Explore our AI courses, practical training for non-technical teamsExplore courses →Explore AI courses →
AI RisksHalloweenAI Governance

Five AI Monsters Hiding in Your Workday (and How to Banish Them)

The Confident Liar, the Shadow, the Doppelganger, the Zombie Workflow and the Rubber Stamp: five real AI risks, five controls, ten minutes a week.

TLDR: Five common AI risks at work show up as ordinary help: a confident answer, an unapproved tool, a familiar face on a call, a workflow nobody owns, a review that can’t fail. Each has a documented public case and one plain control with a named person behind it. Sources read October 29, 2026.
5monsters, each with one control you can start this week.
4documented public cases, each checked against a primary source or reputable news report on October 29, 2026.
10minutes for the Friday workday check.

Share this article

The Short Version

Give each risk a name and one control: open the source, keep a data list, call back on a channel you opened, name an owner, make reviews able to fail. Run the ten-minute Friday check. The cases are documented; the controls are Future Factors’ judgment.

Why AI risks at work look like monsters

Most AI trouble at work doesn’t arrive looking like a threat. It arrives looking like help: a confident answer, a handy tool, a familiar face on a video call, a workflow that runs itself, a review that says yes. Halloween is a good excuse to give those five a face, because a risk with a name is easier to spot on an ordinary Thursday.

So here are five monsters, each tied to a documented public case and each with one control you can start this week. The cases come from a tribunal decision, a federal court order and two news reports, all read on October 29, 2026. The controls are our own judgment about what would have helped, not findings from those sources, and I haven’t tested any tool for this piece.

My thesis is short. In every case below, the technology did something predictable, and what was missing was a plain control with a named person behind it. That’s good news, because a control is cheaper than a policy and faster than a rollout.

Five monsters, five documented cases, five controls

MonsterWhat it isDocumented caseOne control
The Confident LiarAnswers that sound sure and aren’t trueFake case citations filed in Mata v. Avianca (2023)A person opens the source before anything leaves the building
The ShadowWork data pasted into a tool nobody has ruled onSamsung code and meeting notes in ChatGPT (reported 2023)A short list of what data may go into which tools
The DoppelgangerA familiar face or voice that isn’t themArup video-call fraud (reported 2024)Confirm money and access requests on a channel you opened
The Zombie WorkflowAn AI process still running that no one ownsAir Canada’s chatbot and the tribunal ruling (2024)One named owner, a weekly sample, a review date
The Rubber StampA review that approves because it can’t do anything elseSame Mata order on lawyers’ gatekeeping roleA check that can fail, and a record of what it caught

The five monsters at a glance. Cases are described in the sections below with sources; controls are Future Factors’ judgment.

Which monster to meet first depends on your seat:

  • A finance lead or accounts payable team should start with the Doppelganger.
  • A marketing or communications manager who publishes should start with the Confident Liar.
  • A support or operations lead running bots and automations should start with the Zombie Workflow.
  • An IT, security or HR leader setting the rules should start with the Shadow and the Rubber Stamp.

Each section stays short on purpose. When you want the long version of a topic, these are the pieces to read: managing the responsible use of AI for the manager’s view, human-in-the-loop guidelines for writing review rules, and when not to use AI at work for deciding what stays off the tool altogether. I won’t repeat them here.

The Confident Liar: sure of itself and wrong

A communications manager asks an assistant for supporting figures for a draft, with sources. Back comes a tidy list with links and one-line summaries. Two of the links open, and one of those pages doesn’t contain the figure the summary promised. The tone of the answer was identical for the part that was right and the part that wasn’t.

That gap is the whole monster. A language model is built to produce fluent text, and fluent text is not the same as checked text. It’s the reason AI hallucinates in the first place, and the tone gives you no warning.

A well-documented workplace version is Mata v. Avianca. A lawyer used ChatGPT to supplement his research, and his filing cited cases that don’t exist. When he asked ChatGPT whether the cases were real, it assured him they could be found in legal databases.[1] The court’s June 22, 2023 order found that non-existent opinions, with fake quotes and citations, had been submitted, and fined the two attorneys and their firm $5,000 jointly.[2] Notice what he did right: he asked a follow-up question. What he asked it of was the same system that had produced the problem.

The judge’s order didn’t say AI has no place in legal work. It said “there is nothing inherently improper about using a reliable artificial intelligence tool for assistance”, and that existing rules put a gatekeeping role on attorneys to make sure filings are accurate. That’s the lesson for anyone who publishes, sends or signs something. The tool can draft. Someone has to be the gate.

The Open-the-Source Rule

If AI supplied a fact, quote or citation, a person opens the source before it leaves the building.

In practice that means a few specific habits, by role:

  • The marketing manager checks every statistic in a campaign draft against the page it came from.
  • The finance lead reconciles any number back to the ledger.
  • The HR partner reads the actual policy text, not the summary of it.
  • Anything you can’t trace gets cut or labelled unverified. Asking the same chatbot “are you sure?” doesn’t count as a check.

For the longer toolkit, see the anti-hallucination toolkit.

The Shadow: work data in a tool nobody has ruled on

The Shadow is the least dramatic monster and probably among the most common. Usually nobody means any harm. Someone has a long document, a stubborn bug or a meeting to summarise, and a public AI tool is one tab away.

Samsung is the case people point to. In May 2023 Fortune reported, citing Korean media, that an engineer had put confidential source code into ChatGPT to get help with a faulty database, and that another employee had shared a full meeting so the chatbot could write minutes. Samsung responded by barring employees from generative AI tools at work. Its memo cited concern that data sent to these platforms sits on external servers and is hard to retrieve or delete, and it described the ban as temporary while the company worked on its own tools.[3] The report doesn’t say whether those staff were breaking an existing rule, and I don’t want to guess. What it shows is what the Shadow looks like: ordinary work text going somewhere the company hadn’t decided about.

A blanket ban is one answer, and Samsung’s was explicitly temporary. A narrower answer works better for most teams, because people go on using the tools when the rules are too blunt. The control is a short, written list of what kind of data may go into which kind of tool, and it needs to be short enough that people can remember it without opening a document.

A data traffic light you can adapt

Data typePublic AI toolApproved company toolWho decides edge cases
Public or already published (published copy, public product facts)FineFineAnyone
Internal, low risk (draft agendas, generic templates)Remove names and figures firstFineTeam lead
Customer, employee, financial or unreleased (contracts, source code, payroll, forecasts)StopOnly if the tool has been approved for that data classIT or security lead, in writing

A starting point, not a standard. Which tools count as approved, and what they may hold, depends on your contracts and your counsel.

It also helps to say why the list exists. People follow a rule more readily when they know the reason, and “data sent to an outside tool can be hard to get back” is a reason a developer or an analyst will recognise from their own work.

The edge-case column is the part teams skip. A marketing manager with a customer list, a finance analyst with a forecast and a developer with a stack trace all hit grey areas quickly, and if nobody is named to answer, they guess. For more on spotting and handling it, see Shadow AI and using AI without leaking company data.

The Doppelganger: a face or voice that isn't who it seems

For years, a message that looked and sounded like your boss was a reasonable thing to trust. That assumption is what the Doppelganger attacks, and it targets whoever can move money or grant access: an accounts payable lead, a finance analyst, an IT administrator, an executive assistant.

The case here is Arup. A finance worker at the engineering firm’s Hong Kong office joined a video call with people he believed were the chief financial officer and other colleagues. Police said they were deepfake re-creations, and the worker made 15 transfers totalling HK$200 million, about US$25.6 million at the time of CNN’s report in May 2024. He had first suspected a phishing email before he joined the call. Arup said fake voices and images were used and that its internal systems weren’t compromised.[4] By Arup’s account, no internal system was compromised. A person was persuaded, in a meeting that looked real.

That’s why the control can’t be “look more carefully”, because the whole point of the fake is that it passes a careful look. The control sits outside the face and voice.

The Callback Rule

Any request to move money or grant access gets confirmed on a channel you opened yourself, whoever appears to be asking.

Opened yourself means a number from the company directory or a message in a system you start, not the link, number or call the request came from. Pair it with a second approver above a limit your finance lead sets. Neither step needs a tool, and both can feel rude the first time. That’s worth saying out loud in advance, so a junior team member who calls the CFO back is thanked for it. Our piece on deepfakes at work covers the wider picture.

The Zombie Workflow: still running, owned by no one

An operations lead notices that a weekly summary email is still going out to a distribution list. The automation behind it was built months ago with an AI step in the middle. The person who set it up has moved teams, the source it summarises has changed format, and nobody can say who would fix it if it broke.

That’s a Zombie Workflow: it still runs, so it looks alive, and nobody is responsible for what it says. The risk grows with how public its output is, which is why customer-facing bots are the sharpest example.

In Moffatt v. Air Canada, a customer asked the airline’s website chatbot about bereavement fares after a death in his family. The chatbot said he could apply for a reduced fare within 90 days after travel, while a separate page on Air Canada’s own site said the policy didn’t apply to requests made after travel. Air Canada argued it wasn’t liable for what its chatbot said, which the tribunal said suggested the chatbot was a separate legal entity. In its February 14, 2024 decision the tribunal called that “a remarkable submission”, found negligent misrepresentation, and ordered Air Canada to pay C$812.02 in total.[5] The sum was small. The principle wasn’t: the company answered for what its website said, whichever part of the website said it.

Here is my reading, and it’s an inference, not a finding. A bot and a policy page that disagreed for long enough to cost someone money is the kind of mismatch a routine comparison would surface. The tribunal found Air Canada hadn’t taken reasonable care that the chatbot was accurate. A weekly check would have been one way to show that care.

The Named-Owner Rule

Every AI workflow has one named owner, a weekly sample and a review date, or it gets switched off.

In practice each part is small:

  • Owner: a person with a role, such as the customer service lead for the support bot or the demand generation manager for the campaign automation.
  • Weekly sample: read a handful of real outputs against the current source of truth.
  • Review date: the day you decide to keep, change or retire it. Retiring is the part that often gets skipped, and it’s what stops a Zombie.

The Rubber Stamp: a review that can't fail

The Rubber Stamp is the monster that can undermine the other four. A review step exists, someone approves, and the approval means nothing because the reviewer can’t see the source, doesn’t have time, or hasn’t sent anything back in a long time.

The same Mata order shows the pattern. One attorney did the research and wrote the affirmation, and another, who had signed it, was the lawyer of record. The court sanctioned both. It records that the defendant’s lawyers said in March they couldn’t locate most of the cases, that the court ordered copies in April, and that the two attorneys “doubled down and did not begin to dribble out the truth until May 25”, after an order to show cause.[2] A review that can’t change the outcome is a signature, not a check.

The fix is to make the review able to fail. Give the reviewer the source, not just the output. Give them three things to check, not “looks good?”. And ask them to note what they changed or rejected, because a reviewer who has changed nothing in a month is either working with a very good process or not really checking, and you’d like to know which. One test is to slip a known error into a batch and see if it’s caught. If it isn’t, the review needs repair before the AI does.

Reviewers need time as much as access. A support lead who is asked to approve a long queue of AI drafts between other tasks will approve them. If the volume is too high to check properly, the options are to check a sample, slow the process down or take the task off AI, and “approve everything quickly” isn’t among them.

It also helps to be explicit about who owns what, so nobody assumes the tool covers the human’s part.

What AI does, what the human owns, how it gets checked

What AI doesWhat the human still ownsHow it gets checked
Drafts a customer reply from the case notesWhether the reply promises anything the company can’t deliverSupport lead compares the draft to the policy page, three replies a week
Summarises a supplier contractWhat the clauses actually commit the company toContract owner reads the termination and renewal clauses in the original
Pulls figures for a campaign reportWhether the numbers are right and what they mean for budgetMarketing manager reconciles each figure to the source dashboard
Writes a first-pass job descriptionThe legal and fairness review of what it requiresHR partner checks it against the role profile and legal guidance

Examples of the split for four roles. Replace the checks with ones your team will actually run. For a deeper treatment of review design, see the human-in-the-loop guidelines linked above.

Your ten-minute check, and one line for each monster

You can run this check on your own, and plenty of teams do. A facilitated workshop is useful when you want a whole team to learn the same habits against its own real work, and our corporate workshops are built for that. Either way, the Friday check below takes about ten minutes.

The ten-minute workday check

MinutesMonsterLook at thisPasses if
2Confident LiarAnything that went out this week with a fact, quote or citation from AIYou can point to the source you opened for each one
2ShadowWhich tools your team pasted work into this weekEach one is on your list for that kind of data
2DoppelgangerAny payment or access request that came by call, video or messageEach was confirmed on a channel you opened
2Zombie WorkflowOne AI workflow or bot, picked at randomIt has a named owner and a sample read this week
2Rubber StampThe last review of an AI outputThe reviewer had the source and changed or rejected something

Score one point for each pass. Four or five: keep the habit. Two or three: fix the failures this week, starting with whichever touches money or customers. Zero or one: pause new AI use in that area until a control exists. The bands are our judgment, not a validated scale.

When one of the monsters does turn up, the first minutes matter more than the investigation. Keep these one-liners somewhere people can find them.

One line for each monster, for the moment it happens

MonsterFirst response
Confident LiarPull the item back, tell everyone who received it, and replace the claim with one you have opened the source for.
ShadowStop pasting, write down what data went into which tool, and tell your security or IT lead the same day.
DoppelgangerHalt the transfer or access, call your bank and security lead on known numbers, and preserve the call and message records.
Zombie WorkflowPause it, then either name an owner and a review date or retire it.
Rubber StampRe-check the last batch against the sources, then add a test the review can actually fail.

Written for the first hour. Legal and regulatory duties differ by country and sector, so check yours with counsel.

This guide has limits. Four cases are a small and unrepresentative sample, because they’re the ones that became public, and they tell you nothing about how common each monster is. The controls are judgment, and I haven’t tested them in your setting. Pick the one monster closest to money or customers, run its row of the check next Friday, and name the person who’ll run it after that.

Hina Mian
Hina Mian, Co-Founder of Future Factors AI

Hina is a marketing strategist with over a decade of hands-on campaign experience across B2B and consumer brands. She writes about using AI to run leaner, sharper marketing without losing the human touch. Future Factors helps professionals and teams build practical AI capability through role-based training, workflow design, and hands-on adoption.

More about Hina →

Frequently Asked Questions

Should we ban public AI tools to stay safe?

A blanket ban is one option, and Samsung’s was described as temporary while it built its own tools. For most teams a written list of which data may go into which tools, with a named person for edge cases, is easier to follow and harder to quietly bypass.

Is a deepfake call a realistic risk for a small company?

The Arup case involved a large firm, but the method needs only a convincing request and a person who can move money. Smaller teams often have fewer approvers, which makes a callback rule and a second approver more useful, not less.

If a chatbot gives a customer wrong information, is the company responsible?

In Moffatt v. Air Canada, a British Columbia tribunal held the airline responsible for its own chatbot’s statement. It is one small-claims decision in one province, not a rule everywhere, and this is not legal advice, so ask counsel about your market.

How can we tell whether our review step is real?

Slip a known error into a sample batch and see whether the reviewer catches it, and look at whether reviewers ever send anything back. A review with no record of changes may be fine or may be a signature, and the test tells you which.

We can only fix one thing this month. Which?

Start with the monster closest to money or customers. For a finance team that is usually the callback rule. For a team that publishes, it is opening the source. For a team running bots, it is naming an owner.

About This Article

I read these sources on October 29, 2026: the British Columbia Civil Resolution Tribunal decision in Moffatt v. Air Canada (2024 BCCRT 149), a Westlaw copy of Judge Castel’s June 22, 2023 sanctions order in Mata v. Avianca, a law firm’s summary of the lawyer’s affidavit, CNN’s May 2024 report on the Arup fraud, and Fortune’s May 2023 report on Samsung. The Mata order was only partly readable in the copy I could fetch, and the Samsung account relies on Korean media as relayed by Fortune. I left out scare statistics about how often these incidents happen because I couldn’t confirm them to a primary source. This is not legal advice. I did not test any tool hands-on for this article. The five monsters, the controls, the traffic light, the Friday check and the one-line responses are Future Factors working methods, not research-validated models, and the scenes are illustrations, not client stories.

Sources

  1. Hinshaw & Culbertson LLP, New York Lawyer Who Cited Non-Existent Cases Generated by ChatGPT Now Faces Sanctions Hearing (law firm summary of the lawyer’s affidavit, 2023). Counsel said he consulted ChatGPT, asked whether the cases were real, and was assured they could be found in legal databases. Secondary summary of a court filing, read October 29, 2026. https://www.hinshawlaw.com/en/insights/lawyers-for-the-profession-alert/new-york-lawyer-who-cited-non-existent-cases-generated-by-chatgpt-now-faces-sanctions-hearing
  2. U.S. District Court, S.D.N.Y. (Judge P. Kevin Castel), Mata v. Avianca, Inc., 678 F. Supp. 3d 443, Opinion and Order on Sanctions, June 22, 2023 (hearing held June 8, 2023). Non-existent judicial opinions with fake quotes and citations generated by ChatGPT were submitted; $5,000 sanction imposed jointly and severally on two attorneys and their firm. Court filing (docket 54, CourtListener copy), read in full for the cited points on October 29, 2026. https://storage.courtlistener.com/recap/gov.uscourts.nysd.575368/gov.uscourts.nysd.575368.54.0.pdf
  3. Fortune, Samsung bans ChatGPT among employees after sensitive code leak (May 2, 2023). Reports, citing Korean media, that an engineer put confidential source code into ChatGPT and another shared a full meeting to generate minutes; Samsung barred employees from generative AI tools at work and described the ban as temporary. Reputable news report, read October 29, 2026. https://fortune.com/2023/05/02/samsung-bans-employee-use-chatgpt-data-leak
  4. CNN, Arup revealed as victim of $25 million deepfake scam involving Hong Kong employee (May 16, 2024). A finance worker at Arup’s Hong Kong office joined a video call with people he believed were the chief financial officer and colleagues; police said they were deepfake re-creations, with 15 transactions totalling HK$200 million (about US$25.6 million). Arup said fake voices and images were used and that its systems were not compromised. Reputable news report, read October 29, 2026. https://edition.cnn.com/2024/05/16/tech/arup-deepfake-scam-loss-hong-kong-intl-hnk
  5. Civil Resolution Tribunal of British Columbia, Moffatt v. Air Canada, 2024 BCCRT 149 (File SC-2023-005609), decision of February 14, 2024, Tribunal Member Christopher C. Rivers. Finds negligent misrepresentation by Air Canada over its website chatbot’s statement about bereavement fares and rejects the argument that the chatbot was a separate entity. Orders C$812.02 in total (damages C$650.88, interest C$36.14, tribunal fees C$125). Primary source, read October 29, 2026. https://decisions.civilresolutionbc.ca/crt/crtd/en/525448/1/document.do

Psst, Hey You!

(Yeah, You!)

Want helpful AI tips flying Into your inbox?

Weekly tips. Real examples. Practical help for busy professionals.

We care about your data, check out our privacy policy.