Automate Back-Office Work: AI Assistants, Local AI or Plain Scripts

Path A: A Commercial AI Assistant With Data Controls

Assistants such as Claude, ChatGPT and Microsoft Copilot are the fastest route to results, especially for judgment-heavy work. They increasingly connect to your mailbox, files and calendar, which is what makes them useful and what makes the plan you choose matter: consumer and business plans are not the same product. Before anyone uses one for work, check these points against the vendor's current published terms:

• Training on your data. As of November 2026, Anthropic, OpenAI and Microsoft each state that business-tier prompts and outputs are not used to train their models by default. Read the exceptions: Anthropic, for example, notes that conversations a user submits as feedback (the thumbs up or down buttons) may be used, and lets administrators switch that off.
• Retention. How long are prompts, outputs and uploaded files kept, and can an administrator shorten that? Some vendors offer zero-retention options on certain plans.
• Administrator controls. Single sign-on (SSO), multi-factor authentication (MFA), the ability to add and remove users centrally, and control over which features and connectors are switched on.
• Connector scope. An assistant connected to a mailbox can usually read the whole mailbox. Can you limit that? Treat every connector as a new account with access to your data.
• Audit logs. Can you see who used the tool, when, and with which connectors?
• A Business Associate Agreement (BAA) for HIPAA. If PHI is involved, you need a signed BAA before any PHI goes in. Vendors limit this to specific offerings: Anthropic lists its application programming interface (API) with a HIPAA configuration and HIPAA-ready Enterprise plans; OpenAI lists ChatGPT for Healthcare, ChatGPT Enterprise with a regulated workspace, and API accounts set up with modified retention; Microsoft says Copilot supports HIPAA compliance when properly configured, but notes that web search queries are not covered by its BAA. Check which features the BAA covers, not just which plan: Anthropic, for example, excludes connectors and beta features from its BAA, so an assistant reading a mailbox through a connector is outside that coverage. No BAA, no PHI.

These terms change often. Confirm them on the vendor's own privacy and trust pages on the day you decide, keep a dated copy, and recheck at renewal.

Path C: Scripted Automation, No AI Required

Long before AI assistants, offices saved hours with Python, PowerShell or Bash scripts, Ansible playbooks and scheduled jobs. If the task follows fixed rules, or a policy, contract, CJIS agreement or auditor rules AI out, this is the right tool.

If you already manage servers this way, our Ansible automation guide shows the same habits applied to infrastructure. For anything you build yourself, the companion guide on building and maintaining custom tools safely covers version control, secrets handling, testing and who owns the code after it ships.

The Same Job With No AI

These receipts come from known senders in known layouts, so most of the job is rule-based. The illustrative sketch below (not production code) connects to a mailbox over the Internet Message Access Protocol (IMAP), searches each sender for one year, saves every original message as evidence, pulls the amount with one pattern per vendor, writes a comma-separated values (CSV) file and prints totals. Anything it cannot read is marked for a person.

# Illustrative sketch: a year of receipts from known senders, no AI involved.
# Credentials come from environment variables, never from the script itself.
import csv, email, imaplib, os, re
from pathlib import Path

VENDORS = {  # sender -> pattern that finds the total in that vendor's template
    "receipts@appstore.example":   r"Total[:\s]+\$([\d,]+\.\d{2})",
    "billing@carrier.example":     r"Payment amount[:\s]+\$([\d,]+\.\d{2})",
    "invoices@accounting.example": r"Amount due[:\s]+\$([\d,]+\.\d{2})",
}
OUT = Path("receipts-2025"); OUT.mkdir(exist_ok=True)

box = imaplib.IMAP4_SSL(os.environ["IMAP_HOST"])
box.login(os.environ["IMAP_USER"], os.environ["IMAP_APP_PASSWORD"])
box.select("INBOX", readonly=True)  # read-only: the script never changes the mailbox

rows, totals = [], {}
for sender, pattern in VENDORS.items():
    _, ids = box.search(None, f'(FROM "{sender}" SINCE 01-Jan-2025 BEFORE 01-Jan-2026)')
    for num in ids[0].split():
        _, data = box.fetch(num, "(RFC822)")
        msg = email.message_from_bytes(data[0][1])
        body = "".join(p.get_payload(decode=True).decode(errors="ignore")
                       for p in msg.walk() if p.get_content_type() == "text/plain")
        found = re.search(pattern, body)
        amount = float(found.group(1).replace(",", "")) if found else None
        name = f"{sender.split('@')[1]}-{num.decode()}.eml"
        (OUT / name).write_bytes(data[0][1])  # keep the original as evidence
        rows.append([msg["Date"], sender, amount if amount is not None else "CHECK BY HAND", name])
        totals[sender] = totals.get(sender, 0) + (amount or 0)

with open(OUT / "receipts.csv", "w", newline="") as f:
    csv.writer(f).writerows([["date", "vendor", "amount", "file"], *rows])
for sender, total in totals.items():
    print(f"{sender}: {total:,.2f}")
box.logout()

A real version would also read HTML-only emails, convert saved messages to PDF if your bookkeeper prefers, and run once per mailbox (which handles the changed address from the example). Microsoft 365 no longer accepts basic password sign-in for IMAP, so a production script against a Microsoft 365 mailbox needs modern sign-in (OAuth 2.0), either over IMAP or through the Microsoft Graph API. Gmail accepts an app password only on accounts with 2-Step Verification, and the Gmail API with OAuth is the sturdier choice there too.

Where the script wins. It gives the same answer every time, its logic can be reviewed by an auditor, it keeps every document on your own systems, and once it works it runs on the first of each month without anyone watching. For a SOX (Sarbanes-Oxley Act) environment, that reviewable, version-controlled logic is the kind of evidence change-management testing under IT general controls (ITGC) looks for; see our guide to SOX IT general controls.

Where it loses. Every new vendor needs a new pattern, and every redesigned receipt email breaks an old one. The AI assistant handled seven vendor layouts it had never seen, and asked a question when it was unsure. A sensible hybrid: the script does the monthly run for known vendors, and anything marked "CHECK BY HAND" goes to a person, or to an approved AI tool if your policy allows it.

An AI Acceptable-Use Policy and Staff Training

Your staff are probably using AI already, approved or not. One or two written pages covering these points turns that into something you can manage:

• Approved tools and plans. Name them, including the plan. "ChatGPT" is not an approval; "our business workspace, with work accounts" is.
• What never goes in. CJI, PHI (unless the tool is covered by a BAA), cardholder data, passwords and other credentials, and anything a contract says must stay with you.
• Human review. A named person checks AI output before it is sent, filed, published or used for a decision. AI drafts; people sign.
• Connectors and new tools. Who approves connecting an AI tool to email, files or business systems, and who approves a new tool.
• Record retention. Where prompts and outputs are kept and for how long.
• Incident reporting. Who to tell, and how fast, when someone pastes the wrong thing or an AI output causes a mistake. Never punish reporting, or it stops.

For Ohio public bodies, the policy belongs inside the cybersecurity program ORC 9.64 requires; see our ORC 9.64 readiness guide. Retention matters more for you: Ohio's definition of a record (ORC 149.011) covers items in any form, created or received by the office, that document its functions, decisions and activities, so AI prompts and outputs used for official business may be records. (Personal notes kept and used only by their author are excluded, but do not assume a shared AI workspace qualifies.) Ask your records commission and legal counsel how your retention schedule applies.

For Ohio businesses relying on the ORC 1354 safe harbor, AI use needs to fit the written program you say you follow. And for everyone, a written policy gives you an honest answer when a cyber-insurance questionnaire asks how you handle sensitive data.

Short, practical training topics

Fifteen minutes each, with real examples from your own work, beats an hour of slides. Fold them into your existing security awareness training:

1. Which tools we use, which account to sign in with, and why a personal or no-cost consumer account is not the same thing.
2. The "never paste" list, with examples of what CJI, PHI and cardholder data look like in your office.
3. How to check an AI answer: confident is not the same as correct, and numbers always get verified.
4. Writing a clear request: give context, the format you want, and what to do when unsure.
5. What to do after a mistake, and who to tell.