Junction

How to prepare your job-cost data for AI: a 7-step checklist

By the Junction team · Published 24 October 2026 · 3 min read

The short answer

In Dodge and CMiC's research on contractors and AI, only 26% of contractors rated their data as high quality, according to the published release. That is the usual reason a first trial disappoints: the assistant is fine, but the data it reads has quirks that everyone in the office knows about and nobody wrote down.

This checklist finds those quirks before the assistant does. Each step takes an hour or two for a controller or an accounting lead.

1. Make job numbers consistent

List every open and recently closed job with its number, name and status. Look for duplicates (the same job entered twice), numbers reused after a job closed, and jobs whose names differ between accounting and project systems. A shared, unique job number is the key that lets answers from different systems line up.

2. Agree what "open" means

Most systems have more than two statuses (open, soft-closed, closed, and sometimes on-hold or warranty). Decide which count as open for budget questions, and write it down. Without this, "all open jobs" will return a different total from your usual report.

3. Check phases and cost types

Questions like "which jobs are over budget on labor" depend on cost types being used the same way on every job. Spot-check a dozen jobs: is subcontract cost sometimes coded to material? Are equipment charges under their own type? Fix the worst offenders or note them.

4. Tidy the vendor list

Duplicates in the vendor master split spend and aging across entries, so "what do we owe this vendor" gives a partial answer. Merge or flag duplicates, and mark inactive vendors.

5. Review commitments and change orders

Committed cost is one of the most useful measures and one of the easiest to get wrong. Check that subcontracts and purchase orders are entered, that approved change orders are applied to the commitment and not left pending, and that closed commitments are really closed.

6. Decide who may see what

  • Keep payroll and personal records out of reach of any general-purpose assistant. Job cost rarely needs them.
  • Limit people by system, kind of data and job. A project manager needs their jobs, not the company's whole ledger.
  • Use a database login or API key that can only read for anything that connects.

Our security checklist covers what to ask of any connector on these points.

7. Write ten test questions with known answers

Choose ten questions you can answer from reports you already trust: a few totals, a few lists, a few comparisons. Record the correct answers. When you connect an assistant, ask them first. If a result differs, you will know whether it is the data or the connection. Start from our 40-question cheat sheet.

When you are ready to connect, see how to connect Claude to your construction accounting data, or request early access to set it up with us.

Frequently asked questions

Do I need clean data before using AI with my accounting system?
You do not need it to be perfect, but you should know where it is weak. Check job numbers and statuses, cost types, vendor duplicates and commitments first, because those drive the most common wrong answers.
What is the most common data problem in job-cost reporting?
Inconsistent definitions: which jobs count as open, which cost types are used for what, and whether approved change orders have been applied to commitments. These produce totals that differ from your usual report.
How do I test an AI assistant's answers about my accounting data?
Write ten questions whose answers you already know from trusted reports, ask them first, and compare every figure. If one differs, ask the assistant which system and rows it used.

Sources

Ask Claude or ChatGPT about your own systems

Junction connects Claude and ChatGPT to construction accounting, project and property systems. Early access, set up with you.