Business plan for the data-centre scheduler

You were right in two ways

• We under-counted use: plan on about 12 delay analyses a project-month, not 1
• Schedule rows are heavier than we assumed (measured: see the chart)
• The 12 is the test bank's design density (20 delays planted in 7 weeks), not an observed rate
• Your original figure was 1 to 2 big delays; Janika scopes the first product to big delays only (as relayed, not read at source), nearer the Light column. We carry both.
• If big delays run 1 to 2 a month, the checked decisions included at bands C and D (4 and 6) exceed demand: safe on cost, but size the offer after buyer interviews

Measured: a schedule never fits in one read

• Caveat: those rows carry names, tags and link reasons; the packet the product sends may be smaller (not yet measured)
• Code-first assumes 45 tokens an activity; at the measured 86 it adds about $0.04 an analysis

One delay analysis, 9,000 activities

What the rulebook allows

• Agentic builds are forbidden: "the AI has no tools and never acts" checked: SPECS.md, Security Invariants
• Multi-agent without tools breaks today's $5 per analysis cap at its central case checked: SPECS.md, Numeric Constants
• Test set before each release (140 delay reports x 10 runs): about $3,100 code-first, $18,600 multi-agent, $100,000 disciplined agentic estimate
• Next measurement: one code-first analysis on a 9,000-activity schedule, timed and costed, before any figure leaves the founders

What the market does

• Unlimited users is standard for tools a whole project team shares checked: pricing pages, 11 Oct
• Per-seat survives in tools one specialist drives: Primavera P6 £244 a user a month, minimum 25; Foresight £24,000 a named user a year
• Where AI is unmetered, a paid seat usually bounds it (Intercom, Salesforce, Microsoft 365 Copilot); Outbuild includes AI agents with unlimited users and publishes no limit
• Where users are unlimited, the AI is usually metered: Procore says "We don't charge by the seat" and sells credits per task checked
• Flat fees without limits lose money: GitHub Copilot reportedly lost $20 a user a month, up to $80 on heavy users one anonymous source; Microsoft disputed it

What drives our AI bill

• Analyses and questions, not log-ins
• One extra person asking 60 questions a month: $2.27 code-first, $4.54 multi-agent, $134 agentic; 12 times a light user estimate
• Your fear is right for one design: an open AI box for everyone, built agentic
• A seat limit alone does not fix it (GitHub, Cursor); a monthly AI allowance does

What one more person costs a month

How the allowance works

Band A stays

• Kept (Munim, 11 Oct: "we may eventually be selling to them")
• Software only, sold later, is the research's proposal judgement
• $1,500 a project-month for contracts under $50M; no checked decisions included; unlimited viewers
• An allowance of 5 analyses and 150 questions a month (about $24 of AI and hosting when fully used) estimate
• At 4 analyses and 1 hour of support a month it keeps 77%: AI and hosting about $19 house, $26 blind-scaled, which the margin uses; AI alone about $9
• At most about $14,000 to win a customer (12-month payback): it needs self-serve sign-up, not founder-led selling
• With full support (4 hours a month) it would keep only 26% use: judgement
• At 3.6 times SmartPM with no checked decision, its premium has to rest on data-centre rules and costed options judgement, for the founders

What is out

• Deep (multi-agent) analyses are not in the allowance unless the founders raise the rulebook's $5 and $3 caps: they stay within both in only about 3% of runs estimate

Revenue types

• Subscriptions are recurring
• Pilots and extra checked decisions are services revenue, which investors leave out of ARR
• The checked decisions inside the subscription are also a service; investors will carve them out how the bundle splits: unverified
• A company-wide annual contract comes once a contractor runs 3 or more projects

Gross margin by delivery and band, $153 a delivered hour

If a customer needs our checks every month

• The margins rest on the AI running on its own after the pilot. If it cannot, the number that decides everything is our hours per checked decision
• At 4 decisions a month, 50% margin needs under 2.1 hours at band B, 5.2 at C, 10.7 at D; margin reaches zero at 5.4 / 11.8 / 23 hours. Nobody has measured it; the first pilot must
• Why $153, not $100: a planner is paid for every hour but delivers about 120 a month. At $100 the 4-decision row is 47% / 73% / 85%
• Band A, software only: 77% at 4 analyses and 1 hour of support; payback 55.9 months at a $60,000 cost to win, so at most about $14,000 to win one
• Band B with our checks every month works only at one decision a month: 67% and 27 months' payback; with four decisions, 21%
• Planner capacity: at 20 hours a project-month, one planner carries about 6 live projects
• At a typical software company's cost to win: LTV to CAC 1.7 to 2.1 with the AI alongside, 1.4 to 1.6 with our checks every month; payback 44 to 51 months. The business needs founder-led, efficient selling

Two margins, two questions

• Gross margin is per project: price minus the cost of delivering it
• Operating margin is the whole company: gross profit minus the team, selling and overheads
• Break-even needs about 57 live projects at the 2031 cost base, against 37 on the path estimate
• Profit needs faster customer growth, more projects per customer, the second price case, or construction beyond data centres (market section)
• The margins rest on the AI running on its own after the pilot; if a customer still needs our checks every month, its margin falls to the checked-decision rows in the unit economics section

What the pre-seed buys

• YC's standard deal: $500,000 ($125,000 for 7% plus $375,000 on an uncapped SAFE) checked: ycombinator.com/deal
• The demo at real schedule size
• 3 paid pilots; 8 live projects
• Measured hours per checked decision

Live projects against break-even

The path, in short

• 3 / 5 / 7 / 9 / 11 new contractor customers a year, each starting with one project and growing at NRR 104% judgement
• Live projects at year end: 3 / 8 / 15 / 25 / 37
• The path assumes each customer books its next project as the last one ends: 52 in 2031 needs NRR about 127%; 110% gives 40
• 2027 is a pilot year billed at $4,000 a month; from 2028 every new customer starts with the 3-month pilot (Janika checks, fee credited to year one), then the AI runs on its own
• Janika delivers up to 60 hours of checks a month in 2027 and 2028 at no extra cash cost; our delivery hours stay at 15 to 33 a month
• Band mix B 30%, C 40%, D 30%; band A is not in the path (sold later)
• Team 3 / 7 / 11 / 15 / 19 people costing $0.45M to $4.6M a year; $60,000 to win each new customer

Why the pilot is checked by us

• It sells with the least product, teaches us the work and earns the trust that lets the AI run on its own afterwards
• Warren (YC): "You have to sell outcomes, not seats or tokens." checked
• Two YC request rounds asked for AI that does the service (Spring and Summer 2026)

The trap

• Our people quietly covering product gaps
• Measure our hours per decision every month
• Cap pilots at two or three. Warren: "cap your first pilot customers to a small handful" checked

What the rulebook needs

• Today the founders may see customer data "only for options flagged as wrong" or for support checked: SPECS.md
• The checked pilot needs a recorded, customer-switched-on "our analyst on this project" role that keeps the shared-library firewall

The narrative

• Problem: a late chiller or switchgear package stops the critical path; finding what can move takes days of planner work in P6
• Janika: a planner "has to manually think about it for multiple days". No buyer has confirmed it yet
• Why now: $84.95B a year, up 73% checked: Census; projects get bigger, not more numerous: about 20 to 23 US starts a month, 701,000 sq ft on average in 2026 against 454,400 in 2025 checked: ConstructConnect
• Why us: Janika plans these jobs and knows which links can really move; Munim builds it. Blomfield: "70 team, 20 idea, 10 traction"
• First: the scheduler's job on data-centre builds, after a paid pilot Janika checks
• Where it goes: the same system on every large construction project in the US and Europe

The money answer, in the shape of the accepted 2025 answer

• Price: $4,000 to $15,000 a project-month by contract size, about half to one planner
• Gain: a delay analysis that takes two or three planners a week, done in hours
• Market, last: about 450 US data-centre projects (about $31M a year, up to about $80M as projects grow), then every large project in the US and Europe, about $1B a year; $100M a year is 10% of it
• How YC sizes markets: Seibel prefers bottom-up; Graham: be matter-of-fact and disclose the flaws; partners argue AI is paid from the wage budget, about 4% to 10% of a customer's spend against about 1% for software
• Say plainly that data centres alone are small: that is why they are the start, not the market

YC's own request, and the YC companies a partner will name

• Spring 2025 "Datacenters" request asked for data-centre "project management"; none of YC's 40 construction companies from 2025 and 2026 has scheduling as its main product checked
• "AI agents that can quote complex work and schedule teams" (Fall 2026, Warren) checked: ycombinator.com/rfs: it is about crews, not planning teams. Do not overstate it
• Friedman calls project-management tools "simple targets": never pitch a dashboard. Warren: "You have to sell outcomes, not seats or tokens."
• Applied Kinetics (F26): agents that "chase suppliers, track equipment approvals and delivery dates"; Marengo (S26): an AI engineering firm for data centres; Karmen (F24): "AI scheduling assistant for construction"
• Ours starts from the reported delay and ends with a checked recovery plan; one line each, before the interview

The two objections

• "ALICE and SmartPM do this": SmartPM sells time impact analysis and managed services; ALICE claims a data-centre recovery "Within hours" checked
• Our answer: link meanings from a planner's rules; every option costed and checked by code; the statement kept separate from the recovery, so entitlement is never given away
• Proof that settles it: a contractor who has seen them and pays us
• "Isn't this a consultancy?": only the pilot is checked by us; after it the AI runs on its own

Avoid, and say plainly

• Avoid: "a $16 trillion construction industry" (industry size, not our market), project-management language, per-seat pricing, and the flagged words ("wedge", "AI-native", "agentic", "AI-powered platform")
• The 28 Oct demo uses the 138-activity schedule; real ones run 9,000 to 50,000. Say so, and name the size the product is built for

Prove before 2 Nov

• Five or more contractor conversations with numbers, before 2 Nov
• One written commitment: a letter of intent, a paid pilot, or a re-analysed past delay
• Before the interview: one checked decision delivered for money, with hours, days recovered and our hours spent

US data centres by year: the count stays flat

2031 if projects grow bigger: the growth is in size

• US starts have run at 20 to 23 a month since late 2024 while the dollars rose about 5.5 times checked: ConstructConnect
• About 412 live today, 460 at the 2029 peak, 416 in 2031
• If every project paid the top price: about $80M a year

What sets the size

• Big projects are a minority: about 30% of US data-centre projects are over $250M (15% to 50%) estimate
• July 2026 had 23 starts worth $2.7B in total, so at most 4 were over $250M checked: ConstructConnect
• Our draft prices with half the over-$250M projects paying $15,000: $31M; $26M if none does

The second price case

• Fits only the end-to-end system: the delays product leaves the scheduler in place
• Best on projects over $250M: on those alone, $36M in the US
• It assumes $220,000 loaded per scheduler load factor: unverified
• That needs a salary near the top tenth of construction project managers: 90th percentile $160,510, mean $104,890 checked: BLS May 2025

Against the bar

• A fund looks for a market of at least $5B, ideally a path to $15B, and $100M a year of revenue checked: Euclid Ventures
• Nothing clears $5B in the central case
• $100M a year is 10% of the US plus Europe end-to-end market, 21% of the US one

Not counted, and the weakest inputs

• Not counted: small builders' job-management software (a different product, about $147M in the US), home builders ($564M), China
• Weakest input 1: the share of projects over $250M (no project database pull by value band)
• Weakest input 2: the share we win: "about 60 data-centre contractors" is the length of a published top-60 list, not a count
• Weakest input 3: every price is a desk estimate until buyers name one

Buyers and downturn

• Buyers are concentrated: about 15 US general contractors build most of the volume judgement
• Downturn risk: 75 projects worth $130B blocked or delayed in Q1 2026 checked via Bain

Team

• Full-time status and the CTO's depth

Whether the AI can run on its own after the pilot

• The margins assume it
• If a customer needs our checks every month, band C falls to 60% at 4 hours a decision, below 50% above about 5.2 hours, and to zero near 12 hours

Accuracy at real size

• One impossible option, or dates that do not match P6, ends a pilot; the demo is 138 activities

Liability and who signs

• Our planner checks in the pilot; the contractor's project lead signs
• If our planner signed the statement the contractor submits, we would author a contractual submittal
• Contract terms, professional liability insurance and who signs need counsel before the first paid pilot

Revenue ends with each project

• Fees stop when the project ends, and data-centre starts can stall

One model provider

• Top-tier list prices flat to rising since 2025; a price or terms change hits every project judgement

Access to customer data

• A contractor may not let our planner see its schedule (the founders already expect clients to resist automatic reading of their emails and chats: scheduler decision DL-023); then checked decisions become self-serve

Founders' decisions, for after you have both read this

• Settled by Munim on 14 Oct, the framing: an AI that does the scheduler's job, data centres first
• Settled, the delivery: Janika checks in the pilot, then the AI works alongside and replaces scheduling work over time
• Settled: the second price case shown beside the draft prices
• None of the five below needs an answer now.

1. Users: unlimited viewers plus an AI allowance, limited named seats each with an allowance, or unlimited viewers plus a few named analysts who start AI runs?Section 3: both norms exist
2. Band A: kept (Munim, 11 Oct)Software only, sold later, is the research's proposal
3. Which AI build?Research recommends: code-first by default; multi-agent only if it proves better on the test set and the caps are raised; never an open agentic chat
4. Rulebook changes for Janika: an AI stop at the limit that spares the core delay analysis (reverses "alert rather than block")Also: deep-tier caps; the "our analyst" role for the pilot; a question box or not
5. Still open: the product name before YC; who signs (counsel, risk 4); whether the draft prices move toward the second price case once pilots measure the value

Not established

• No buyer interview yet: willingness to pay, who signs, delays a month, planner hours per recovery, damages rate and cap, AI acceptance by owners
• Whether the AI can run on its own after a 3-month pilot: the delivery model; only pilots show it
• Our hours per checked decision: only a pilot measures it. A planner's delivered hours a month (120 assumed) sets the hourly cost
• Janika's delivery hours a month (60 assumed) set how many early projects cost no extra cash
• Whether contractors will let our planner see their schedules
• Real cost per analysis on a 9,000-activity schedule (row size measured; the product's packet and a full run not yet)
• Real schedule size: no published activity count for a real data-centre schedule; our bands assume 9,000 / 25,000 / 50,000
• Oracle's P6 sizing guide tiers projects at 7,500 / 25,000 / 75,000 activities; realistic schedules carry 2 to 4 links an activity secondary: test-bank research
• Delays a month on a real job: about 13 delay-causing issues per weekly report on one London high-rise (not data-centre), most of them small secondary: test-bank research
• Unknown: real CAC; pilot conversion in construction; projects per customer; how investors value per-project revenue; how the bundle splits between software and service revenue
• Not found: prices for ALICE, nPlan, Buildots, Document Crunch; US delay-consultant rates; the median data-centre project value
• Not found for the market: the share of projects of each size that run a formal critical-path schedule (the input that swings it most); the size mix of US data-centre projects (30% over $250M assumed)
• No product code yet: specification, checks, research and the test bank; demo due 28 Oct