Insider misuse of AI: when approved tools help dishonest work

The risky moment can look like a clean handoff.

A sales manager asks the company's approved AI assistant to summarize active accounts, renewal dates, pricing history, decision-makers, objections, and open follow-ups. The request sounds normal. Sales leaders prepare handoff notes all the time.

The same summary can also become a portable client playbook.

Insider risk existed before AI. What changes is the amount of effort required. AI can turn scattered CRM entries, emails, proposal notes, pricing sheets, and customer history into a clean package in minutes. The tool may be approved, the user may have legitimate access, and the work may look like normal productivity until the business understands the intent.

What the risk is

This risk is staff using approved AI tools to take company information out of the business, fabricate records, impersonate people, bypass business controls, inflate activity, or create a plausible cover story for misconduct.

The approved AI tool already supports legitimate work:

• Summarizing accounts.
• Drafting client emails.
• Cleaning up CRM notes.
• Preparing handoff documents.
• Reviewing expense details.
• Rewriting messages into a more professional tone.
• Turning scattered information into a clear plan.

That is what makes this risk hard to spot. The audit log may show an approved user, using an approved AI tool, against data the user could normally access. The difference is intent, and intent is usually invisible in the tool log.

Common patterns include:

• Exit theft: a departing employee uses AI to summarize client lists, pricing, renewal dates, and decision-maker relationships into portable form.
• Fabrication: AI helps create fake receipts, references, resumes, delivery notes, customer approvals, or other business records.
• Impersonation: AI drafts messages that sound like a manager, vendor, customer, or colleague.
• Control evasion: AI rewrites questionable activity into softer, more legitimate-sounding language.
• KPI inflation: AI generates plausible customer notes, follow-ups, or support interactions to make activity metrics look better than they are.
• Cover for misconduct: AI drafts the CRM update, customer note, follow-up email, or internal explanation that makes the activity look routine.

This is separate from Phishing and payment fraud and Voice and video impersonation, where an external attacker uses AI to deceive the business. Here, the person is inside the business and has legitimate access.

It is also separate from confidential data entering AI and When AI sounds right but is wrong. Those articles cover accidental data exposure into AI and wrong AI output. This article covers intentional misuse of an approved tool.

How it happens in a normal SMB

A sales manager at a 60-person Alberta industrial services company has accepted a job with a competitor. She has not resigned yet.

Her access still looks normal. She can open the CRM, shared proposal folders, pricing sheets, customer email history, and renewal trackers. She needs that access for her role, and nothing about it looks unusual by itself.

The company has approved an AI assistant for productivity work. Staff use it to summarize account history, draft customer emails, clean up CRM entries, and prepare internal handoff notes.

The manager starts with a reasonable request:

Prepare transition notes for my active accounts, including open issues, renewal timing, pricing history, key contacts, decision-maker preferences, and likely next steps.

The AI assistant turns scattered CRM fields, old email threads, proposal notes, renewal dates, and service complaints into clear account briefs. The output is useful. It is also much cleaner than the raw data.

The manager keeps going. She asks the AI to group accounts by renewal window, pricing sensitivity, relationship strength, and competitor mentions. She asks it to identify customers that may be open to a new provider. She asks it to turn the summaries into a territory plan.

The AI has no way to know the purpose changed. It is doing approved summarization work against data the manager can access.

The manager also uses AI to make the activity look ordinary. It drafts CRM updates that sound like pipeline cleanup. It writes short internal notes explaining why files were exported. It prepares polite handoff messages that make the account summaries look like responsible transition work.

A week later, she resigns.

The outreach lands close to renewal dates, references service frustrations that were never public, and shows unusual awareness of pricing pressure. The owner asks how the competitor knew so much.

The investigation is frustrating. There is no malware, no strange login from overseas, and no obvious compromised account. The logs show the former manager using approved systems and the approved AI assistant before resignation. They show account summaries, CRM updates, and file access from a user whose job required customer access.

The business can see activity. It struggles to prove purpose, scope, and whether the AI summaries were used outside the company.

The failure path

The failure path looks like this:

Case file Sequence 13 · Insider misuse

1. 01

   A staff member has legitimate access to company information.

2. 02

   The business has approved AI for normal productivity work.

3. 03

   The staff member's intent changes because of resignation, conflict, pressure, fraud, or personal gain.

4. 04

   The staff member uses AI to summarize, repackage, rewrite, fabricate, impersonate, or explain activity.

5. 05

   The output looks like ordinary work product: account briefs, CRM notes, expense explanations, customer messages, or handoff material.

6. 06

   The audit trail shows approved users, approved tools, and data the person could normally access.

7. 07

   The business discovers the issue through client loss, record mismatch, suspicious timing, a complaint, or a later investigation.

8. 08

   The investigation has to separate legitimate productivity work from preparation for theft, fabrication, impersonation, or cover-up.

Tool approval answers one question: whether the software belongs in the business. Intent, authorization, and role fit still have to be judged from access, behaviour, timing, and business context.

Business consequence

The first consequence is loss of business control over client intelligence.

In the sales-manager example, the customer list is only one part of the value. The useful material is the interpretation: renewal timing, relationship history, price sensitivity, decision-maker preferences, open complaints, service weaknesses, and likely next pitch. AI makes that interpretation easier to create, cleaner to carry, and harder to distinguish from legitimate handoff work.

Other consequences depend on the misuse:

• Customer loss when a departing employee uses AI to prepare a competitor-ready account playbook.
• Pricing damage when renewal dates, discount history, margin pressure, or negotiation weaknesses are summarized for outside use.
• Record integrity problems when AI-generated CRM updates, delivery notes, approvals, receipts, references, or activity notes enter business systems.
• Internal confusion when managers cannot tell whether records were created from real activity or generated after the fact.
• Fraud exposure when AI helps fabricate expense details, supplier explanations, applicant material, or customer approvals.
• Impersonation harm when staff use AI to draft messages that appear to come from a manager, customer, vendor, or colleague.
• Investigation cost when the business has to reconstruct intent from logs that show ordinary access and approved AI usage.

The evidence problem is serious. The business may know that an employee summarized accounts, exported documents, or drafted messages while still lacking the details that matter: which AI outputs were copied elsewhere, which drafts became external messages, which records were fabricated, or whether the employee used personal devices or personal accounts after creating the summaries.

That ambiguity matters in disputes with former staff, new employers, customers, insurers, and lawyers. The company may believe misconduct occurred and still struggle to show exactly what happened.

Controls that interrupt the failure path

The first control is to treat AI approval as a software decision, then govern use through access, behaviour, and records.

An approved AI tool still needs access rules, behaviour monitoring, record controls, and exit procedures. The tool can be legitimate while a specific use violates company rules.

The control should focus on behaviour and access. A tool log may show that a sales summary was created; the surrounding pattern shows whether it was normal handoff work or part of a competitor move. Timing, volume, access scope, export path, role change, and later customer activity all matter.

Because most staff use AI honestly, added review should follow the company's written procedure and attach to role risk, access level, notice periods, unusual volume, or specific investigation triggers.

Policy rule this creates