AI-first preconstruction

MeltPlan introduces AI-first preconstruction — a construction-native AI system that handles the grunt work accurately, and hands over where expert judgment makes the difference.

01The Problem
02The Solution
03The Outcome
01

Frontier model apps like ChatGPT and Claude aren’t useful for preconstruction.

General AI from OpenAI, Google, and Anthropic is built to answer everything with above average IQ, not necessarily focusing on high standards of construction IQ.

Precon requires construction–native intelligence – reading drawings, classifying scopes, reasoning through trade sequencing – not a broad model pointed at a nuanced industry.

Frontier models, projected to 20300%25%50%75%100%Capability score (%)202220232024202520262027202820292030LanguageVisionReasoning
Language
Summarize docs, search specs
MMLU (like SAT for AI)
Vision
Read drawings, match takeoff
AECV Bench
Reasoning
Scope gaps, constructability calls
Humanity’s Last Exam
02

Three layers turning general intelligence → AEC intelligence.

Layer 1: Small Language Models → Built small language models expert at specific construction tasks.
Layer 2: Harnesses → Built structured harness workflows that prevent AI from derailing.
Layer 3: Guardrails → Guardrail systems that catch errors before they reach your team.

Frontier Model
Small Language Model
Construction ontology
Harness
Domain & workflow understanding & orchestration
Guardrails
Multiple checks to keep it on track
Trusted output (95%+)
Clears the trust bar, today
MeltPlan AI
Stack
03

Precon AI accurate enough to save you valuable time

MeltPlan AI works on a strict principle → we hand over nothing to you until the AI is 95%+ accurate. That’s the bar.

Melt Bid

Expert estimators leveled 100 bid proposals across 40 packages, capturing 8,500 data points. Each AI’s answer was scored against the estimator’s answer

Melt Bid
95.3%
accuracy, 8,500 data–points
Claude, OpenAI, Gemini
75–80%
accuracy, 8,500 data–points
Melt Code

2,000+ questions from official building inspector certification exams across 4 discipline tracks. Each question scored against certified inspector answer keys.

Mechanical (M2)98%
Plumbing (P2)97%
Residential (B1)94%
Commercial (B2)93%
01 The Problem

Frontier model apps like ChatGPT and Claude aren’t useful for preconstruction.

  • General AI from OpenAI, Google, and Anthropic is built to answer everything with above average IQ, not necessarily focusing on high standards of construction IQ.
  • Precon requires construction–native intelligence – reading drawings, classifying scopes, reasoning through trade sequencing – not a broad model pointed at a nuanced industry.
Frontier models, projected to 20300%25%50%75%100%Capability score (%)202220232024202520262027202820292030LanguageVisionReasoning
Language
Summarize docs, search specs
MMLU (like SAT for AI)
Vision
Read drawings, match takeoff
AECV Bench
Reasoning
Scope gaps, constructability calls
Humanity’s Last Exam
02 The Solution

Three layers turning general intelligence → AEC intelligence.

Layer 1: Small Language Models → Built small language models expert at specific construction tasks.
Layer 2: Harnesses → Built structured harness workflows that prevent AI from derailing.
Layer 3: Guardrails → Guardrail systems that catch errors before they reach your team.

Frontier Model
Small Language Model
Construction ontology
Harness
Domain & workflow understanding & orchestration
Guardrails
Multiple checks to keep it on track
Trusted output (95%+)
Clears the trust bar, today
MeltPlan AI
Stack
03 The Outcome

Precon AI accurate enough to save you valuable time

MeltPlan AI works on a strict principle → we hand over nothing to you until the AI is 95%+ accurate. That’s the bar.

Melt Bid

Expert estimators leveled 100 bid proposals across 40 packages, capturing 8,500 data points. Each AI’s answer was scored against the estimator’s answer

Melt Bid
95.3%
accuracy, 8,500 data–points
Claude, OpenAI, Gemini
75–80%
accuracy, 8,500 data–points
Melt Code

2,000+ questions from official building inspector certification exams across 4 discipline tracks. Each question scored against certified inspector answer keys.

Mechanical (M2)98%
Plumbing (P2)97%
Residential (B1)94%
Commercial (B2)93%

What AI-first Preconstruction means for you.

AI takes the first pass on every precon workflow, handling the grunt work you shouldn’t have to do, then hands it over to you for the judgment calls that only your experience can make.

01AI–first, not AI–powered02AI that works for your team, not against it

MeltPlan AI takes the first pass – reading plans and specs for design issues, counting quantities, normalizing sub proposals, and hands it over for two things:

  • To flag its gaps where your experience is needed.
  • To make the decisions that require your precon judgment.

Most traditional construction software added AI as an afterthought – a chatbot, an autofill, a summary button bolted onto existing workflows. These don’t take work off your plate; they give you more things to check.

When MeltPlan hands over the work, it is accurate, cited, and honest about its gaps. Your expertise finishes it. Four things are always true at the handoff:

  • Accurate enough that your review improves it, not corrects it:Below 85%, fixing AI output takes longer than starting fresh. Above 95%, your estimator adds judgment, not proofreading.
  • Every number traces to a source:Drawing sheet, spec section, sub quote, all verifiable in seconds.
  • Gaps called out, not hidden:Low confidence and missing coverage surfaced explicitly so you know exactly where to step in.
  • Handed over at the right stage:A near–complete deliverable where your experience closes the last gap between a good estimate and a winning one.
01

AI–first, not AI–powered

MeltPlan AI takes the first pass – reading plans and specs for design issues, counting quantities, normalizing sub proposals, and hands it over for two things:

  • To flag its gaps where your experience is needed.
  • To make the decisions that require your precon judgment.

Most traditional construction software added AI as an afterthought – a chatbot, an autofill, a summary button bolted onto existing workflows. These don’t take work off your plate; they give you more things to check.

02

AI that works for your team, not against it

When MeltPlan hands over the work, it is accurate, cited, and honest about its gaps. Your expertise finishes it. Four things are always true at the handoff:

  • Accurate enough that your review improves it, not corrects it:Below 85%, fixing AI output takes longer than starting fresh. Above 95%, your estimator adds judgment, not proofreading.
  • Every number traces to a source:Drawing sheet, spec section, sub quote, all verifiable in seconds.
  • Gaps called out, not hidden:Low confidence and missing coverage surfaced explicitly so you know exactly where to step in.
  • Handed over at the right stage:A near–complete deliverable where your experience closes the last gap between a good estimate and a winning one.

How we pressure–test our AI system

Expert–built evaluations

Every product goes through a structured evaluation process before and after every model update.

01
Golden sets

The correct answer is established first – by practitioners in the top 2% of their discipline. Not crowdsourced through ChatGPT.

02
Diversity of data

Eval sets cover the full range avoiding bias: different project sizes, building types, firm sizes, subcontractor quote formats, and more. A model that only works in one context doesn’t ship.

03
Continuous evaluation

Every model update triggers a full eval run. If accuracy drops for any reason, we catch it before it reaches production. The numbers published reflect the current production model.

04
The 95% threshold

Below 85% accuracy, AI costs more time than it saves. Between 85% and 95%, the experience is hit or miss. Our target is 95%+ because consistency is what makes the workflow real.

01

Golden sets

The correct answer is established first – by practitioners in the top 2% of their discipline. Not crowdsourced through ChatGPT.

02

Diversity of data

Eval sets cover the full range avoiding bias: different project sizes, building types, firm sizes, subcontractor quote formats, and more. A model that only works in one context doesn’t ship.

03

Continuous evaluation

Every model update triggers a full eval run. If accuracy drops for any reason, we catch it before it reaches production. The numbers published reflect the current production model.

04

The 95% threshold

Below 85% accuracy, AI costs more time than it saves. Between 85% and 95%, the experience is hit or miss. Our target is 95%+ because consistency is what makes the workflow real.

See what happens to your precon with the right technology

Bring a set of drawings, a spec book, or a stack of sub quotes — we’ll show you how MeltPlan AI works on your project.

MELTPLAN