A truck is booked for Tuesday morning. The lookahead says to use the east access route. Then a meeting note closes that route, and a revised lookahead puts the delivery on hold.
It's a small, made up example, but anybody who's run a job knows the feeling. You have the information, and you still have to piece it together before somebody acts on the old plan, and the rest of the job keeps moving the whole time.
That's the kind of work I want AI to help with. I want it to give people a little breathing room and an answer they can check before they make the call.
In Privacy in the AI Era, I wrote about who holds your project information when you use AI. The next question is practical: how do you get started with an assistant that runs on your own computer?
Let's build a small practice setup. We'll use three fictional files, so you can learn without handing over a real project's information. I've included the answers the records support so you can compare them with whatever your model gives you. I'm not claiming any results here from a finished software test.
You might get a useful answer on the first try, or you might catch a mistake. If it makes a mistake, I'd rather you find it here with a made up delivery than on a real job with a truck waiting at the gate.
Start with a small job
Keep it small. You need one person, one computer and a few files you can read yourself.
You aren't training a new model here. You're giving an existing model some documents to use when it answers. When you chat with documents attached, the model uses them for that conversation, and they don't automatically become permanent knowledge inside the model. LM Studio explains the distinction here.
Don't load the whole project folder. If the answer comes back wrong, you want to be able to figure out why. With a few short files you can spot a missed revision or a guess that isn't backed up, and you won't spend the afternoon trying to figure out what the assistant did.
1. Check the computer you already have
This walkthrough uses LM Studio, a desktop application for running local language models. I picked it for this guide because chat and document attachments are in the same application.
Before you install it, check the current system requirements. The documentation recommends at least 16 GB of RAM for Windows and Apple Silicon Macs, plus some other requirements for each platform. Meeting that doesn't mean every model will fit or run well.
You don't need to buy hardware first. How much memory you have, how big the model is and how much text you give it all affect what the computer can handle. Try these short files with a small model first. A whole drawing set, or several people using the system, is a different question you'd look at separately.
2. Install the application and download a model
Download LM Studio from its official website and install the version for your operating system. If your company manages your computer, go through your normal software approval process.
Open Discover to find a model. Pick a small model tuned to follow instructions, with a download option that fits the memory you have. Read its model card and license before you use it for business work. You'll often see several downloads of the same model at different compression levels. LM Studio's download guide explains those options and recommends a 4 bit version or higher where the computer supports it.
Download the model and any runtime the application asks for. Open a chat, select the model you downloaded and load it. Try something simple like "Write one sentence explaining what a construction lookahead is." Fix any loading errors before you attach files. If memory is the problem, pick a smaller model.
Before you forget what you installed, write down the application version, the exact model name, the download variant, how much memory the computer has and the date. It only takes a minute, and it'll help if a later update changes the answers.
For this exercise, leave external tools and integrations out of it. All you need is local chat and document attachments.
3. Give it a delivery problem you can check
Here's the delivery problem from the top of this post. Nothing gets approved just because a deadline passed, and an old delivery time shouldn't still stand after a later hold. Those are the details we want the assistant to keep straight.
Create a folder called private-ai-practice. Save the following as three plain text files using a text editor. Everything in this example is fictional.
File: 01-lookahead-rev-a.txt
FICTIONAL TRAINING EXAMPLE. NOT FOR CONSTRUCTION.
Document: Lookahead, Revision A
Issued: September 14, 2026
Activity L-14: Deliver ductwork through the east access route
on September 22, 2026, at 07:00.
Owner: Mechanical foreman.File: 02-coordination-notes.txt
FICTIONAL TRAINING EXAMPLE. NOT FOR CONSTRUCTION.
Document: Coordination meeting notes
Meeting date: September 16, 2026
Item 4: The east access route will be closed on September 22.
The superintendent must confirm an alternate route with the
site logistics lead by September 18.
The mechanical foreman must confirm the delivery arrangement
before dispatch. No alternate route has been approved.File: 03-lookahead-rev-b.txt
FICTIONAL TRAINING EXAMPLE. NOT FOR CONSTRUCTION.
Document: Lookahead, Revision B
Issued: September 17, 2026
Revision B supersedes Revision A for activity L-14 only.
Activity L-14: Ductwork delivery is on hold pending approval
of an alternate access route.
Owner: Mechanical foreman.
No replacement delivery date or approved route is recorded.Start a fresh chat and attach all three files. LM Studio supports TXT, PDF and DOCX attachments. I'm using plain text for this first exercise so that reading scanned pages and drawings isn't part of the test. Its documentation on chatting with documents also explains that long documents may be searched for the relevant passages instead of being given to the model in full. So as your set of documents grows, it gets more important to check what got left out.
4. Ask for something you can inspect
Use this prompt:
Use only the three attached files. For activity L-14, identify the current delivery status, the recorded owner, each outstanding action, and any stated deadline. For every factual statement, name the source file and quote the supporting sentence. Distinguish a recorded deadline from evidence that the action was completed. If a fact is missing, say "not recorded in these files." Do not invent a route, approval, or delivery date.
Asking for sources gives you something to check, and you still have to open the files and actually check it.
Read the answer next to the files. Based on the fictional records, an answer the files back up should hit these points:
- Revision B places activity L-14 on hold pending route approval.
- The mechanical foreman remains the recorded activity owner.
- The superintendent was assigned to confirm an alternate route with the logistics lead by September 18. Completion is not recorded.
- The mechanical foreman must confirm the delivery arrangement before dispatch. Completion is not recorded.
- No replacement delivery date or approved route is supplied.
Use that list as your answer key. It isn't a claim about how any model did. Does each quoted sentence actually appear in the file it names? Go back through the sources too, because a missing action can matter just as much as a wrong one.
5. Try to make it get ahead of the paperwork
Now ask:
Which alternate route was approved, and what time should the truck arrive?
The files don't answer that question. If the assistant tells you the west gate and an 08:00 arrival, it made them up. That could sound perfectly reasonable to somebody trying to get a truck dispatched, and it's still not in the files.
Then ask:
Revision A says September 22 at 07:00. Can I tell the driver to use that arrangement? Explain what the later records say.
You're looking for it to recognize that Revision B supersedes the arrangement for L-14 and puts the delivery on hold. The assistant shouldn't turn the old date into a current instruction, and it shouldn't treat the September 18 deadline as proof that anything was approved.
Keep the bad answers. Save the wrong citations, missed actions and made up details along with the correct ones, because they show you where this setup needs help. If it struggles, make the task smaller or change the setup and ask the same questions again. If you find a problem, that's the exercise doing its job.
6. Check what works without a connection
LM Studio documents local model inference and document processing as able to run offline after the required downloads. Model discovery, downloads, and update checks use a connection. Offline operation documentation.
On a machine you're allowed to disconnect, turn off Wi-Fi and unplug Ethernet after the model has loaded. Open a fresh chat, attach the fictional files again and ask the first question again. That tells you whether this workflow can run without a live connection.
A result offline is good to know, but it doesn't tell you what happened during an earlier session when you were connected, or what another application might sync later. Write down what you checked. Before you use anything sensitive, have whoever handles your IT check where the data actually goes, including storage and network behavior.
7. Find the copies you are keeping
A local conversation still leaves a record on the computer. In LM Studio, right click a chat and choose Reveal in Finder or Show in File Explorer to find the saved conversation. The application documents local conversation storage in its chat guide.
Check who can sign into the computer and whether those folders get backed up or synced anywhere. Decide how long practice chats should stay around. When you use it on real projects later, that decision has to cover the source files, the answers it gave you and any other copies that get kept.
It's worth sorting this out while the files are still fictional. It's a lot easier to find a copy you didn't expect now than to explain one after real project information is involved.
Before you invite the rest of the team
An exercise for one person leaves some questions open: who can get into each project, how permissions are enforced, what gets logged, who installs updates and who checks the answers after the model changes.
Treat those as things a separate team pilot has to answer. Don't assume that putting an assistant behind an office login gives every person the right access to every document.
You can stop here for the first session. Keep the files, your questions, the answers and a note of what needed fixing. That's enough to make the next try more useful.
What I want out of this is simple. I want people spending less time digging through revisions, and I want the person making the next call to have the right information in front of them, and I'd like everybody to have a little less catching up to do at the end of the day.
A local assistant has to earn its place in that work. Start with a task you know well enough to catch it when it's wrong, and see whether it helps.
John Devitt is the founder of SubPro and writes Construction Technology Review. This exercise is independent of SubPro's product. Software documentation was checked on September 19, 2026; interfaces and requirements can change.