The owner said it on one of our first calls, almost apologizing:

"I pay somebody almost full time to get these four pieces of information."

Four details. The general contractor, the project, the bid due date and the building's square footage. Pulled from every bid invitation, typed into a spreadsheet, the email filed, the attachments saved to a project folder.

About 150 times a month.

I hear a version of this in almost every company we talk to. The AI project that pays off first is almost always the one that gives a person their day back.

Why nobody had fixed it

The company is a specialty construction contractor in the United States, somewhere between 10 and 50 people. General contractors invite them to bid on their part of a project, mostly by email, into one shared inbox.

On paper, the job sounds easy to automate. In practice, every sender uses a different format. Some put the due date in the subject line, some bury it in a PDF, some send a link to a portal. Square footage means measuring from plans.

So the job stayed manual. One person, all day, every day.

What we built

Nothing fancy. We built on top of the inbox and the file storage the team already used.

Every 10 minutes, the system checks the inbox. It sets aside the emails that aren't bids (reminders, newsletters, replies). In a sample of 200 inbox emails, about 4 in 10 were screened out that way.

For each real invitation, AI reads the email and its attachments and pulls out the contractor, the project, the due date, the site address, the dates and the contacts. It flags jobs with special wage rules. It finds or creates the right project folder. Then the invitation shows up in a simple queue, with a link back to the original email.

The team kept working the way they already did, with a cleaner version of the same inbox.

What stayed human

People stay in charge.

The team checks every record. They fix anything missing and mark it completed or cancelled. Square footage is still measured by a person, by their choice, because most plans sit behind secure links. We could have pushed for it. It wasn't worth the risk.

And the records aren't perfect. The contractor shows up in 98% of them. The bid due date in 74%, because some invitations simply don't include it. The team knows what to check, and checking a ready record takes a couple of minutes instead of seven.

What happened after launch

The system went live in March 2026. Since then it has logged 979 bid invitations, and the team has closed out 93% of them in the new queue.

At launch, the owner told us: "your team can manage the whole thing." So we do. We watch it, fix it and improve it.

Twice, in August and October, the team reported emails that seemed stuck. Both times we found the cause and shipped a fix the same day. In between, small things kept getting better: editing names in the queue, more rules to screen out non-bids, documents filed into each project folder without overwriting anything.

None of that is exciting either. It's what keeps a system in use a year later.

What it saves

Our estimate is about 14 hours a month saved, after the time the team spends checking each record. Realistically it's somewhere between 8 and 24, depending on the month.

I'd rather tell you that than round it up. The bigger change is harder to count: the person who used to type all day now spends that time deciding which bids are worth chasing.

Find your "four pieces of information" job

Every company has one. Look for work that:

  1. Happens every day, the same way.

  2. Moves information from one place to another.

  3. Somebody quietly hates.

That's your first AI project. Boring, specific, and worth real hours from week one.

If you want help finding it, here's how we build, and the full case study with every number. Or book a 30-minute call and tell me which job came to mind while you were reading.

Aldrey