AI Connectors Need a Job-Folder Cleanup

Connecting AI to company files sounds like a shortcut. For many service businesses, it is really a filing-cabinet test.

Claude's Microsoft 365 connector security guide says the connector can use delegated permissions to read Microsoft 365 services such as Outlook, SharePoint, OneDrive, Teams, and calendar data during active queries. OpenAI's June 2026 Codex help article also points business users toward workspace app controls, role-based access, plugin controls, and compliance logs.

That is useful product progress. It also raises a plain question for a Regina contractor, Saskatoon repair shop, construction subcontractor, equipment dealer, or field-service office: are your job folders clean enough for AI to search?

If the folder is messy, AI will find the mess faster.

Start with the folder, not the model

Most office teams already know where the problem lives. It is the shared drive, the estimator's desktop, the email thread nobody filed, the old quote version in the wrong folder, or the closeout photo sitting in someone's phone.

AI connectors can help staff search across that material. They can also surface files a person technically has access to but should not use for a specific customer job.

Before turning on a connector, pick one workflow where the files matter:

• •quote preparation
• •change-order backup
• •warranty records
• •service history lookup
• •closeout packets
• •supplier communication
• •invoice explanation
• •safety or training documents

Then open five real job folders. Do not clean them yet. Look at what AI would see if it searched today.

You will usually find duplicate estimates, old photos, vague filenames, private customer details, unsigned approvals, and orphaned emails. That is the real setup work.

Write one access rule per job type

The Claude guide is clear that the Microsoft 365 connector uses user-level permissions and that users cannot bypass Microsoft 365 sharing settings or folder permissions. That is good, but it is not the same thing as a business rule.

Microsoft may know whether a dispatcher can read a folder. It does not know whether that dispatcher should use an old site photo when writing a customer update.

A practical access rule says:

• •which folder is the source of truth
• •which files can be used for customer-facing drafts
• •which files are internal only
• •which files need owner, estimator, foreman, or office-manager approval
• •which files should never be sent into an AI tool

For a plumbing company, the rule might say AI can summarize service notes, approved photos, and the final quote. It cannot use private technician comments, unapproved price assumptions, employment notes, or anything from another customer's job folder.

For a construction subcontractor, the rule might say AI can help prepare a draft closeout packet from approved photos, signed change orders, and the latest drawing set. It cannot decide whether work is complete, change the scope, or promise a completion date.

Make old files obvious

AI search gets risky when old files look current.

Many local businesses have a folder that includes:

• •quote-final.pdf
• •quote-final-v2.pdf
• •quote-final-new.pdf
• •a scanned supplier quote
• •an email attachment from two weeks ago
• •photos from before the repair
• •photos from after the repair

A person may understand which file matters because they were in the conversation. AI sees names, dates, content, and permissions. If the naming is poor, the tool may bring the wrong document into the answer.

The fix is not fancy. Use file names that tell the truth:

• •2026-06-22-approved-customer-quote.pdf
• •2026-06-22-internal-estimate-not-for-customer.pdf
• •2026-06-22-before-repair-photos
• •2026-06-22-after-repair-approved-photos
• •superseded
• •internal-only

This is boring work. It is also the difference between a helpful AI summary and a confident mistake.

Keep customer data on purpose

The Office of the Privacy Commissioner of Canada's PIPEDA principles include accountability, identifying purposes, limiting collection, safeguards, openness, and individual access. You do not need a legal lecture to apply the idea.

Before staff connect AI to job records, decide what customer data belongs in the workflow.

A field-service office might need the customer's name, address, service history, warranty record, photos of the equipment, and the approved quote. It probably does not need unrelated emails, payment-card details, a photo that shows a child's bedroom, a staff dispute, or a private note about the customer's financial situation.

The clean rule is simple: only connect the information needed for the job.

That rule matters for offices where everyone wears five hats. The same person may handle dispatch, billing, collections, and customer support. Broad access may feel normal inside a small team, but AI can make broad access easier to misuse.

Use connectors for drafts, not decisions

The Government of Canada's generative AI guide tells teams to verify outputs and avoid using generative AI without proper review when mistakes could cause serious consequences. That maps cleanly to service and construction work.

Good first uses for connected AI:

• •summarize a job folder for the estimator
• •draft a customer update from approved notes
• •list missing closeout documents
• •compare a work order against required photos
• •pull together supplier communication for a quote review
• •prepare a warranty packet for human approval

Bad first uses:

• •approve a warranty claim
• •decide final price
• •promise a completion date
• •interpret a legal term
• •change a safety procedure
• •send a customer commitment without review

Use AI to prepare the work. Keep the decision with the person who owns the customer, the crew, the price, or the risk.

Run a small permission test

Do not turn every connector on for every staff member first.

Start with one role. An office manager is often the best fit because the work crosses dispatch, billing, customer updates, and job records.

Give that role one task:

• •"Find what is missing from this closeout packet."
• •"Draft a customer explanation from the approved notes."
• •"Prepare a quote-review summary from this folder."
• •"List the documents needed before billing."

Then check three things.

First, did AI use the right files? Second, did it skip files it should not use? Third, could the reviewer tell where the answer came from?

If the answer is no, fix the folder before adding more people.

Where Prairie AI can help

This is the kind of AI setup that pays off before the business buys more tools. Prairie AI can help map the workflow, sort the files into practical source-of-truth rules, write the staff permission card, and build a connector pilot around one real job folder.

If your team wants connected AI but the shared drive is already messy, book a call with Prairie AI. We can pick one workflow, clean the handoff, and decide what AI is allowed to read before it starts answering.

For related help, see AI automation services, AI help in Regina, AI help in Saskatoon, and AI help across Saskatchewan. If you would rather send the problem in plain language first, use the Contact Prairie AI form and describe the folder, inbox, or job record that slows the office down.

What I would do this week

Pick one job type that creates rework. Choose a quote packet, warranty folder, closeout packet, or service-history lookup.

Open the last five examples and mark four things: the source-of-truth folder, the files AI can read, the files AI must ignore, and the person who approves the final customer-facing output.

Then test one AI question against one folder. If the answer pulls from stale, private, or confusing files, that is useful evidence. The first job is not to write a better prompt. The first job is to clean the folder.