Turnaround and shutdown contractor · Baytown, TX
Billing a plant turnaround in six days instead of five weeks
A Gulf Coast turnaround contractor produced 2,800 tags in a three-week outage, then spent a month billing them. The bottleneck was not the volume.
Representative engagement. This describes work of a type we build, with figures modeled rather than measured at a named client. We would rather tell you that than invent a customer you cannot phone.
Outcome Measured after go-live
- 34 → 6 days
- From outage close to final invoice
- Measured across the two turnarounds following go-live.
- 22% → 7%
- Tags requiring rework
- Incomplete, unsigned, or missing a cost code on first submission.
- $1.4M
- Working capital released
- Average receivables reduction across the turnaround cycle.
The situation
A turnaround contractor running planned outages for refiners and chemical plants along the Houston Ship Channel. Four to six major turnarounds a year, each running two to four weeks with up to 400 craft on site.
The economics of turnaround work are unusual. Revenue arrives in enormous concentrated bursts, and the billing that follows an outage determines the company’s cash position for the next quarter. They were closing an outage and then taking another five weeks to finish invoicing it.
What the audit week found
The obvious hypothesis was volume: 2,800 tags in three weeks is a lot of paper. That was wrong, and it was worth being wrong about early.
The actual constraint was that nothing could be billed until everything was verified, and verification could not start until the outage ended. Tags accumulated in boxes for three weeks, then four people spent a month reconciling them against cost codes, scaffold tags, and the owner’s daily work packages.
Three things made that reconciliation slow:
No signal in the unit. Any digital system already tried had assumed connectivity. Inside a process unit there is none, so crews reverted to paper within days and the system became shelfware. This had happened twice before we arrived, which made the client reasonably sceptical of us.
Cost codes were assigned after the fact. A foreman writing a tag at 2am did not know which of 60 cost codes the owner wanted it against. Somebody in the office guessed later, and the owner rejected roughly one in five guesses.
Scaffold tags and work tags lived in different worlds. Scaffolding was tracked by a separate crew on a separate sheet, and marrying the two happened manually at the end.
What we built
Capture that assumes no connectivity. Rugged devices, full offline operation, sync at the gate or on the bus. A foreman can work a twelve-hour shift underground and the device does not care.
Cost code suggestion at the point of entry. The system proposes a code from the work description, the location and the owner’s work package, and the foreman confirms or overrides. It is right most of the time and, more usefully, it flags when it is not confident rather than guessing silently.
Daily rolling billing packets. This was the change that mattered. Instead of waiting for the outage to end, verified tags are packaged and submitted daily against the owner’s work packages. By the time the outage closes, most of the work is already billed.
Scaffold reconciliation. Scaffold tags and work tags are matched automatically by location and date, with exceptions surfaced for a human.
What happened
The headline number, 34 days to 6, understates what actually changed. The company stopped financing a month of a turnaround out of its own balance sheet four times a year. Their line of credit usage fell enough that their bank asked what they had done.
Rework fell from 22% of tags to 7%. Most of the remaining 7% is genuine ambiguity that a human should be looking at anyway.
What went wrong
The first device rollout failed. We specified devices that were rugged enough but had a screen nobody could read in daylight with safety glasses on. That cost two weeks and a device swap. The lesson generalises: specify hardware with the people who will hold it, in the conditions they hold it in.
Cost code inference was worse than expected on the first turnaround. It ran around 71% accurate against a target of 90%, because the training examples we had came from one owner and the first live outage was for a different one. It improved as it saw more work, but the client had to live with a rough month first. We should have said that more clearly in advance.
Two superintendents never adopted it. They ran their crews on paper through the whole first turnaround. That resolved itself when the billing team started returning their tags for rework while everyone else’s went through clean, but it took a full cycle and it was uncomfortable.
What it cost
| Workflow Audit | $7,500, one week |
| Build | $86,000, seven weeks, fixed price |
| Care and feeding | $7,000 a month |
| Devices | 24 units, purchased directly by the client |
What it was built with
- Offline-first capture on rugged Android devices
- Postgres and object storage in the client’s AWS account
- Claude for extraction and cost-code inference
- Viewpoint Vista integration