–599/month. If takeoff from clean plans is their bottleneck, tell them to buy Buildxact and save your fee.
Not a replacement for Xero, or for their job management tool in the first phase.
Not autonomous. Nothing goes to a client or moves money without a named human pressing the button.
Not a compliance guarantee.

Who is in the room

Stakeholder map — first meeting
Person	What they fear	What wins them
Owner / MD the buyer	Losing money on jobs they thought were profitable; the business being unsellable because it lives in three heads	The number in §01. Then: the estimator being off sick no longer stops the company quoting
Estimator / QS the gatekeeper	Being replaced, or being shown to have been wrong	Authorship. The labour norms are their norms, extracted and made durable. Their name goes on the audit report
Office / accounts manager the enabler	More systems, more double entry	Nothing changes for them in Phase 0. In Phase 3, supplier invoices code themselves
Site lead / foreperson not in this room yet	Surveillance, more paperwork	Deliberately excluded until Phase 2. Do not mention voice capture in meeting one — it sounds like monitoring before it sounds like relief

The decisive person

The estimator decides whether this succeeds. If they are also the owner — common at this size — the pitch is easier to land and harder to deliver, because there is no one to delegate the behaviour change to. Note it and price the implementation accordingly.

The narrative — five beats

Rehearse these as five sentences before you build any slide.

Your estimating is fine. Your memory isn't. Every quote you write draws on what one or two people remember about jobs you finished two years ago. Nothing you own turns those finished jobs into a price for the next one.
The market can't fix this for you. AI plan-reading got very good in 2026 — 70–90% faster takeoff, above 95% accuracy on clean plan sets. It also stops working exactly where renovation lives: ambiguous scope, and cost hidden behind existing linings. Your risk isn't in the drawing. It's in the seven similar villas you've already done.
Which means the only useful data is yours. Nobody can buy your realised labour rates on scrim-lined ceilings, or your actual overrun distribution on wet-area relocations. That's a moat, and right now it's sitting in Xero and a shoebox of invoices doing nothing.
The enterprise end already proved the architecture. Procore spent 2026 building a single trusted data environment specifically so AI agents could work on top of it — and cut off a third-party agent vendor's API access first, then bought their own. Whoever owns the data controls which agents exist. At your size nobody has built this. That's the opening.
So we start with a report, not software. Give us 24 months of your finished jobs. We'll tell you what they actually cost by stage against what you quoted. If that report isn't worth what you paid for it, we stop there.

Why it's ordered this way

Beat 5 is the ask. Everything before it exists to make beat 5 feel small.

The 45-minute first meeting

Running order
Min	What happens	Why
0–5	No slides. Ask: "How many detailed quotes did you write last year, and roughly how many did you win?" Then: "Of the ones you won, do you know which ones actually made money?"	The second question is the whole pitch. Most owners answer with a feeling, not a number. Let the silence sit
5–15	The diagnostic questions (§05). Write the answers down visibly	You are building their baseline in front of them. This is also your scoping
15–25	Beats 1–4. Two slides maximum. Show the verified 2026 finding that AI estimating fails on hidden-scope remodel work	Credibility comes from the limitation, not the capability. You become the person who told them what doesn't work
25–35	The proof artifact (§06) — the back-cost of one of their own finished jobs	Concrete beats abstract, always
35–42	The audit proposal: fixed fee, ~3 weeks, ~6 hours of their people's time, they own the report either way	Small, bounded, reversible
42–45	One ask: "Can you send me the last two years of completed job files and read-only access to Xero?"	A single, specific next action

Do not

Do not demo software. There is nothing to demo yet, and a mockup invites feature debate instead of a data conversation.

The proof artifact: one job, back-costed

Do this before the second meeting, free, on a single completed job. Ask for the accepted estimate, all supplier invoices, subcontractor invoices, timesheets, and any signed variations. One job, ideally one they were pleased with.

Deliver one page:

Quoted total versus actual total, and the realised margin against the quoted margin
The variance broken down by trade stage — where the money went, and where it went missing
Labour hours estimated versus actual for the two or three largest stages
Variations: how many, total value, how many were signed before the work was built
One sentence naming the single largest source of variance

Two outcomes, both useful. If the job made its margin, the finding is "you priced this well — can you reproduce it deliberately, or was it luck? We can only tell you that with more jobs." If it did not, you have their attention and you did not have to argue for it.

The rule

Do not soften the report. A padded number destroys the only asset you're selling, which is that the number is true.

The measurement instrument

Agree this table with the client before Phase 0 finishes. Unmeasured value is unrenewable.

Baseline and year-one targets
#	Metric	Baseline source	Year-1 target	Attribution risk
M1	Quote cycle time hrs, site visit → proposal issued	Estimator's recall, then timed for 5 quotes pre-launch	−40%	Low — directly observable
M2	Quote throughput detailed quotes/month, same headcount	Last 12 months of their records	+30% or the M1 reduction — not both	Low
M3	Estimate-vs-actual variance per job, per class	Phase 0 back-cost of 24 months	Narrow the spread for the top 3 classes; set the threshold from the baseline distribution, not in advance	High — see below
M4	Realised vs quoted margin	Phase 0	+1 percentage point, weighted average	High
M5	Variation capture rate signed before built ÷ all changes	Usually unmeasurable at baseline — that is the finding	≥90%	Medium — behavioural
M6	Period end → claim served	Their claim records	≤3 working days	Low
M7	Win rate by class	Last 12 months	Report it first; only target it once segmented	Very high — do not promise
M8	Quote-dependency people who can price a full-home reno unaided	Count today	≥2	Low, and it is the succession argument

Be honest about attribution

Two facts will corrupt this measurement if you do not name them first.

The market is moving

NZ dwelling consents for the 12 months to February 2026 were up 11.7% year on year, and February alone was up 22.9%. Revenue growth in the measurement window is not your doing, and claiming it will eventually be found out. Never present revenue or profit growth as a return on this system. Use per-job, per-class variance (M3, M4) — far more robust to market conditions.

Selectivity works without software

Being choosier about what you bid improves win rates by roughly 5–10% with no change to pricing at all. If the client gets more selective because your audit told them which jobs lose money, that is a real return — but it came from the report, not the software. Saying so builds more trust than claiming it for the platform.

The defensible claim is narrow and strong: for jobs quoted using the company's own back-costed history, estimate-vs-actual variance tightened relative to a like-for-like historical baseline.

The ROI model

Assumption

All figures illustrative. Rebuild this table with the client's actual numbers from §05 before showing it to them — a generic ROI table reads as a brochure.

Illustrative firm: 25 staff, NZ$7.6M revenue, 60 detailed quotes/year, 30% win rate, 18 jobs won, NZ$420k average contract, 12% target gross margin.

Modelled year-one return
Line	Basis	Year 1
Estimating labour recovered	60 quotes × 6 hrs saved × NZ$70/hr fully loaded (verified NZ$52/hr base for an NZ construction estimator, loaded ≈1.35×)	NZ$25k
Margin variance recovered	1 percentage point of the assumed 2pp slippage on NZ$7.6M	NZ$76k
Variations recovered	Half of an assumed NZ$6k per job × 18 jobs built-but-unpriced	NZ$54k
Modelled year-1 return		NZ$155k
Win-rate upside	+3pp = 1.8 extra jobs × NZ$420k × 12% — excluded	NZ$91k

Why win-rate upside is excluded. Winning more work requires the capacity to build it. This firm is already running 6–15 concurrent jobs; two extra full-home renovations may need another crew. Presenting NZ$91k of margin that carries hidden hiring cost is exactly the overclaim that kills a renewal. Mention it as optionality — "this either gives you back time or gives you more shots at bat, and you should decide which" — and keep it out of the total.

The framing that matters

Point at the table and say: the estimating time saving is the smallest number here. Owners expect an efficiency pitch. The value is in variance and evidence. If you win the deal on time saving, you will lose the renewal when someone counts the hours.

Investment shape

Three stages
Stage	Duration	Structure
Cost Memory Audit	~3 weeks	Fixed fee. Deliverable: the back-cost report, which they own outright. Includes the five-job reconciliation spike that determines whether the rest is even possible
Implementation Phases 1–3	Staged	Fixed fee per phase, released on the phase's completion state from blueprint §12.2 — not on activity
Subscription Phase 4	Ongoing	Monthly, per company not per seat

Hold the line

Refuse to quote implementation before the audit. You do not yet know whether their history reconciles (blueprint Q1), and quoting a build against unknown data quality is how fixed-fee projects lose money. Say this plainly — it reads as discipline, not evasion.

Anchoring. They are likely paying NZ$500–900/month for job management today (NextMinute from NZ$199/mo for 3–9 users; Buildxact NZ$199–599/mo). A subscription in the low thousands cannot be justified against that comparison and must not be argued on features. Justify it against the NZ$155k table, and be ready for "so why is it more than Buildxact?" — the answer is that Buildxact prices from a catalogue and this prices from their jobs, and the catalogue does not know what their crew achieves on a villa.

Objection handling

Nine objections and the response
Objection	Response
"We already have NextMinute / Buildxact / Fergus."	Keep it. Phase 0 doesn't touch it — we read history, we don't replace your workflow. Those tools do job administration well. None of them turn your finished jobs into the price for your next one; that's a different problem and it's the only one we're solving.
"AI can't price a villa renovation. Half the job is behind the walls."	Correct, and that is our thesis. Independent 2026 assessments say AI estimating works on clean plan sets and falls apart on exactly your work. We're not asking a model to guess what's behind your wall. We're asking your last seven villas what usually is — with the job numbers attached so you can check them.
"My estimator has 25 years in his head. A computer won't replicate that."	We're not replicating it, we're recording it. His rates, his allowances, his judgement about which villas bite — extracted from what he actually built and made durable. His name goes on the report. The question isn't whether he's good; it's what happens to the business the day he retires.
"What if it prices a job wrong and we lose money?"	Nothing prices anything. Every line shows where its number came from — a specific invoice, a labour norm from named jobs, or a subbie quote. Where there aren't at least three comparable jobs, the system refuses to recommend a contingency and tells you so. Your estimator issues the quote; he always did.
"Who owns our data? Will you sell this to the guys down the road?"	Have the answer written down before this meeting. Recommended position: your corpus is yours, irrevocably, exportable in open formats any time. We own the software. Nothing you generate trains anything shared. If regional exclusivity matters to you, let's price it.
"We're flat out. We don't have time for a project."	Phase 0 needs about six hours of your people's time in total, spread over three weeks: one session mapping your cost codes, and answering questions about five old jobs. Everything else is us reading documents you already have. Nothing changes on site.
"Can't I just use ChatGPT for this?"	For drafting a scope-of-works letter, yes — do that today, it's free. For pricing: you can't paste two years of supplier invoices and timesheets into a chat window and get back a defensible labour rate per square metre with the source jobs cited. That's a data problem, not a writing problem.
"What happens if you two get run over by a bus?"	Open formats, full export on demand, source escrow in the agreement, and documentation as a deliverable rather than a promise. Ask the same question of the vendor whose product has no public API — the honest answer there is that you can't leave.
"How much?"	The audit is [fixed fee]. I'm not going to quote the build until the audit tells us both whether your history is usable. If it isn't, I'll tell you, and the report is still yours.

The objection you should raise yourself

If their work is genuinely all one-off — no repeating job classes, no three comparable villas or bathrooms or extensions — the comparables model will not have enough data, and the honest recommendation is a different sequence (capture-first, with estimating intelligence 12 months out). Raise this before they do. The five-job spike is designed to find out, and telling them you might disqualify them is the most persuasive thing you can say.

What not to promise

Any specific percentage improvement in estimating accuracy before the audit has shown you their data.
That AI will read their renovation sketches reliably. Renovation drawings are frequently partial or hand-marked; quantity extraction will need human confirmation and the pitch should say so.
Compliance. You can say the system is designed to support CCA payment claim requirements and produces the evidence a reviewer would inspect. Do not say "CCA compliant" until counsel has reviewed a generated claim, and make no claim at all about the retention regime until blueprint Q6 is answered.
Revenue or profit growth. See §07.
Council integration. It does not exist across 66 BCAs.
Timelines for Phases 2–4 during the first meeting. Sell Phase 0.

Disqualification criteria

Walk away, or reprice heavily, if any of these are true. Each one breaks a load-bearing assumption in the blueprint.

No usable accounting history. No Xero/MYOB with job-level coding for 24 months → there is no corpus and no wedge.
The owner won't share financials. Understandable, and fatal. The product is their financial history.
Estimates were never written down. If quotes were verbal or lump-sum with no line detail, there is nothing to compare actuals against.
The five-job spike fails. More than 20% of historical cost unallocatable to a trade stage, or labour hours untraceable to jobs. Stop, deliver the spike findings, refund nothing and promise nothing further.
Fewer than three comparable jobs in their main job class. The contingency model cannot work.
The estimator is openly hostile and is the owner's relative or long-term partner. You will not win this politically, and the build depends on their cooperation.

Why this is worth the lost fee

Disqualifying early costs you a fee. Delivering a failed implementation costs you the reference case, which at this stage of the business is worth more than the fee.

After the meeting

Within 24 hours: a one-page email restating their own numbers from §05 back to them, the three questions they couldn't answer, and the audit proposal as an attachment. Nothing else. No deck.

The audit scope of work should specify

Inputs required, itemised: completed job files for 24 months, accepted estimates, supplier and subcontractor invoices, timesheets, variation records, read-only Xero access
The five-job reconciliation spike, run first, with its pass/fail test stated in advance — all five reconcile to Xero within ±1%; ≥95% of cost lines allocate to a trade stage; labour hours tie to stages
An explicit stop-point after the spike, with the client's decision either way and the spike findings delivered regardless
The deliverable: back-cost report by job and by class, variance analysis, win/loss and profitability by class and suburb, the agreed cost-code taxonomy mapped to their Xero tracking categories, and a stated recommendation on whether the estimating wedge is viable for them
Explicit data handling: NZ residency, single-tenant, no cross-tenant learning, deletion on request, full export
Who owns what: they own the report and the corpus; you own the software and the taxonomy

Meeting two

The second meeting is not a sales meeting. It is the audit report presentation. Bring the numbers, present them without commentary for the first ten minutes, and let the owner reach the conclusion. Then ask one question: "Which of these job types do you want to stop losing money on first?"

Leave-behind one-pager — draft copy

What your last two years of jobs actually cost

You quote from experience. That experience lives in two or three people, and it doesn't get any sharper with each job you finish — because nothing you own turns a finished job into the price for the next one.

AI didn't fix this, and 2026 made that clear. Plan-reading tools now cut takeoff time by 70–90% on clean drawings. On renovation work — partial plans, unknown substrates, scope behind existing linings — independent assessments say they fall apart. Your risk isn't in the drawing. It's in the similar jobs you've already built.

Which means the only data that helps is yours. Nobody can buy your realised labour rates on villa reline work, or your actual overrun distribution on wet-area relocations. That's the asset. It's sitting in Xero and a folder of invoices, doing nothing.

We start with a report, not software.

Give us 24 months of completed jobs. In about three weeks — and roughly six hours of your team's time — you get:

What each job actually cost by trade stage, against what you quoted
Realised margin by job type: which work you win, and which work makes money (rarely the same list)
Where variance actually comes from, ranked
How much scope got built without being priced and signed
A straight answer on whether your history is good enough to price your next job from

You own the report either way. If your data won't support the next step, we'll tell you and stop.

[Name] · [Contact] · Wellington

In one line

Sell a paid back-cost audit of their own history; the audit is both the proof and the corpus; estimating is the wedge because it is where the owner feels pain and where variance is measurable; everything else in the blueprint is downstream of owning that data.

Selling the Job Spine,
estimating first