When AI Is the Wrong Tool - The Case for Boring, Deterministic Systems

There is a particular kind of project failure that has become common over the last two years. A business identifies a process that feels slow or expensive, puts a large language model in the middle of it, and declares the problem solved. Six months later the process is slower, more expensive, and harder to audit than it was before, and somebody quietly restores the old system.

The mistake is rarely the technology. It is the assumption that AI is a general-purpose upgrade rather than a specific tool with a specific shape, and that anything a human used to do is therefore a candidate for it.

The scale of the rollback

This is not a fringe pattern. MIT’s “The GenAI Divide: State of AI in Business 2025” report examined 300 public deployments alongside interviews and surveys, and found that roughly 95% of generative AI pilots delivered no measurable impact on profit and loss. S&P Global Market Intelligence’s 2025 survey found that around 42% of companies abandoned most of their AI initiatives that year, up from 17% the year before.

The staffing side tells the same story. Talent consultancy Robert Half found that nearly a third of companies that cut roles because of AI have since reopened and rehired for those exact positions. Klarna is the best documented example: after announcing in early 2024 that its AI agent was handling the work of 700 people, the company began rehiring human agents in 2025. Its CEO, Sebastian Siemiatkowski, put it plainly, saying the business had focused too much on efficiency and cost and that the result was lower quality. Commonwealth Bank of Australia went through a compressed version of the same cycle, cutting 45 call centre roles in favour of a voice AI system and reversing the decision within weeks after call volumes rose and service quality fell.

None of this means the technology does not work. It means a lot of organisations put it somewhere it did not belong.

What goes wrong, and why it is predictable

Look closely at the well-known failures and they share a structure.

Air Canada’s chatbot told a customer they could buy a full-price ticket and claim a bereavement discount retrospectively. The airline’s actual policy required the claim to be made before travel. A Canadian tribunal ruled the airline had to honour what the chatbot said. The underlying task, “state the company’s refund policy correctly”, is a lookup. There is one right answer, it is written down, and it changes rarely. A generative model was asked to produce that answer probabilistically, and it produced a plausible one instead of the correct one.

McDonald’s ended its AI drive-thru trial in 2024 after orders came back with extra drinks nobody asked for and ketchup packets added to ice cream. Order capture is a constrained problem: a finite menu, a finite set of modifiers, and a hard requirement that the output match what the customer said. The speech recognition part is genuinely an AI problem. The order construction part is not.

Deloitte refunded the Australian government $290,000 of a $440,000 fee in 2025 after a commissioned report was found to contain fabricated academic references, a made-up quote attributed to a Federal Court judgment, and references to case law that did not exist. Citations are the most verifiable content in any document. They either resolve to a real source or they do not, and checking that is a database query.

The pattern in all three: a task with a single verifiable correct answer was handed to a system optimised for producing fluent, likely answers. Fluency and correctness overlap most of the time, which is exactly what makes the failures hard to catch. A rules engine that does not know the answer returns an error. A language model that does not know the answer returns a sentence.

The actual dividing line

The useful question is not “could AI do this?” It is “what does failure look like, and can I tell when it happens?”

Deterministic systems are the right answer when:

• The rules are written down and finite. Pricing tiers, eligibility criteria, VAT treatment, shipping bands, refund windows. If you can express it as a decision table, build a decision table. It is cheaper, faster, testable, and it will produce the same output next year.
• The output must be exactly right, not approximately right. Financial calculations, invoicing, tax, compliance checks, medical dosing, anything with a regulator attached. “Usually correct” is a category error in these domains.
• You need to explain the decision. A rules engine can tell you which rule fired. Under the EU AI Act and UK data protection law, “the model decided” is an increasingly uncomfortable position for automated decisions affecting individuals.
• The input is already structured. If data arrives as fields in a form or rows in a database, you do not need a model to interpret it. Validation logic and SQL are decades ahead on reliability and cost.
• Failure is silent. If a wrong answer looks exactly like a right answer and nobody downstream checks, the error rate is not 2%, it is 2% compounding invisibly.

AI is genuinely the right tool when:

• The input is unstructured and messy. Free-text emails, PDFs with inconsistent layouts, call transcripts, scanned documents, customer reviews. Extracting structure from mess is the thing traditional systems are worst at and models are best at.
• The problem space is genuinely open-ended. Drafting, summarising, brainstorming, translating, rewriting for tone. There is no single correct output, so there is nothing for a deterministic system to be deterministic about.
• The rules exist but nobody has ever written them down. Sentiment, intent classification, tone, relevance, “does this support ticket sound urgent”. You can gather examples far more easily than you can specify criteria.
• A human checks the output before it matters. Draft, review, send. Suggest, confirm, execute. The review step is what converts a probabilistic system into a reliable one.
• Approximate is genuinely good enough. Ranking, recommendation, prioritisation, first-pass triage. Getting the ordering roughly right beats not doing it at all.

The hybrid is usually the answer

Most real processes are not one or the other. They are a messy front end and a rule-bound back end, and the failures come from applying one approach to the whole thing.

Take an insurance claim submitted as a scanned letter. Reading the document, pulling out the policy number, the date, and the described incident is an AI problem. Deciding whether the policy covers the incident is not. That should be a rules engine reading the policy terms, and it should be able to explain which clause it applied.

Or customer support. Klarna did not conclude that AI had no place in support. It moved to a hybrid model where AI handles routine, high-volume queries and people handle anything requiring judgement. That is the same split: use the model for understanding what the customer is asking, use deterministic routing and written policy for what the company actually commits to.

The design principle is straightforward. Let AI handle interpretation. Let deterministic systems handle decisions and commitments. When a model generates something the business will be held to, put a verification step between the two, and make that verification deterministic.

Questions worth asking before you build

Before adding AI to a process, work through these honestly:

1. Can I write down the rule? If you can, and it fits on a page, you have just specified a deterministic system. Build that instead.
2. What is the cost of a confident wrong answer? Multiply it by a realistic error rate. Air Canada’s chatbot was cheap right up until it was a legal precedent.
3. How would I detect a wrong answer? If the honest answer is “a customer would complain”, you have no detection mechanism, you have a complaints queue.
4. Does the same input need to produce the same output? If yes, and you cannot tolerate drift when the model is updated, that is a hard constraint against a hosted model.
5. What was the problem before I heard about AI? If you cannot articulate the problem without mentioning the technology, you are solving for the technology.
6. Is a human reviewing this, genuinely? Not “there is a review step in the process diagram”. Does someone actually read the output, and do they have time to?

The unglamorous conclusion

A well-designed rules engine, a validated form, a properly indexed database, and a scheduled job will solve more business problems, more cheaply and more reliably, than most of what was funded as AI transformation over the last two years. That is not an argument against AI. It is an argument for putting it where it earns its keep, which is the messy, unstructured, judgement-heavy work that traditional systems genuinely cannot touch.

The businesses getting real value from AI in 2026 are not the ones that deployed it most widely. They are the ones that were willing to look at a process and say, honestly, that this part does not need it.

If you are trying to work out which parts of your process are which, get in touch. We would rather tell you a simple script will do the job than sell you something more complicated.

Sources: MIT report on generative AI pilot failure rates · Klarna reverses AI push, Forbes · Employers reversing AI-driven layoffs, CNBC · Moffatt v. Air Canada, McCarthy Tétrault · McDonald’s ends AI drive-thru trial, Global News · Deloitte partial refund over AI errors, Fortune