Pre-Bid Operations Automation
Crosno: Turning uncertainty into confident bids
Every bid carries risk. Building a system that helps estimators identify missing information, validate assumptions, and submit more confident proposals.
Crosno Construction
Welded steel storage tanks
Sector
Specialty industrial construction. welded steel storage tanks
Engagement
Discovery + "Bid It" build
Duration
6 months
Bids screened
23 bids · ≈ $31M volume
Team
Nexxt Ideas + Crosno pre-bid ops
Stack
Self-hosted · spec-extraction model · OCR ingestion · governed knowledge base · approval gating
Knowledge entries
34 formalized (12 never written down before)
Headline outcome
+6.1% margin · 0 repeat misses
The client
Two decades of know-how, trapped in people.
— Head of Pre-Bid Operations
“Every loss was something we already knew. It just lived in someone’s head, and that someone was on another bid.”
Crosno Construction builds welded steel storage tanks and competes in a world where a single missed requirement on a 200-page bid package can erase a project’s entire margin. For twenty years, the intelligence that won or lost those bids lived in the heads of a few senior estimators.
Which consultants run long. Which owners demand unusual insurance. Which specs hide expensive surprises. None of it was written down as it was remembered, by people who were always busy.
The challenge
Three ways margin quietly leaked.
01
Read by hand, under deadline
Two-hundred-page packages were read manually against the clock. Requirements got missed, and a miss surfaced as a change order, not a line item.
02
Knowledge that could retire
The best estimator’s memory was the system of record. One vacation, one departure, and a decade of judgment was simply unavailable.
03
No paper trail
Risk calls were made in hallways. When an owner disputed scope later, there was nothing on record to stand on.
— CEO, Crosno Construction
“The most expensive miss is the requirement nobody saw until the project was underway.”
The approach
We mapped the 11 ways the old process leaked — before building a screen.
Discovery catalogued every failure mode in the pre-bid workflow. The product is now being built from a single rule: AI proposes, humans decide, and nothing reaches estimating with an unresolved risk.
01
Ingest the package
Every page OCR’d and normalized — drawings, specs, addendums, insurance requirements, and general conditions.
02
Extract every requirement
AI proposes a structured list. Nothing counts yet.
03
Validate against source
A named human confirms each item against the source document.
04
Screen against the
knowledge base
knowledge base
Auto-match each bid to 34 governed entries of hard-won risk.
05
Gate the bid
No advance to estimating with an unresolved risk — no exceptions.
06
Record the decision
Every approval immutable, signed, and re-checked on every addendum.
Design decisions that made a miss hard.
Authority
AI proposes, humans decide
The model never commits a requirement. A named estimator validates each one against the source so the system is fast without being trusted blindly.
Gating
A hard approval gate
A bid physically cannot advance with an open risk. Overrides are allowed, but logged, signed, and tied to the bid forever.
Memory
A governed knowledge base
Hard-won lessons become curated, versioned entries that screen every future bid automatically. Institutional memory that can’t retire.
Control
Addendum reconciliation
Every addendum re-screens the whole package, so a late change can’t quietly invalidate an earlier decision.
The solution
One screen turns a 200-page unknown into a checklist.
01 · The Bid-It console
AI reads the spec; a human signs off every line
02 · Spec extraction engine
From paper to structured risk
03 · Governed knowledge base
Memory that compounds
04 · Approval & audit
Decisions you can prove
The target · first 8 months
Measuring where the business actually needs it.
Metric
Before
After
Target Change (est.)
Metric
Spec extraction time
Before
8–12 hours
After
23 min
Change
−97%
Metric
Average margin on awarded bids
Before
baseline
After
+6.1%
Change
+6.1%
Metric
“We already knew that” surprises / yr
Before
2–3
After
0
Change
−100%
Metric
Requirements caught per bid (avg)
Before
—
After
+18
Change
n/a
Metric
Knowledge entries formalized
Before
~0
After
23 min
Change
n/a
Metric
Addenda reconciled automatically
Before
0
After
100%
Change
n/a
A junior estimator will have a full package before lunch which is work that used to consume a senior’s entire afternoon.
The knowledge base will capture twelve risk patterns that had never been written down, including three that had each cost a project in the past.
Test across 23 bids and roughly $31M of volume
Aiming for efficiency is one thing, helping produce faster revenue and ROI is the main goal.
"We aim to finish a bid faster and know exactly what we decided, and why. The arguments can then move from the hallway into the record."
"Our best estimator's judgment is an asset the company owns, not a risk that walks out at five o'clock."
Key insights
What this engagement is teaching us.
Advisory AI earns trust that
autonomous AI never will.
autonomous AI never will.
Because the model only proposes and a human signs, the team can build the foundation without second-guessing. Speed from the AI create confidence while the business literates on the output.
The gate is the product.
Extraction is impressive, but the hard part “can’t advance with an open risk” rule is what actually changes the outcome. Discipline here will beat cleverness.
Curation turns memory
into an asset.
into an asset.
Letting only validated patterns into the knowledge base means it will stay trustworthy, and grow by what Crosno needs, not what an algorithm guesses.
A decision record pays
for itself in one dispute.
for itself in one dispute.
The audit trail is sold as governance.
What’s next · Phase 2
Extending Bid-It along three vectors.
01
Margin modelling
Use the structured bid history to forecast the margin band on a job before the team commits to bidding it.
02
Subcontractor intelligence
Score consultants and partners on historical risk so the screen flags relationship-level patterns, not just specs.
03
Live addendum alerts
Push an alert the moment an addendum lands that touches a previously approved risk decision.
This is our impact
Bringing Crosno a memory it can bid on.
Two decades of know-how lived in people, not systems. We are turning it into a governed platform that screens every bid, lets AI propose but never decides, and records every risk decision — so nothing the company already know slips through again.
Nothing slips
Zero “we already knew that” surprises, versus two or three a year before.
Healthier margins
Catching cost risk before it’s priced in.
Decisions you can prove
