Making AI reliable: How to avoid AI guessing in your business

AI is a powerful tool, but its tendency to sometimes 'make things up' – what we call hallucination – can undermine its usefulness. For businesses, this isn't just a minor issue; it can lead to incorrect decisions and erode trust.

Why this matters for SMEs

Misinformation from AI can directly impact core business functions, from customer service and finance to operations and sales. Owner-managers and operations leads need to be confident that the information AI is providing is absolutely correct, especially when dealing with critical data like customer balances or stock levels.

What it looks like in practice

Consider Mary, the Finance Manager at 'Green Living Ltd', a 60-person Irish e-commerce business. Her team frequently gets queries about customer account balances and order statuses. Before implementing a robust AI system, these queries would often require a team member to manually check multiple systems, which was time-consuming and prone to human error when under pressure.

Before: A customer service agent asks an AI assistant, "What's Mrs. O'Connell's outstanding balance?" The AI, using a single model and without direct, verified access to the accounting system, might confidently reply with a figure that's close but incorrect, perhaps because it's based on an old summary or has misinterpreted the query. This leads to customer frustration and extra work for Mary's team to correct the record and apologise.

After: With a 'zero-tolerance' approach to AI hallucinations, the process changes. When the customer service agent asks the same question, the AI system doesn't guess. Instead, it:

1. Directly accesses the source: It connects securely to Green Living Ltd's accounting software, the definitive 'system of record'. It doesn't rely on general knowledge or external data.
2. Uses multiple checks: The query is processed by several independent AI models, all simultaneously querying this same precise record for Mrs. O'Connell's account balance. They look for specific, structured data points like outstanding_amount_in_euro and last_updated_timestamp.
3. Cross-references results: The system compares the data returned by each model. If all models retrieve the exact same, correct balance (e.g., "€125.50 on 2024-10-26"), it's confirmed.
4. Verifies or escalates: If there's any disagreement between the models, or if the data appears inconsistent, the system doesn't try to 'fill in the gaps' or provide a best guess. Instead, it either re-checks the source with stricter parameters or, for complete assurance, flags it for a human team member to review. The human gets the full context of the query and the conflicting AI findings.

This ensures that the answer given to the customer service agent (and in turn, to Mrs. O'Connell) is accurate and fully auditable, reducing errors and building trust.

What could go wrong

• Poor quality source data: If the underlying business systems contain errors, even the best AI system will struggle to give correct answers.
• Over-reliance on AI without verification: Assuming AI is always right, even when it flags uncertainty, can lead to overlooked issues.
• Complexity and cost: Implementing such robust systems can be more involved and potentially more expensive than simpler AI solutions.
• Setup and maintenance: The 'rules' for checking and validating data need careful definition and ongoing maintenance to stay relevant.

What to try this week

1. Identify a critical, factual query: Pick one common question in your business that currently takes time to answer or has a high impact if answered incorrectly (e.g., "What's the current stock of product X?" or "Has client Y paid their last invoice?").
2. Map the data source: Determine exactly which system and field holds the single, accurate answer to that question. This 'system of record' is crucial.
3. Consider a small pilot: Look for an AI tool or consultancy that prioritises accuracy and data verification over speed for critical queries, and discuss how they would handle your identified use case without making things up. Ask about their approach to 'grounding' AI in your specific business data.

Reliability is the foundation upon which the true value of AI for your business will be built.