The MAGE brain

It writes down every call it makes. Then it shows you the misses.

Every duration, price, leak flag and verdict goes into a ledger the moment it's made. When the job closes, MAGE grades itself against what actually happened. The misses stay on the board — that's the only thing that makes the hits worth anything.

The loop nobody else closes

Most construction AI never finds out if it was right.

It hands you a number, you build the job, and nobody goes back to check. The next estimate is exactly as blind as the last one. MAGE closes that loop — on your jobs, in your numbers.

01 — PREDICT

It commits, in writing

Framing takes 6 days. This estimate totals $42,000. Four items on this job are leaking money. Each call is recorded when it's made, not reconstructed afterwards from a generous memory.

02 — RESOLVE

The job answers back

Framing actually took 6. The job closed at $42,840. Three of the four flagged items got billed back and one got eaten. When the job closes, the real outcome lands next to the prediction.

03 — GRADE

Hit, miss, or tie

Every resolved call gets scored — a duration inside a day is a hit, an estimate inside 3% is a hit, a leak nobody billed back is a miss. Nothing is graded on vibes, and nothing quietly disappears because it went badly.

Your track record

The receipts, not just the score.

Any vendor can put a percentage on a slide. This is the itemized list behind it — what was predicted, what actually happened, and what that made it.

Track record predicted → actual
Framing6d 6dbeat the AI's 8dHit
Tile4d 7d3d overMiss
Drywall5d 5dmatched the AITie
Estimate$42,000 $42,840+2% vs actualHit
Profit-leak scan4 flagged 3 recovered$6,180 billed backHit
Margin verdict20% target 17%−3% vs targetMiss
Tracked bid62% win odds lostcall didn't landMiss
Illustrative rows in the real format. Your board is built only from your own graded predictions, and any category with fewer than three graded results shows nothing at all. Track Record is a Business-plan screen; the grading itself runs on every plan.

Earned autonomy

It has to earn the keys.

Plenty of software will happily automate your business on day one, on the strength of a demo. MAGE won't touch a thing until it has proved itself on your jobs — and the thresholds aren't a secret.

n ≥ 5  ·  rate ≥ 60%

Before it fills in a duration

MAGE won't pre-fill task durations for a trade until its learned pace has matched or beaten the generic AI call on at least five graded predictions for that trade, at 60% or better. Per trade — earning it on framing says nothing about tile.

n ≥ 5  ·  billed ≥ 50%

Before it drafts a change order

It won't draft change orders on its own until at least half the profit-leak items it flagged were actually billed back, across at least five graded scans. If your flags aren't converting, it stays quiet and leaves the call to you.

n < 3  ·  shows nothing

Before it claims an accuracy number

Any prediction type with fewer than three graded results is suppressed entirely, and tracked-bid calibration needs five. A hit rate off one or two jobs isn't a hit rate, and rounding it up to look clever would poison every other number on the board.

Why the grading matters

It prices from your jobs, not a catalog.

Once MAGE knows what work actually costs you, that's what goes on the next bid. Drop in a plan set and every priced line says where its number came from — so a guess can't sit next to a real number and look identical.

Takeoff estimatepriced from your history
Framing1,240 SFyour rate · 7 jobs · ±9%Yours
Drywall3,100 SFyour rate · 4 jobs · ±12%Yours
Skylight flashing3 EAno history yetCatalog
A generic AI estimator scores zero in that first column, permanently — it has never seen your books. For what the work costs across the trade, see the free MAGE Price Index, published from real completed jobs.

Ask your software how often it's wrong. Then ask why it can't tell you.

Straight answers

The questions worth asking any vendor.

How does it know whether its calls were right?

It records each prediction when it's made, then compares it to the real outcome once the job closes — durations against actual task spans, estimate totals against actual cost, leak flags against whether the item was actually billed back, bid odds against whether you won. Each one is marked hit, miss or tie.

What if I've only finished a couple of jobs?

Then it shows you almost nothing, on purpose. Categories under three graded results are suppressed, and the autonomy features above stay switched off until their thresholds are met. It gets sharper as your history accumulates — the first few weeks are honest about being early.

Can a competitor just copy this?

They can copy the screens in a quarter. They can't copy fifty of your finished jobs. The scoreboard and the pricing both run on your own outcomes, which is the one input a rival can't ship in a release.

Does my cost data get shared?

Not unless you switch it on. Contributing to the public Price Index is opt-in and off by default, and even then a trade only publishes once at least five independent contractors have contributed — median and quartiles only, never your rates, your name or your jobs.

What does it cost?

Free to start, no card. Pro is $29/mo and Business is $79/mo. Your subs join your projects free on every plan. Full breakdown on the pricing page.