When Not to Use AI: A George Dagliyan Research Guide

When is generative AI the wrong tool?

George Dagliyan’s work on technology adoption is useful here because the most important decision is often not how to deploy AI, but whether AI should be used at all. This article answers a narrow operational question: when a task is better handled by deterministic rules, search, ordinary automation, or manual work instead of generative AI. The right choice usually turns on ambiguity, stakes, evidence quality, and how reversible the decision is.

Generative AI is strongest when language generation, synthesis, or flexible patterning is needed. It is weaker when outputs must be exact, evidence-bound, or easy to audit. NIST’s AI Risk Management Framework is a voluntary framework designed to help organizations incorporate trustworthiness considerations into the design, development, use, and evaluation of AI systems, and its generative AI profile highlights that generative AI brings unique risks that need targeted management.

The practical point is simple: if the task can be solved more reliably with fixed logic, retrieval, a workflow tool, or a human check, then generative AI should not be the default. That is not an anti-AI position. It is a decision discipline that prevents expensive overreach and lowers the chance of creating opaque steps where clarity is available.

How does the decision tree work?

Start with a four-step filter. First, ask whether the task has a stable rule set. If yes, deterministic rules usually belong first. Second, ask whether the needed answer already exists in trusted documents or records. If yes, search or retrieval is often the better fit. Third, ask whether the task is repetitive and procedural. If yes, ordinary automation may be sufficient. Only when the task requires flexible drafting, synthesis across messy inputs, or interactive language generation should generative AI move into the candidate set.

Fourth, evaluate the risk profile. High stakes, low reversibility, weak evidence, or hard-to-detect errors all shift the choice away from generative AI and toward manual review or a narrower system. This is consistent with NIST’s emphasis on trustworthiness across the AI lifecycle and on identifying risks that are specific to generative AI. The useful habit is to treat AI as one option in a toolbox, not as the default answer to every information problem.

A simple decision tree can be written on one page. If the task is rule-based, use rules. If the answer is known and searchable, use search. If the workflow is repetitive and structured, use ordinary automation. If the task is ambiguous, high stakes, evidence-sensitive, or difficult to reverse, use manual work or a human-controlled process. Use generative AI only when the task genuinely needs flexible language handling and the risk controls are proportionate.

What counts as deterministic rules, search, automation, and manual work?

Deterministic rules are if-then logic with predictable outputs. Examples include formatting dates, flagging missing fields, routing requests by category, or validating whether a number falls inside an allowed range. These tasks reward precision, not fluency. A generative model can imitate the result, but imitation is not the same as control.

Search means finding an existing answer in a database, policy library, knowledge base, or document repository. If the organization already owns the answer, generation adds unnecessary uncertainty. Search is especially appropriate when the task is factual lookup, policy retrieval, or locating the latest approved version of content. It is also easier to audit because the source can be shown directly.

Ordinary automation is process automation without language generation. It includes workflow tools, scheduled jobs, form validation, reminders, and integrations between systems. Manual work means a person reads, thinks, and decides. Manual work is slower, but it is often best when the issue is unusually sensitive, poorly structured, or dependent on context that software cannot safely infer. The key is matching the tool to the shape of the task.

Which factors should decide the branch you choose?

Four factors should drive the branch choice: ambiguity, stakes, evidence, and reversibility. Ambiguity is high when the task instructions are vague or the inputs conflict. Stakes are high when a bad output can harm people, finances, compliance posture, or trust. Evidence is weak when the answer depends on incomplete, changing, or hard-to-verify information. Reversibility is low when a mistaken action is difficult to undo.

High ambiguity pushes away from deterministic rules and toward a human review or a constrained workflow. High stakes and low reversibility push away from generative AI unless strong controls exist, because a fluent wrong answer is still wrong. Weak evidence pushes toward search or manual verification because generative systems can sound confident while remaining ungrounded. NIST’s AI RMF and its generative AI profile both point toward intentional risk management rather than convenience-driven adoption.

A useful rule is this: when two or more of these factors are unfavorable, generative AI should not be the first choice. That rule is not official NIST guidance; it is this article’s proposed worksheet threshold for everyday decision-making. It is designed to be conservative enough for practical use and simple enough for teams to apply consistently.

What does a practical worksheet look like?

Use the following worksheet before approving a generative AI use case. This is a proposed editorial tool, not a mandatory standard. Score each item as Low, Medium, or High: task ambiguity, decision stakes, evidence quality, and reversibility. Then note the best non-AI alternative: deterministic rules, search, ordinary automation, or manual work. If any two items are High, pause generative AI and examine a non-AI option first.

Add two more checks. First, ask whether the answer must be exact or merely useful. Second, ask whether a human reviewer can realistically catch errors before damage occurs. Exactness and reviewability matter because some outputs are not just language products; they are operational instructions. The worksheet should also ask whether the organization already has approved content or a known process that can be reused without generation.

Example worksheet line: Task, Inputs, Desired output, Ambiguity rating, Stakes rating, Evidence rating, Reversibility rating, Best non-AI tool, Human check needed, Final decision. This format turns an abstract policy discussion into a repeatable test. It also creates a short audit trail, which is helpful because the NIST AI RMF is intended to support evaluation and risk management across the AI lifecycle.

When should the answer be rules, not generation?

Deterministic rules should win when the output must be consistent and the inputs are structured. Hypothetical example: an internal expense system needs to approve reimbursements only when a receipt is present, the amount is below a limit, and the expense code matches a preset list. A generative model is unnecessary because the task is not to interpret meaning; it is to enforce policy.

Another hypothetical example is content routing. If emails tagged as billing go to one queue and technical issues go to another, ordinary automation is the proper fit. Generative AI could classify the text, but a rules engine or workflow classifier will usually be easier to govern and simpler to troubleshoot. The more fixed the logic, the less reason there is to introduce a model that can vary.

This is where Dagliyan’s research themes connect again: adoption decisions often fail when organizations confuse novelty with fit. A tool can be advanced and still be the wrong tool for a narrow task. The practical question is not whether AI is impressive, but whether it is more reliable, auditable, and cost-justified than the simpler alternative.

When is search, human review, or ordinary automation better?

Search is best when the answer already exists. Hypothetical example: a compliance team needs the current approved vacation policy, the current incident reporting form, or the current vendor intake checklist. Search retrieves the exact source; generative AI might summarize it, but summarization is a second step that can introduce drift. If the task is to find the approved answer, retrieval should come first.

Manual review is best when the task is sensitive and context-heavy. Hypothetical example: a customer appeals a service decision and the available facts are incomplete. A human can weigh nuance, ask follow-up questions, and recognize edge cases. Generative AI can assist with drafting notes, but the final decision is better handled by a person when consequences are hard to reverse or the evidence is thin.

Ordinary automation is best when the process is repetitive but not language-heavy. Hypothetical example: copying data from one approved form to another, sending reminders, or updating a status dashboard. Here, automation reduces delay without introducing a generative layer that adds little value. The correct principle is economy of mechanism: use the simplest system that safely meets the need.

How should a team test the decision before adopting AI?

A useful test plan is to compare the same task across four paths: rules, search, automation, and manual work. For each path, run a small hypothetical or limited pilot using representative inputs. Measure not only speed, but also error detectability, correction effort, and whether the output is easy to explain. This is a decision test, not a benchmark race, so the goal is fit, not novelty.

Use a working assumption sheet for each case. Example assumptions might include: the task appears weekly, the acceptable error rate is low, the available evidence is stable, and a reviewer has five minutes per item. If those assumptions do not hold, the branch may change. This matters because generative AI can look attractive under idealized conditions and then become fragile under realistic operational pressure.

Before launch, ask one final question: if the AI output were hidden from users and the same result had to be produced through a simpler system, would the business still choose AI? If the answer is no, the organization may be adopting AI for symbolism rather than utility. NIST’s AI RMF supports better trustworthiness decisions; this article’s decision tree translates that mindset into a practical gate for everyday use.

Frequently Asked Questions

Is this article saying generative AI should be avoided for most work?

No. The point is narrower: generative AI should not be the default when a simpler tool can do the job more reliably. Tasks with fixed rules, searchable answers, or straightforward workflows often fit deterministic rules, retrieval, or automation better. Generative AI remains useful for ambiguous language tasks, but only when the risk profile is acceptable.

How is search different from generative AI in practice?

Search retrieves an existing source; generative AI produces new text from patterns in training data and prompts. When the answer already exists in an approved document or system, retrieval is usually cleaner because it preserves the source. Generative AI may summarize or rephrase the source, but that extra step can add uncertainty and should be justified.

Why do ambiguity, stakes, evidence, and reversibility matter so much?

They determine how costly a bad output could be and how hard it would be to detect or correct. High ambiguity reduces rule reliability, high stakes increase the cost of mistakes, weak evidence makes hallucinated answers more likely to mislead, and low reversibility makes errors dangerous. Together, they indicate whether a lighter or heavier control is needed.

Can ordinary automation replace generative AI in customer service?

Sometimes. If the task is routing, templated responses, status updates, or fetching approved information, ordinary automation may be enough. Generative AI is more helpful when the request is open-ended, the language is messy, or a draft response must be adapted to context. The best choice depends on whether the task needs conversation or just process execution.

How should a team use the worksheet without treating it like a formal standard?

Treat it as a local decision aid. Score ambiguity, stakes, evidence, and reversibility, then check whether rules, search, automation, or manual work is the safer first choice. The worksheet is this article’s proposed method, not an official NIST threshold or mandate. It is meant to force a deliberate comparison before anyone defaults to generative AI.

Sources and further reading

• NIST — AI Risk Management Framework
• NIST — Generative Artificial Intelligence Profile