Most guidance about AI in construction ends at the same instruction: keep confidential material out of public AI tools. That is reasonable advice and it is close to unusable if you coordinate crews, because the crew roster, the certification list and the schedule problem you need help writing up are all confidential, and they are also the exact things you want help with.
Table of Contents
ChatGPT has two settings that make a bounded version of this workable, and because they do different jobs and live in different places, most people who set up the first one assume they have also handled the second. This walks through what a Project is, how project memory actually separates one general contractor’s information from another’s, where the training setting sits and why it is a separate question, and how to structure projects when you are working for four contractors at once.
What a ChatGPT Project is
A Project is a container that holds related chats, uploaded files and its own instructions together, so work inside it starts with the files already attached and the rules already set.
Three parts matter for crew work: project files, which are documents any chat in that project can reference; project instructions, the rules that apply inside that project; and project memory, which governs what ChatGPT carries between conversations and is the part doing the confidentiality work.
Capacity is generous enough that structure rather than storage becomes the constraint, and OpenAI’s documentation on projects puts file limits at 5 per project on Free, 25 on Go and Plus, and 40 on Edu, Pro, Business and Enterprise, with a maximum of 10 files per simultaneous upload. The number of projects you can create is unlimited, which is what makes the one-project-per-contractor approach below practical.
The two settings people confuse
Almost everything written about AI confidentiality collapses these two controls into one, which is how a coordinator ends up flipping the first, assuming the second is covered, and solving half the problem without knowing it.
| Project memory | Training setting | |
|---|---|---|
| What it controls | What ChatGPT carries between your own conversations | Whether OpenAI may use your content to improve its models |
| What it stops | The GC-A project drawing on anything in the GC-B project, or on your general chats | Your material contributing to model training |
| What it does not do | Anything about training | Anything about context bleeding between projects |
| Where it lives | Project settings, per project | Account settings, under “Improve the model for everyone” |
| Default on Business, Enterprise, Edu | Set to project-only automatically when a project is shared | Project content not used for training by default |
| Default on Free, Plus, Pro | Your choice per project | Content may be used if the setting is on |
| The detail worth knowing | Sharing a project turns project-only memory on automatically, and shared projects cannot reach any member’s outside context | For shared projects, OpenAI trains only if every contributor and the owner has the toggle on. One person with it off protects everyone |
That last row is the most useful fact in this piece and it appears almost nowhere else. On a shared project, one person’s setting protects the whole thing.
Setting up projects for multi-contractor work
1. Build one project per general contractor rather than one for everything
The instinct is a single project called “work.” Given that projects are unlimited and memory is bounded per project, the better structure gives each general contractor its own, which keeps their information separate by design rather than by discipline.
Two situations complicate this and are worth deciding before you build. A job with a joint venture or a construction manager between you and the owner is genuinely a job-level project rather than a contractor-level one, and scoping it to the job is the honest answer. Your own internal planning, the roster and certification picture across everything you run, belongs to you rather than any client, so it gets its own project and is the one most likely to hold personal data about your people.
2. Set project memory to project-only
Inside each contractor project, set memory to project-only, after which conversations draw context solely from that project and cannot reach your general chats or anything held in a different project.
If you share the project with a colleague, this happens automatically. OpenAI is explicit that sharing turns project-only memory on and that shared projects have no access to any member’s context or memories from outside the project, which is a sensible default and one most people would not predict.
3. Check the training setting, which is a different place entirely
Open your account settings and find “Improve the model for everyone.” On Business, Enterprise and Edu, OpenAI states that information accessed from projects is not used for training by default. On Free, Plus and Pro it may be used, and whether it is depends on that toggle.
This takes about a minute once you know it exists, and it is the step almost everyone skips because they believe project memory already covered it.
4. Write instructions that name your terminology and your limits
Instructions inside a project let you state once what you would otherwise repeat weekly, and the useful ones combine your vocabulary with explicit refusals.
Example project instructions: “You help me coordinate crews for a mechanical contractor working as a subcontractor to this general contractor. Their manpower report goes out Thursdays in the format in the attached example, hours grouped by phase. A crew is a foreman plus three to five installers moving between jobs as a unit. Never estimate an hours figure I have not given you. Never state that a certification is current unless it appears in the attached list. If information is missing, ask rather than assume.”
The internal planning project needs different instructions, because its job is comparison across contractors rather than work for any one of them.
Example instructions, internal planning project: “You help me see our whole labor picture across every job we are running. Do not draft anything client-facing here. When I ask where we are short, answer by crew and by week, name the jobs competing for the same crew, and state which certifications gate the scopes involved. Never merge this with anything from a client project. If the roster attached here is older than two weeks, say so before answering.”
That last instruction is doing something specific and worth copying. Asking the assistant to flag the age of its own source material will not make the file current, though it does put the staleness question in front of you at the moment you are about to rely on it.
5. Load the files that repeat, and leave out what should not be there
Attach that contractor’s standing job facts and site constraints, the report format they asked for, the phase sequence, the crews assigned to their jobs, and the certifications gating those scopes. Tracking who has what certification and when it lapses belongs in your system of record rather than a chat window, though a project that knows which qualifications a scope requires will stop proposing a crew that cannot legally perform the work.
What stays out is anything whose exposure would be a genuine problem regardless of any setting: personal data beyond what planning requires, anything under a specific confidentiality undertaking, anything you would struggle to explain having uploaded.
6. Test each project before you trust it
Give a new project a situation you already handled and compare its output against what you actually sent. You are checking three things: whether it holds the format, whether it invents anything, and whether it correctly refuses the things you told it to refuse.
Example test prompt: “The level 4 rough-in on the Halston job moved up two weeks. Draft the manpower notice in our Thursday format, flag any crew this puts in conflict, and tell me what you had to assume.”
The refusal test matters most and is the one people skip. Ask it something it should decline, like requesting an hours estimate you never provided, and confirm it says so rather than producing a confident number. An instruction set that has never been tested against the thing it is supposed to prevent has not really been tested.
Where this falls short
A configured boundary is a design intent rather than a guarantee, and it is worth holding that distinction. Project memory is not encryption and it does not constitute a compliance position, so a contractor carrying contractual data-handling obligations to a general contractor needs the right plan tier and the terms that come with it rather than a toggle somebody switched.
The structure also needs maintaining once it exists, because jobs finish and a project full of a completed contract’s material is worth archiving rather than leaving open indefinitely, which means somebody has to own that on a schedule or it will not happen.
The same question is worth asking of every system holding this material rather than only the AI one. The security of the software you rely on gets assessed on its terms, its certifications and its defaults rather than on a setting an individual user remembered to change. A tool that already knows your crews and your jobs should keep each contractor’s picture separate as a property of how it is built, which is one of the things that distinguishes purpose-built AI workforce planning from a general assistant somebody configured carefully on a Tuesday.
Opinion: the prohibition was never going to hold
The standard advice, keep client data out of public AI tools, has the shape of a rule nobody follows. Specialty contractors working across several jobs need help with exactly the material the rule excludes, so what actually happens is that people use the tool anyway, in whatever configuration their account shipped with, and stop thinking about it.
A bounded setup is a better outcome than a prohibition that gets quietly ignored, and two settings plus a project per contractor takes twenty minutes while putting a real boundary where there was previously an unexamined default. It is not a compliance programme and nobody should present it as one, but it is considerably better than the honour system most coordination teams are running on right now.
