What AI implementation actually looks like

In asphalt paving, the most expensive decision of the day happens around 5:30 in the morning: do the crews roll or not? Get it wrong one way and you've dispatched trucks, fired up a plant, and put a crew on the clock for a job the rain shuts down by nine. Get it wrong the other way and you've parked a full crew on a day that turned out perfectly paveable.

That decision is where our platform for a major South Jersey asphalt paving company starts. This article walks through the three tools we built for them — what each one is, how it works, and why it's built the way it is.

Tool one: a paving call, not a weather report

Every morning, before dispatch, each field manager gets a text message for their specific job site: the conditions that matter for laying asphalt — temperature, precipitation, wind — and a plain, emoji-coded rating on whether it's a paving day.

Under the hood, the system pulls from three independent weather sources — a government feed and two commercial APIs — and reconciles them metric by metric. Temperatures take the median of the three. Precipitation takes the most conservative reading, because the cost of a surprise shower beats the cost of caution. The paving rules themselves — minimum temperatures, precipitation thresholds — are configured per job by the company's own admins, because a highway job and a parking lot have different tolerances.

Why a text message? Because field managers live in their text threads, not in apps. The average weather app answers "what's the sky doing?" This answers the only question that matters at 5:30 AM: "can we pave today?" One consistent call in the same format every morning, instead of three managers checking three different apps and arguing from three different forecasts.

Tool two: ninety seconds at the truck, live numbers at the office

At the end of each shift, the foreman gets a link by text. It opens a short form — tonnage placed, crew hours, equipment used, any issues — and most of it is already filled in. Crew and equipment carry over from the job's roster and yesterday's answers; the foreman confirms, corrects, and submits from the driver's seat.

Those submissions feed a KPI dashboard that leadership actually checks: production by job, trends over time, which jobs are ahead and which are bleeding hours.

The interesting engineering choice is what's missing: logins. Foremen never create an account. Each day's link is cryptographically signed, single-use, and expires after 24 hours — the security lives in the link, not in a password someone forgot in a truck. The first submission for a job locks the day; anyone tapping the link afterward sees a read-only view. Miss the form and the system nudges: a reminder in the evening, an escalation to the office after that.

The status quo this replaces is the industry default: paper forms in a glovebox, numbers texted to the office, and someone stitching spreadsheets together at month-end — when it's too late to fix anything. Data quality doesn't come from policing the field. It comes from making the form so fast there's no reason to skip it.

Tool three: the plant report the CEO doesn't have to ask for

The third tool is a daily plant-performance report for the CEO: what the asphalt plant produced, at what rate, against what it should have produced — delivered every day without anyone compiling a thing.

The engineering decision worth stealing here: we didn't write production code against the plant's own software until we'd proven the integration. A one-week spike connected to the plant system and confirmed the data path first. If that spike had failed, we'd have redesigned around what was actually available — instead of discovering the problem three weeks into the build.

A surprising amount of what gets called "AI implementation expertise" is exactly this: the discipline to not build the exciting thing until the boring dependency is verified.

So where's the AI?

Here's the part that surprises people. In version one of this platform, the AI is small: Claude drafts the executive summaries on reports and powers a handful of admin tools. The monthly AI bill is dollars, not thousands. Nothing about reconciling weather feeds or moving field data requires a large language model, so no large language model was invited.

The buzzword version of this project would have bolted a chatbot onto the dashboard and called it AI transformation. The next phase holds the genuinely ambitious AI piece: an extraction engine that reads the messages crews already send — "placed 480 tons, wrapped by 3" — and turns them into structured records with a validation loop, no form required. We built a first version of it early on, then deliberately held it back, because it only pays off on top of a foundation people already trust: clean job data, reliable delivery, a dashboard leadership actually checks.

That's the expertise, honestly stated: not knowing how to add AI, but knowing the order in which to add it — and having the restraint to sequence it.

The principles underneath

Wondering what the equivalent build looks like for your operation? That's what the free Bottleneck Audit finds: the decision or routine that eats the most of your week, and what it would take to hand it to a system. No pitch — you leave with a plan.