What companies keep doing wrong when implementing AI | MarTech

AI implementations don’t always go as planned. While the technology promises efficiency and innovation, real-world implementations often create new problems – and more human work.

When the AI ​​promise meets business reality

Before we worry about AI replacing humans, it’s worth examining how AI actually performs in the real world – where automation often creates more work, not less.

Ten years ago, IBM announced with much fanfare that Watson for Oncology was as accurate as human doctors at reading X-rays, CT scans and other reports. In some regions lacking oncologists, IBM even promoted Watson as a potential replacement for physicians.

But reality soon surfaced. According to ASH Clinical News, internal documents showed that Watson made unorthodox and unsafe recommendations when provided with synthetic (rather than real) patient data. Ultimately, IBM sold Watson Health’s data and analytics division to a private equity firm for $1 billion in 2021 – after investing more than $5 billion.

IBM was not alone. Remember Zillow Offers?

Zillow built an AI model to predict home values ​​and aggressively bought homes based on those predictions. The algorithm consistently overpaid, leading to half a billion dollars in losses and massive layoffs. The program collapsed in less than a year when the algorithm failed to adapt to a cooling housing market.

Dig deeper: Implementing AI seamlessly is a fast track to failure

Even recent implementations still miss the mark

You might say: “But those examples are years old. Companies have learned their lessons by now.”

Indeed, the capabilities of AI have improved dramatically in a short time. But the rush to release AI-powered updates hasn’t slowed – and not all deployments are being handled well. Unfortunately we had a front row seat due to a more recent misstep.

Like many small businesses, we rely on QuickBooks Online from Intuit to run our operations. Recently, QuickBooks rolled out an AI-powered version of the platform. For us it has been nothing short of a disaster.

This is what we encountered:

• Forced adoption: Unlike other platforms where customers could sign up or test new features, Intuit pushed us to the AI ​​version.
• Faulty machine learning: Although QuickBooks is trained in transactions, it often miscategorizes payments based solely on dollar value. If a supplier sent one invoice for €1,000, all invoices for that supplier were recorded as €1,000.
• Coding issues: Payments to contractors were recorded under QuickBooks payment instead of the contractor’s name.
• Hallucinations in accounting: Categories were randomly assigned in ways that neither we nor our accountants could explain or resolve.
• Passing on costs to customers: The problems got so bad that we had to pay our accountants thousands of dollars to fix the problem, with no solution.
• Poor communication: No notification of the change, no documentation, and no guidance on how to roll back.
• Broken workflows: Critical functions, such as billing, were disrupted. At one point, email addresses were dropping invoices altogether (including mine and the customer’s) and emails were being marked as spam.

The biggest sin is that QuickBooks is the core of our business. Cash flow, payroll, and customer billing all depend on it. When AI upgrades destabilize that core, the consequences ripple through the entire organization.

And this isn’t unique to QuickBooks. These examples – IBM Watson, Zillow, Intuit – remind us that AI implementation is not just about technology. It’s about trust, communication and responsibility.

Dig deeper: Your AI strategy is stuck in the past. Here’s how to fix this

Key insights for companies deploying AI

Each example shows that AI implementations fail not because the technology is underpowered, but because the implementation lacks care. These are the principles that can prevent innovation from turning into disruption.

• Don’t force change on customers: Allow opt-ins and pilots before mandating a new version.
• Validate in the real world, not just in the lab: Extensive testing with real customer data and workflows.
• Design a rollback path: Customers need a quick way back if something breaks.
• Prioritize communication: Explain what is changing, why, and how users should adapt.
• Respect the business-critical nature of your tool: The more essential the product, the higher the standard for reliability should be.
• Measure the impact downstream: An upgrade to AI can impact payments, compliance, or customer relationships in ways that go far beyond the software itself.

Building AI that deserves trust

AI has the potential to transform industries, but poor implementation can cause real damage. The costs of rushing AI into production without testing, communication and accountability are not borne by software companies. It is a matter for the companies and the people who depend on them.

Correcting AI’s mistakes often requires more human work, not less. The real winners will not be those who ship first, but those who build systems that are trustworthy, transparent and reliable.

Technology must make companies possible. When intelligence in AI is not supported by thoughtful design, it becomes both a technical and business failure.