A Layman's Guide to AI Software

Canaveral was not the first company to think of adding AI capabilities to estimating software, and we certainly won't be the last. As the proliferation of AI begins to saturate the software market for all sectors of the economy, AI will be part & parcel of any modern software offering. So what happens then? How do you decide what tools are best for your team?

Test 1: Unwrap the wrappers

A good place to start is learning what kinds of products to avoid. A pattern we've been noticing in construction software products is becoming quite frequent: a single or batch of AI chat features at its core, with a thin UI wrapper around it, a.k.a. an "LLM wrapper". Often times the AI features in these kinds of products are simply a middleman between the user and one of the popular models like ChatGPT or Claude. Users of these thin wrappers are essentially paying extra for glorified LLM routers with an "estimation software" label. As AI for software engineering itself continues its expansion, we expect these kinds of products to grow in quantity.

Common tells for this kind of architecture are chat-only interfaces, where you upload a file and receive results back in text format. These kinds of products are entirely probabilistic, where you sacrifice the accuracy and predictability of your estimates to the altar of convenience. It makes for great demos, but you shouldn't build a bid off of it.

Test 2: AI output quality

If a product passes the LLM wrapper test, the next test is AI output quality. Products that reach this level are often more useful than their wrapper contemporaries as they may implement different & more advanced AI technologies like computer vision models to predict objects, or AI chat features that actually pre-process context from your documents before sending over to a model provider. This is undoubtedly a step in the right direction, but it's important to look at the output.

Are you passing in your plans to a "black box" and getting back a long list of data, only to be required to pore over the plans to double-check for due diligence? After all, if a tool decreases your takeoff time but increases your validation overhead, then the advertised time-saving benefits are essentially a wash. Products that pass this test will not only produce predictions or insights, but they will be able to tie them directly to your plans in context, whether visually or textually.

Test 3: Real-world data requirements

Finally, if a product passes the first two tests, the final test is one of answering the core question: does the output accurately capture the real-world data you require for a quality estimate? AI takeoff predictions can be extremely accurate at products of this tier when it comes to object identification and list-making, but if they are limited to simple bounding boxes or generic object titles, that still won't get you closer to a bid.

The winners of this tier will not be AI-centric, but data-centric, built on a solid product foundation that captures all the minutiae of construction that is implicit but important:

• Does this software capture how many hangers we need? Nuts? Bolts?
• How many square feet of insulation? Fireproofing?
• For pipes, can we tie the pricing to real-world parts in real-time?
• Are we capturing reinforcement apparatus for ductwork?

...and the list goes on and on. If an estimation product cannot answer these questions in the affirmative, it does not pass this test.

In short

• Pass on the LLM wrappers
• Make sure the product's AI output is delivered in-context, verifiable, without requiring more validation work
• Ensure the product can capture the robust data requirements you need to make an accurate bid for the real world

We've spent years thinking about this at Canaveral. Where other competitors are building an AI-centric offering, we're flipping the script, and have built the best manual takeoff and estimation tool for mechanical estimators, even without AI. Then, once we nailed that experience, we built useful and context-aware AI tools injected into your existing workflows, ensuring the human always stays in control, but is accelerated to the "speed of thought" as we like to say.

AI centric vs Human centric; we have our opinions on which philosophy will win in the end. We'll see what the estimators decide.