AI for SME Success logo
RESOURCE HUB
NEWSLETTER

One AI Use Case Delivers

Another Backfires

Real up-to-date examples of AI successes and failures. Resources & guides.

One AI Use Case Delivers

Another Backfires

Real up-to-date examples of AI successes and failures. Resources & guides.

Why Knowing Where AI Fails and Excels Matters

AI is multipurpose. It excels at some tasks while failing badly at others. Data from SQ Magazine’s LLM hallucination statistics 2026 shows how widely hallucination rates vary by task:

  • Document summarization grounded in specific sources: 2%
  • Legal research: 58 to 88%

Even the companies building AI uncover its capabilities through experimentation and testing! For a small business owner, knowing where AI is safe and where the risk is not worth it is critical.

Where AI Works Exceptionally Well

Austin Lau, a non-technical growth marketer at Anthropic, used AI to manage Anthropic’s entire multi-channel paid growth stack single-handedly. By splitting workflows into specialized sub-agents, using Figma plugins to auto-populate ad templates, and connecting Claude directly to the Meta Ads API via MCP, he cut production from hours to minutes while scaling output.

Why it works: AI makes high-volume ad experimentation cheap and fast. It analyzes performance data well and generates hundreds of constrained, on-brand copy variations. Small businesses can produce copy, image, and video variations in hours instead of weeks, spot winners faster, and scale what works.

Rebel Cheese, an Austin-based vegan cheesemaker profiled by Fast Company, used Claude to analyze a year of shipping invoices and recovered $250,000 in overcharges. They then used Manus AI to build an automated billing audit tool that flags discrepancies and generates refund requests in real time. Fortune later reported the system has saved them $400,000 against tool costs of about $3,000 per year for AI subscriptions.

Why it works: AI excels at scanning huge volumes of repetitive billing data for anomalies. The carrier, an outside party, verifies every claim before paying out. Risk is minimal, and the upside over doing it manually is huge.

On an episode of The 1 on 1 with CNN, Pfizer CEO Albert Bourla said he uses multiple AI tools to critique each other on major decisions. He asks Claude for an opinion, then uploads the response into Gemini and asks it to challenge the reasoning. This surfaces blind spots, counterarguments, and alternative perspectives before he makes the final call.

Why it works: AI is outstanding at generating multiple viewpoints, stress-testing assumptions, and debating opposing arguments at scale.

1. Fortune profiled Rick Chorney, a 29-year-old high school dropout in Abbotsford, BC, who runs Echo Janitorial Services. AI has cut his workload by more than half. He uses Synthesia to produce videos that walk new employees through procedures.

2. The Wall Street Journal profiled another small business owner, Michael Salvatore, who feeds data from both his point-of-sale system and his QuickBooks bookkeeping into Google’s AI tool NotebookLM. The tool turns that data into a podcast on how the business is faring and where it could improve, which he shares with his managers.

Why it works: When AI tools are grounded ion your own materials, they produce the most reliable results. Moreover, you can easily verify the accuracy of these results.

Where AI Falls Short

1. Reuters reports Starbucks scrapped its AI inventory tool nine months after rolling it out to more than 11,000 stores across North America. The app kept miscounting and mislabeling items on shelves, failing to recognize syrup bottles and confusing similar types of milk.

2. In Stockholm, an AI agent named Mona runs Andon Café. Mona bought 120 eggs for a café with no stove and ordered 3,000 disposable gloves for a café serving about one customer an hour, on top of 6,000 napkins, four first-aid kits, industrial trash bags, and 22.5 kilograms of canned tomatoes. Staff built a customer-facing ‘Hall of Shame’ shelf to display the odder purchases.

Why it fails: AI lacks physical-world common sense. It doesn’t know you need a stove to cook eggs, or that a painter in another country can’t show up to do the work. Worse, when pressured, AI agents lie, hide, and overreach instead of flagging that they are stuck.

New NBER research found chatbot-built portfolios give you no real edge over passive index investing. AI picks skew heavily toward large-cap tech because LLMs lean on media coverage when recommending stocks, producing less diversified results.

Why it fails: LLMs reflect patterns in their training data, not market fundamentals. Stocks that get written about more get recommended more, regardless of valuation, risk, or fit with your situation.

A Quebec realtor and his agency had to publicly apologize after using AI-generated images to market a home.

Why it fails: AI image tools generate plausible visuals that are not real. In a regulated industry where listings are legally binding representations of a property, this is not an aesthetic problem. It is misrepresentation, with real legal exposure.

CNN reports that an Australian travel company’s AI-generated blog content sent tourists to hot springs that do not exist. The founder said the online hate and damage to the business’s reputation was soul-destroying.

Why it fails: AI hallucinates locations, addresses, hours, and prices when generating content. Such errors can seriously damage small business reputation.

A CFOtech article reports Canadian businesses incurred financial losses and compliance issues after relying on AI tools for tax, bookkeeping, and financial guidance.

Why it fails: Tax and accounting rules are jurisdiction-specific, change every year, and reward precision. AI models are not reliably updated to these.

A New York lawyer used ChatGPT to draft a court filing. The AI fabricated case citations that looked real but did not exist. The lawyer was sanctioned and the case made international headlines.

Why it fails: The hallucination rate for legal research is 58 to 88%. AI invents case law confidently, with plausible names and citation numbers.

Sinch report covered by ITPro polled more than 2,500 customer service leaders and found three in four companies have pulled back or abandoned AI agents in customer service. The top concerns were customer data exposure (31%), hallucinations and brand-reputation risk (22%), and a lack of auditability and oversight (16%).

Here is a specific example. The Swedish company Klarna replaced its entire customer support team with an AI chatbot in 2024. Two years later it is rehiring human agents. As Forbes reported, Klarna’s CEO admitted that cost-cutting had pushed service quality too low.

Why it fails: AI is fast and cheap on routine questions, but it cannot read emotional tone, and at scale can put your customer data, accuracy, and oversight at risk.

Where AI Can Be Dangerous

1. Veracode’s Spring 2026 review of more than 150 AI models found that secure code was generated in just 55% of coding tasks.

2. Georgia Tech’s Vibe Security Radar reported a steady rise in vulnerabilities linked to AI-assisted software development.

3. Escape analyzed 5,600 live applications and identified over 2,000 critical security issues, hundreds of leaked authentication secrets, and numerous cases of sensitive customer information being publicly accessible.
Common findings included weak authorization mechanisms, inadequate protection of sensitive keys and tokens, and cross-site scripting flaws.

Why it fails: AI coding tools generate code that runs, passes tests, and looks polished, then leaks customer data through basic access-control flaws.

Give an AI agent the keys to live systems and it will eventually use them. Four recent cases:

1. A Cursor agent powered by Claude encountered a credential issue at PocketOS, an automotive SaaS provider. Rather than asking for guidance, it located a high-privilege API token and erased the company’s production database along with its backups in under ten seconds. The agent later documented the safeguards it had ignored during the incident. (Source: Inc.)

2. A developer instructed Google’s Antigravity agent in Turbo mode to clear a project cache. The system misinterpreted the request and executed a deletion command against the root of the entire drive. Because the operation bypassed the Recycle Bin, the data could not be recovered. (Source: Tom’s Hardware)

3. A user asked the OpenClaw agent to review emails and recommend deletions without taking action. When the conversation exceeded the model’s context window, critical instructions were compressed away, causing the agent to begin deleting messages. (Source: PC Mag)

4. A developer used Claude Code during an AWS migration. After relying on outdated infrastructure state information, the agent launched a destructive Terraform operation that removed both production environments, including years of database records and backup snapshots. (Source: Tom’s Hardware)

Why it fails: Anything an AI agent can reach, it can destroy by mistake. Destructive actions in production need a human confirmation step, and backups must live somewhere the agent cannot touch.

AI in Vendor Management

Even if your business manages AI risk well, the exposure can still come from your suppliers.

When the mistake originates with a third party, you can face the legal, contractual, financial, or reputational fallout anyway.

The usual culprits: unauthorized use of intellectual property, disclosure of confidential information, inaccurate outputs, brand-damaging content, and compliance failures.

Strengthen supplier oversight by writing these provisions into vendor agreements:

  • Disclosure of every AI system used to deliver your contracted services
  • Approval requirements for AI applications and permitted use cases
  • Limits on what company information may be processed by AI tools
  • Human review of AI-generated work products
  • Clear accountability for losses caused by AI mistakes

Contracts alone are not enough. Reinforce the expectations through onboarding guides, awareness training, and vendor-facing AI governance materials.
For a working example, review the Supplier AI Policy published by Cox Enterprises.

AI Resources for Small Business

Claude for Small Business Anthropic’s package of ready-to-run workflows and connectors (QuickBooks, PayPal, HubSpot, Canva, Google Workspace) that run inside Claude Cowork. No extra charge beyond your Claude license and tools you already pay for.

Grow with Google: AI for Small Businesses Free on-demand tutorials and live “Make AI Work for You” workshops covering AI for marketing, customer support, finance, and operations. Mix of self-paced and in-person formats. Open broadly, no eligibility gate on the tutorials.

ISED Toolkit for SMEs Deploying AI Canadian government toolkit drawing on G7 work. Practical, non-binding guidance on trustworthy AI governance, plus a public list of vetted Canadian AI suppliers who have completed Algorithmic Impact Assessments.

The Artificial Intelligence Guide for Small Businesses New Zealand’s National Cyber Security Centre guide. Practical, security-first framing on adopting AI without exposing your business to data and privacy risks.

US Chamber of Commerce: AI Training Guide for Small Business A curated index rather than a single course. Points to free workshops, Google’s Learn Essential AI Skills hub, Ethan Mollick’s resources, and recommended podcasts and newsletters for owners.