When NOT to Use AI: A Decision Framework for Business Leaders

In May 2026, Kinney Drugs launched an AI phone assistant named “Burt” to handle prescription refills. By August, they pulled it back. Customers had reported wrong dosages, missed prescription notifications, incoherent calls, and an inability to locate longtime accounts. The pharmacy returned to its old touch-tone phone system.

In July 2026, Commonwealth Bank of Australia replaced 40+ customer service staff with an AI voice robot. Within a month, the system couldn’t handle call volumes, causing operational disruption. CBA reversed the layoffs, apologized to affected employees, and reinstated human workers.

Also in 2026, Medicare’s WISeR AI program for prior authorization caused 6-8 week delays for patients needing procedures like epidural injections. Doctors reported AI hallucinations in denial letters — one denial cited a condition the patient didn’t have. A University of Washington medical system had nearly 100 patients waiting for procedures due to WISeR-related delays.

Three deployments. Three failures. One pattern: AI was deployed where it shouldn’t have been.

The Pattern Behind AI Failures

AI failures rarely begin with a broken model. They begin when prediction is mistaken for judgment, when fluency is mistaken for accountability, and when “can we build this?” is asked before “should we proceed?”

The organizations that failed shared a common error: they treated AI as a universal upgrade rather than a specialized tool with specific trade-offs. AI introduces probabilistic outcomes instead of deterministic ones, opacity instead of transparency, and operational overhead instead of simplicity. Every AI deployment is a trade-off — and if you’re not explicitly evaluating those trade-offs, you’re not making a responsible decision.

The specific failure pattern in all three cases: AI was deployed in trust-critical, high-stakes, regulated interactions where being wrong costs more than being slow.

• Kinney Drugs put AI between patients and their medication. Wrong dosages and missed notifications aren’t inconveniences — they’re safety risks.
• Commonwealth Bank put AI between customers and their money. Call volume failures aren’t efficiency losses — they’re trust erosion.
• Medicare put AI between patients and medical procedures. Denial letters with hallucinated conditions aren’t technical errors — they’re healthcare failures.

The 4-Question Framework

Before deploying AI in any business process, pressure-test the decision with four questions:

1. Can a deterministic solution achieve 90% of the value at 10% of the complexity?

If your problem can be solved with rules, heuristics, or better data pipelines, AI is likely unnecessary. A rules-based system is auditable, predictable, and doesn’t hallucinate. An AI system is none of those things.

If a deterministic solution gets you 90% of the way there, use it. The remaining 10% isn’t worth the operational overhead, risk, and maintenance burden that AI introduces.

Example: If your customer routing problem can be solved with a decision tree based on 5 variables, don’t deploy an LLM to “understand” customer intent. Use the decision tree. It’s faster, cheaper, auditable, and never hallucinates.

2. Does the task run 50+ times per month?

AI automation carries fixed costs: implementation, integration, testing, change management, and ongoing maintenance. For high-volume processes, those costs amortize easily. For low-volume processes — tasks that occur dozens of times per month rather than thousands — the economics frequently don’t close.

Below 50 instances per month, the per-task time savings look attractive in isolation but don’t cover the total cost of ownership. Above 500 instances per month, automation economics typically improve substantially.

Example: If you’re processing 200 invoices per day, AI extraction makes economic sense. If you’re processing 30 invoices per month, a human with a spreadsheet is cheaper, more accurate, and doesn’t require model monitoring.

3. Does the decision require auditability and defensible reasoning?

In regulated environments — healthcare, finance, legal, insurance — decisions must be explainable, documented, and defensible to regulators, auditors, or counterparties. Current AI systems, including LLMs, do not inherently provide auditability, consistency, or defensible reasoning.

The appropriate AI role in regulated decision-making is decision support, not decision execution. AI can surface relevant information, flag anomalies, and draft initial analyses. But the decision — and the documented rationale — must sit with an accountable human.

Example: Medicare’s WISeR program used AI for prior authorization decisions. When the AI denied a procedure citing a condition the patient didn’t have, there was no accountable human who could explain why. The denial was an AI output, not a reasoned decision. That’s the auditability gap.

4. Is the interaction trust-critical?

For B2B firms and healthcare providers especially, customer relationships are long-term, high-value, and personally managed. Automated interactions at critical moments — a medical diagnosis, a financial decision, a customer complaint — carry disproportionate risk of damaging trust that took years to build.

If the interaction is trust-critical, keep a human in the loop. AI can prepare, summarize, and recommend. The human delivers.

Example: Kinney Drugs deployed AI for prescription communication — one of the most trust-critical interactions in healthcare. When Burt gave wrong dosage information, it didn’t just create a technical error. It damaged the trust between patient and pharmacy that is the foundation of the business.

The MIT CISR AI Decision Matrix

MIT CISR’s 2026 research, based on 30 executive interviews across 9 global companies, provides a complementary framework. The AI Decision Matrix maps decisions across two dimensions:

Ambiguity — how clearly data determines the answer. Low ambiguity means repeatable, predictable decisions. High ambiguity means multiple reasonable interpretations.

Risk — the consequences of getting it wrong. Low risk means reversible, low-impact decisions. High risk means significant financial, operational, or reputational impact.

This creates four decision types:

	Low Risk	High Risk
Low Ambiguity	Routine: AI can decide autonomously	Consequential: AI as decision support, human approves
High Ambiguity	Exploratory: AI as helper, human revises	Strategic: Humans must decide, AI can advise

• Routine decisions (low ambiguity, low risk): Fraud detection in low-value transactions, content recommendations, automated data extraction. AI can operate autonomously with monitoring.

• Consequential decisions (low ambiguity, high risk): Loan approvals, medical prior authorizations, compliance flagging. AI provides analysis; human makes the call.

• Exploratory decisions (high ambiguity, low risk): Marketing campaign ideas, content drafts, research summaries. AI expands options; humans evaluate and select.

• Strategic decisions (high ambiguity, high risk): Market entry, organizational restructuring, brand repositioning. AI can synthesize inputs and frame scenarios; humans must decide and commit.

The Kinney Drugs failure was a Consequential decision treated as Routine. The Medicare WISeR failure was a Consequential decision treated as Routine. The Commonwealth Bank failure was a Consequential decision treated as Routine.

The pattern is clear: most AI failures occur when Consequential decisions are automated as if they were Routine.

What to Do Instead

When you decide NOT to use AI, here’s what to do instead:

Use deterministic automation. Rules-based systems, decision trees, RPA, and structured workflows handle 80% of business process automation needs without AI’s overhead. They’re auditable, predictable, and don’t require model monitoring.

Fix the process first. If your process is broken, automating it with AI produces faster broken outcomes. Simplify the workflow, eliminate unnecessary steps, clean the data. Then evaluate whether AI adds value on top of the improved process.

Invest in human capability. The Robert Half 2026 report found that 71% of tech leaders say skills shortages caused project delays — with AI integration the #1 affected area at 64%. Sometimes the answer isn’t AI. It’s hiring, training, and retaining skilled people.

Use AI internally first. TechTarget’s analysis of AI failures recommends starting with internal applications before tackling customer-facing ones. Internal AI deployments carry lower risk — if the AI makes a mistake, it’s caught internally before it reaches a customer.

The AI Strategy for Business consultation includes an AI Fit Assessment: evaluating your proposed AI use cases against this 4-question framework, identifying which ones are genuinely AI-appropriate, and recommending deterministic alternatives for the rest.

The India Context

For Indian businesses, the “when not to use AI” question has a specific dimension. The Bain 2026 India Enterprise Technology Report found that 90% of Indian business leaders say their data foundations are weak. When your data is weak, AI produces weak outputs — confidently, at scale.

The KPMG Global Tech Report 2026 notes that Indian enterprises are “modernizing the entire IT stack all at once” — data, applications, cloud, cybersecurity, and AI simultaneously. That’s ambitious, but it means AI is often deployed on infrastructure that isn’t ready for it. The result is PoC fatigue: pilots that don’t scale because the foundation isn’t there.

For Indian SMEs especially, the question isn’t “should we use AI?” but “have we earned the right to use AI?” — meaning, is the data clean, are the processes defined, is the team ready? If the answer is no, the best AI strategy is to fix those things first.

The CTO Technology Advisory service helps organizations assess readiness: data foundation maturity, process standardization, and team capability — the prerequisites that determine whether AI will deliver value or become the next failure case study.

The Bottom Line

AI is not universally value-additive. In exception-heavy, low-volume, compliance-sensitive, or trust-critical processes, AI automation frequently increases cost and operational risk rather than reducing it.

The most expensive AI decision is the one you shouldn’t have made. Kinney Drugs learned this when their AI phone assistant gave wrong dosages. Commonwealth Bank learned it when their AI voice robot couldn’t handle calls. Medicare learned it when their AI prior authorization caused 6-week patient delays.

The framework is simple: if a deterministic solution works, use it. If the task is low-volume, don’t automate. If the decision requires auditability, keep humans accountable. If the interaction is trust-critical, keep humans in the loop.

AI is a trade-off, not an upgrade. Treat it that way.