We built an AI estimating platform for lighting projects. We expected the pricing logic to be the hard part. It wasn’t. The hard part was reading what customers send.
A typical retrofit package comes as a fixture count in Excel, a scope of work in PDF and a distributor quote in another PDF. There’s a scanned field sheet with handwriting in the margins and photos of a warehouse ceiling. And there’s a forwarded email from the facility manager changing half the fixtures. Before anyone prices a line, an estimator has to open all of it, reconcile it and type it into a spreadsheet.
Estimators know the math. A spreadsheet multiplies quantity by unit price without error. The time and margin go missing earlier, when people interpret the information, and again each time they retype it.
Five things we learned:
- A lighting project arrives in six formats, and someone retypes all of it.
- The expensive mistakes are confident wrong assumptions, not bad arithmetic.
- Incentives can move the customer’s price more than the fixtures do.
- Every handoff after the quote is another place to lose margin.
- AI should show its work and leave the estimate to the estimator.
1. A lighting project arrives in six formats, and someone retypes all of it
Lighting work sits where manufacturing, distribution, construction and utility programs meet, and each of those groups sends documents its own way. The architect sends a fixture schedule. The distributor sends a PDF quote. The site walker sends photos and a scanned audit sheet. The customer sends an email.
Your estimator becomes the integration layer. They count fixtures in one file, check them against another, look up wattages, find the latest price and catch the email that swapped a 2×4 troffer for a 2×2. The judgment you hired them for, like choosing a fixture, reading a site or setting labor, gets whatever time is left after transcription.
That cost doesn’t show up in a quote. It shows up in the bids you couldn’t respond to that week.
2. The expensive mistakes are confident wrong assumptions, not bad arithmetic
Look at which estimates come back to hurt you. The sum is rarely wrong. What’s wrong is a high-bay job priced without a lift rental, or a missing permit fee or lamp disposal line. It’s a labor rate that defaulted to zero on one row, or a fixture price copied from a quote last spring.
Each of these gets through review because the number looks plausible. A reviewer scanning a 200-line costing sheet catches blanks. A plausible wrong number gets past them.
The field side makes this worse. The U.S. Bureau of Labor Statistics projects about 72,700 electrician openings a year from 2025 to 2035, with employment growing 9%. A missed scope item means an extra trip by a crew that’s already hard to staff, and that cost lands on your margin, not the customer’s invoice.
3. Incentives can move the customer’s price more than the fixtures do
A retrofit proposal sells the customer a net price, meaning the cost after utility incentives. If the incentive is wrong, that price is off by the full incentive amount, even when every fixture and labor line is right.
Incentives also depend on things that change. Program rules get updated. A fixture that qualified when you quoted may not qualify when you install. Check the product’s status and the program’s rules, and record the date you checked.
4. Every handoff after the quote is another place to lose margin
Winning the job starts a second round of retyping. Someone enters quantities and prices into the ERP. Purchasing types them again into a PO. The project manager rebuilds why the estimator picked a fixture, which exclusions they priced and what lift and permit assumptions they made, usually by digging through email.
This isn’t new to lighting. A 2021 survey of 3,900 construction professionals by Autodesk and FMI estimated that bad data, meaning information that’s inaccurate, incomplete, inconsistent or hard to find, may have cost the global construction industry $1.8 trillion in 2020. The same survey linked it to 14% of avoidable rework.
Software has limits here. Better data won’t shorten a six-week fixture lead time or create inventory that doesn’t exist. What a cleaner estimate can do is cut down how often people type the same numbers again, and keep the reasoning behind them attached to the job.
5. AI should show its work and leave the estimate to the estimator
“Auto-generated quotes” sound good until you picture an AI that quietly fills in a missing wattage. A blank cell gets noticed. A confident guess gets priced, approved and sent.
In a trade where one assumption can wipe out a job’s margin, estimators need three things from any tool. They need to see what it read, what it inferred and what it couldn’t find. The useful version of AI here reads faster than a person and then hands the estimator a list of things to confirm.
How we built for this
Those lessons shaped AISquared’s lighting estimation platform and its AI assistant, Sparx Lumen. A quote comes together in four steps.
- Drop in documents. Upload whatever the job came with, in any mix. The platform reads 14 file types: spreadsheets (XLSX, XLS, CSV, TSV), PDFs, scanned PDFs and photos (JPG, PNG, WEBP), Word documents, Outlook emails and text notes. It also fills in the customer, address, building type and scope from the documents.
- Review what was read. The platform pulls every fixture line from every document into one review table. Where the documents leave a gap, it asks a question instead of guessing. It marks inferred labor and material values “estimated” and leaves them unticked until you confirm them. Nothing reaches the quote until you add it.
- Price in the sheet. Lines land in a spreadsheet-style costing grid that calculates labor, material sell, unit price and subtotal. You can paste from Excel, add calculated columns and export to CSV. Fixture suggestions and price checks come from your own past quotes. Workspace defaults cover your labor rate, markup, energy rate and incentive rate, and you can price turnkey, labor-only or mixed jobs.
Sparx Lumen works on the same numbers. Ask “What did we last charge for a 2×4 T8 retrofit?” or “What’s missing from this costing?” and it answers from your job and your history. It runs what-ifs, like labor 10% higher or markup at 20%. It flags missing wattages, zero labor rates and prices far from past jobs. Ask it to add a lift rental and a permit fee, and it stages both changes for your approval. Every AI edit gets staged, and the audit log records whether each change came from a person or from Lumen.
- Send the proposal. One click produces a one-page summary with scope, investment, utility incentives and net price, plus a full quote PDF with the fixture schedule, energy savings, payback, terms and a signature block, on your letterhead. The platform calculates incentives from watts saved. After you send the quote, you record whether you won or lost it.
For managers, owners approve or request changes from one queue. Version history covers every quote, and a manager view shows pipeline, win rate, quotes expiring soon and loss reasons by estimator.
What it doesn’t do. The platform doesn’t replace your ERP, inventory system, field audit app or dispatch board. It covers the stretch from incoming documents to a sent proposal, which is where most of the rework starts.
A checklist you can use without buying anything
- Time the touch, not the calendar. Log the minutes an estimator spends on each quote, split into finding data, retyping it, clarifying it, pricing it and formatting it. Most teams find transcription is the biggest share.
- Mark every inferred value. If someone estimated a wattage or labor hour rather than reading it from a document, make that visible in the sheet with a color, a column or a flag.
- Keep a missing-items list. Lift rentals, permits, disposal, after-hours labor, freight. Check every quote against it before review.
- Record each price’s source and age. “84, distributor quote, March” tells a reviewer far more than just “84”.
- Date-stamp compliance. Record DLC status and program rules as of the quote date.
- Hand off a job record, not an email thread. Give the PM one file with quantities, products, assumptions, exclusions and incentive math.
How to measure success
Don’t measure “quotes generated.” A tool can produce more quotes and still make worse ones. Track these four instead:
- Estimator touch time per quote, from intake to approval, not counting time spent waiting on the customer.
- Critical omissions found after approval: missing material, labor, equipment or incentive items.
- Manager correction loops: how often a quote goes back to the estimator.
- Incentive corrections: quotes where the incentive changed after the first calculation.
Take a baseline on your current process before changing anything, and compare the same kinds of jobs afterward. If touch time drops and omissions don’t rise, you’ve removed retyping, not judgment.
See it on your own project documents. Bring a real job package, in whatever format it arrived, and we’ll walk through it live.
