AI Risks for Business: What Matters and How to Manage Them

AI risk is the chance that using AI harms your business: wrong answers, leaked data, security attacks, biased decisions, legal liability. Manage it by listing your AI tools, rating each use by risk, and having a person check the output.
This guide covers the risks that matter, the numbers behind them, and a simple way to rate each AI use. For the rules and owners that manage them, see AI governance for a mid-size company. For senior help, see the consulting page or my fractional Chief AI Officer service.
What is AI risk?
AI risk is the combination of how likely an AI-related harm is and how serious it would be. The harm can hit your customers, your staff, your data or your reputation.
The phrase covers two very different conversations. One is about the risks a company takes on when it uses AI tools at work, such as a chatbot giving a customer the wrong refund policy. The other is about large-scale risks from advanced AI systems, such as AI-assisted cyberattacks or weapons development. This post is about the first one, with a short section on the second so you can place it.
Which AI risks matter most for a business?
For most companies, eight risks cover almost everything. The table lists each one, what it looks like in daily work, and a first control.
| Risk | What it looks like | First control |
|---|---|---|
| 1. Wrong or invented output | A chatbot or writing tool states a false policy, number, quote or source with confidence | A person checks facts, numbers and sources before anything reaches a customer or a decision |
| 2. Data leaks and privacy | Staff paste customer records, contracts, source code or financial data into a public AI tool | Approved tool list and a rule on which data may go in which tool |
| 3. Shadow AI | Staff use AI tools the company has not reviewed, or AI features switched on inside software you already pay for | A shared inventory of tools and features, plus a simple way to request a new one |
| 4. Security attacks | Prompt injection, stolen access to AI accounts, AI-written phishing and deepfake impersonation | Least-privilege access, multi-factor sign-in, and staff training on AI-made scams |
| 5. Biased or unfair decisions | AI helps screen job candidates, price a service or approve a customer, and treats groups differently | No final decision about a person without independent human judgment, and a record of how it was made |
| 6. Legal and contract liability | A company is held responsible for what its AI system tells a customer, or breaks a vendor or client term | Legal review of customer-facing AI and a read of vendor terms before approval |
| 7. Over-reliance | Staff stop checking, skills fade, and one tool becomes a single point of failure | Spot checks, a fallback process, and a named owner for every important AI workflow |
| 8. Vendor and supply chain | A vendor changes its terms, trains on your data, has its own breach, or shuts down | A short vendor questionnaire and an exit plan for tools that run key processes |
Risk 1 is often called an AI hallucination. Risks 2 and 3 are linked, because shadow AI means staff are putting company data into tools nobody has reviewed.
How common are AI incidents, and what do they cost?
AI incidents are rising, and the best numbers come from breach and incident trackers. Four sourced figures give the picture:
| What the research found | Number | Source |
|---|---|---|
| Reported AI-related incidents in 2024, up from 2023 | 233 (a 56.4% increase) | Stanford AI Index, 2025 |
| Organizations that reported a breach of an AI model or application, and how many of those lacked proper AI access controls | 13% reported a breach; 97% of those lacked proper access controls | IBM, July 2025 |
| Breached organizations with no AI governance policy, or one still in development | 63% | IBM, July 2025 |
| Extra breach cost at organizations with high levels of shadow AI | $670,000 on average | IBM, July 2025 |
Two details in the IBM study are worth knowing. One in five organizations reported a breach caused by shadow AI, and only 37% had policies to manage AI or detect shadow AI. Separately, 16% of the breaches IBM studied involved attackers using AI tools, most often for phishing or deepfake impersonation. The study covered 600 organizations and breaches between March 2024 and February 2025, and the cost figure is an average across that sample.
Stanford’s count comes from the AI Incident Database, which relies on public reports. The real number is probably higher, and part of the rise reflects more AI use and more reporting.
What does a real AI risk case look like?
A good example is the Air Canada chatbot case. In 2024 a British Columbia tribunal ruled in Moffatt v. Air Canada that the airline was liable for negligent misrepresentation after its website chatbot gave a customer wrong information about bereavement fares. Air Canada had argued the chatbot was responsible for its own actions. The tribunal rejected that, saying the chatbot was part of the airline’s website and the airline was responsible for all the information on it.
The lesson is practical. A customer-facing AI system speaks for the company. If it is wrong, the company can own the result, and a disclaimer on a different page may not undo it. McCarthy Tétrault, a Canadian law firm, describes the decision as a reaffirmation that organizations are generally responsible for the acts of the computer systems they use. The case came from a Canadian tribunal, so check how your own state or country treats a similar claim.
What do experts say about the biggest AI risks?
Experts disagree on how likely the worst outcomes are, but they agree on a short list of risks worth watching. In a study from MIT FutureTech and the University of Queensland, reported by MIT Sloan on July 20, 2026, 272 experts rated 24 AI risks on likelihood and severity of harm for 2025 to 2030.
- The five risks with the highest expected severity were AI with dangerous capabilities, competitive pressure, weapons and cyberattacks, concentrated power, and false or misleading information.
- Information, national security and finance were seen as the most vulnerable sectors.
- Under a business-as-usual scenario, experts judged that 18 of the 24 risk areas had at least a 10% probability of catastrophic outcomes over five years, where catastrophic meant more than 1 million deaths, more than $100 billion in losses, or comparable damage. Cost-effective mitigation lowered those estimates but did not remove them.
- The people most exposed to AI risks are often not the ones best placed to fix them.
For a business leader, the study’s co-author Peter Slattery said the right response is continuous, not a one-time fix, because AI is moving fast. The practical question the study raises is whether competitive pressure is pushing your company to deploy AI faster than your risk practices can keep up.
How do you assess AI risk in your own company?
Rate each AI use, not each tool, because the same tool can be safe for one job and risky for another. Use four steps:
- List every use. Ask each team lead which AI tools and AI features they use, and for what. Check the software you already own, since AI is built into email, design, support and accounting tools.
- Note the data. For each use, write down the most sensitive data that goes in: public, internal, or customer, employee and financial.
- Note the decision. Does the output reach a customer, or influence a decision about a person’s job, money or access to a service?
- Sort into tiers. Match the amount of review to the tier, as in the table below.
| Tier | Typical use | Data | Review |
|---|---|---|---|
| Low | Drafting a public blog post, brainstorming, summarizing public pages | Public information | Staff check the output; no formal approval |
| Medium | Summarizing internal documents, drafting customer emails, analyzing company data | Internal data | Approved tool only; a person reviews before sending; vendor reviewed |
| High | Customer-facing chatbots, screening candidates, credit or pricing decisions, tools that read contracts or customer records | Customer, employee or financial data, or a decision about a person | Named owner, legal and security review, human sign-off, logging, and a way to switch it off |
Review the list every quarter, and again after any incident or major change to a tool’s terms. If you want a structured starting point for the inventory, the AI readiness assessment checklist covers where a company stands before it adds more AI.
How do you reduce the main AI risks?
Most of the controls are ordinary management practices applied to a new tool. The ones with the best return are cheap.
- Keep a person in the loop for anything that matters. Staff should check names, numbers, quotes and sources, because AI tools state wrong answers with confidence. Do not let AI make a final decision about a person without independent human judgment.
- Give staff an approved list and a way to ask for more. A flat ban sends AI use out of sight, which makes shadow AI worse.
- Set data rules with examples. Say which kinds of data may never go into a public tool, and which may go into approved ones.
- Limit what AI systems can access. An AI agent or assistant should only reach the files and systems its job needs. OWASP ranks prompt injection first in its 2025 Top 10 for LLM applications. It happens when input, such as text hidden in a web page or file, changes a model’s behavior in ways the developer did not intend, and OWASP says it is unclear whether fool-proof prevention exists. Limiting access limits the damage.
- Train people. Staff cannot follow rules they have never seen explained. Short sessions with real examples work better than a long document. See AI training for employees for what a good session covers, or book a live, hands-on session through my AI workshops.
- Review vendors before approving them. Ask where data is stored, whether it trains the model, who can access it and how it is deleted.
- Plan for an incident. Name a contact, know how to switch a tool off, and keep a log of errors and complaints.
- Test before you roll out. Try a new tool on a small team and a small set of real cases before you open it to everyone. See AI adoption for a rollout approach.
Here is how I keep a person in the loop. I let Claude read my email on its own, but anything it writes needs my approval first. Claude will not send emails for you unless you tell it to. If you say reply, it replies and sends, so I always say draft, or tell it not to send.
I limit access the same way with scheduled work. In my agents video, I explain that I run more than 15 routines every day and none of them is allowed to message me. They only write an output file, and one chief of staff agent reads those files and sends me a single summary of what needs my attention.
Which framework should you use to manage AI risk?
The NIST AI Risk Management Framework is the common starting point. It is voluntary and meant for organizations of any size. NIST released version 1.0 on January 26, 2023, and a Generative AI Profile on July 26, 2024. NIST says it is revising version 1.0 as part of the White House AI Action Plan, so confirm the current version before you quote it to a customer or auditor.
The framework groups the work into four functions:
| NIST function | In plain English | How it shows up in the steps above |
|---|---|---|
| Govern | Decide who owns AI risk and what the rules are | Named owner, approved list, data rules |
| Map | Understand where AI is used and what could go wrong | The use list and the tiers |
| Measure | Check how well the AI and the controls work | Output checks, logs, vendor review |
| Manage | Act on what you find | Fixes, incident plan, retiring tools that fail |
This reading is plain English, not NIST’s wording, so check the framework itself for the exact definitions.
What is the difference between AI risk and AI governance?
AI risk is the question of what could go wrong and how badly. AI governance is the system a company builds to answer it: owners, policies, approval steps, training and review dates. You need to understand the risks to design the governance, and you need the governance to act on the risks.
If you have finished the risk list in this post, the next step is the program. The AI governance guide covers the six parts of a working program, the employee usage policy checklist and a four-week start plan.
When should you get outside help with AI risk?
Bring in a senior person when the work matters every week but does not fill a full-time role, or when a high-tier use is about to launch and nobody on staff has run this before. Some companies fill the gap with a fractional Chief AI Officer. For the role, see what a Chief AI Officer does, and for timing, see when to hire a fractional Chief AI Officer.
What should you do next?
Do three things this week. Ask every team lead which AI tools they use and for what. Mark each use low, medium or high with the tier table above. For every high-tier use, name an owner and confirm a person reviews the output before it reaches a customer or a decision. If you want help with it, my fractional Chief AI Officer service owns the plan, the policy, the training and the shared systems.
Sources
- MIT Sloan, These are the most urgent AI risks, according to 272 experts, published July 20, 2026.
- IBM, 13% of organizations reported breaches of AI models or applications, 97% of which reported lacking proper AI access controls, published July 30, 2025.
- Stanford HAI, AI Index Report 2025, Responsible AI chapter, 2025.
- McCarthy Tétrault, Moffatt v. Air Canada: A Misrepresentation by an AI Chatbot, published February 19, 2024.
- OWASP, LLM01:2025 Prompt Injection, read November 12, 2026.
- NIST, AI Risk Management Framework, read November 12, 2026.


