Independent evaluation of construction AI.
We test AI agents built for construction project controls against real schedules and real source data, check whether the same question gets the same answer twice, and publish what we find. Starting with scheduling agents.
Latest run · Oct 2026
Total matched, 8 categories
We don't sell AI scheduling tools.
Every result is checked against the source schedule, not just against the vendor’s own answer.
Methodology published before any result is.
Vendors see their own results first, and can respond before anything goes public.
First place, AI in Project Management Hackathon, Global Project Management Forum, Riyadh 2026.
How we test
The same discipline we used on our own tool, applied to every agent we evaluate.
01
Start from the source file
We take a real Primavera P6 schedule and compute the correct answer directly from the raw XER (calendar-aware duration and float, DCMA-14 checks, relationship counts), independently of any AI.
02
Ask the same question many ways
The same underlying question, phrased several different ways, should get the same answer. When it doesn’t, that’s a finding, not noise.
03
Grade against ground truth
Numbers, dates, and activity IDs are graded automatically against the source data, exact match or not. Open-ended answers are graded against a rubric a project-controls professional writes in advance, and every grade is confirmed by a person before it’s published.
04
Share results with the vendor first
Before anything is published, the vendor sees the findings and has a chance to respond. Public results, when we publish them, cover general-purpose models, not named commercial products, unless a vendor chooses to publish their own result.
Starting with scheduling agents
Scheduling is where we have the deepest domain background: 1.5 years spent building a deterministic schedule-analysis engine ourselves, and the specific experience of finding real, non-obvious correctness bugs in our own tool by checking it against raw P6 data. More categories (cost, risk, contracts) may follow as the methodology holds up.
First results: three OpenAI models
Each model, given read-only query access to a real Primavera P6 schedule and no other help, was asked 25 scheduling questions across 8 categories, each phrased three to four different ways and graded exactly against the raw XER.
3
Models tested
8
Categories tested
25
Questions per model
12/25
Best score
WBS element count
Critical path activity count
Finish-to-start relationship mix
Missing predecessor/successor logic
Longest activity by duration
Zero-duration activity count
Hard constraint count
High-duration activity count
Total matched
Both newer models fixed the finish-to-start relationship question that GPT-4o got wrong, and improved on “missing logic.” But all three models still score zero on hard-constraint counting and high-duration activities. A newer model isn’t a uniform upgrade: it can close one gap while leaving another untouched.
Two further categories (negative float, cost performance index) were run against all three models but excluded from this result: one is limited by a known gap in our own test setup rather than any model’s answer, the other requires cost data our test tool doesn’t yet expose. Both will be added back in a future run.
Questions, answered plainly.
What is Schedira now?
Schedira independently evaluates AI agents built for construction project controls. We started as a company building one of those agents ourselves; we now test the category, starting with scheduling agents.
Are you actually independent? You used to build a competing product.
That's the right question to ask. We're winding down our own scheduling-AI product specifically so this evaluation work isn't run alongside a competing one. Our background building and testing a scheduling engine ourselves is why we can tell real correctness problems from noise, not a conflict we're hiding.
How do you check an AI agent’s answer?
We compute the correct answer directly from the source P6 file (calendar-aware duration and float, DCMA-14 checks, relationship counts), independently of any AI, then compare the agent's answer against that ground truth. Numbers, dates, and IDs are graded automatically; open-ended answers are graded against a rubric a project-controls professional writes in advance, and every grade is confirmed by a person before it's published.
Do vendors know they’re being tested?
Yes. Vendors see their own results before anyone else does, and have a chance to respond before anything is published. We don’t run undisclosed tests.
Which AI tools do you evaluate first?
We’re starting with AI agents built for construction scheduling, where we have the deepest domain background. Public results, where we publish them, cover general-purpose models rather than naming commercial products, unless a vendor chooses to publish their own result.
How is this funded?
This is an early pilot, not a funded company. We're finding out whether this is useful to anyone before building more than a methodology and a first round of results.
I build a scheduling AI agent. How do I get evaluated?
Write to us. We’ll walk through the methodology, and you’ll see your own results before anyone else does.