A customer email arrives: “Read this message, find the account in the CRM, decide its priority, draft a reply and update the ticket.” It sounds like one AI use case. It is at least five different jobs: generating text, classifying a request, retrieving a record, applying rules and taking an action. Calling the whole system “generative AI” makes the technology—and its controls—harder to choose.
Start with the job, not the label
Generative AI produces or transforms content—such as text, code, images and audio—from patterns learned in data. It fits tasks where the desired result is a draft, summary, explanation, translation, variation or prototype, and where that result can be checked in proportion to the consequence of an error.
By itself, it is not a database, search engine, deterministic rule, reliable forecast or authorized workflow actor. A product may combine all of those capabilities, but the generative model is only one part of the system.
The useful question is therefore not “where can we add generative AI?” It is: what result must this step of the work produce, and what evidence will show that the result is acceptable?
Generation is not knowledge
A language model produces a plausible sequence of tokens for the context it receives. That mechanism can create a clear, relevant and useful answer. It is not the same as retrieving a fact from a verified source.
Compare two requests:
- “Rewrite this procedure so a new colleague can follow it.”
- “What is the current status of order 4821?”
The first is mainly a content-transformation task. The second requires access to a precise, current record. A model can phrase either answer well; only a trustworthy connection to the source can establish the real order status.
Retrieval and generation remain separate even when they appear in one chat response. Retrieval places an existing source or record into context; the model turns that context into an output. The combination can improve relevance and make checking easier. It does not guarantee that the right evidence was retrieved, that it is current or that the answer represents it accurately.
Fluency is therefore not evidence. When an answer can affect money, rights, customers, security or operations, the source, data and reasoning need consequence-based verification. How to verify a generated answer is a different problem from making it read well.
Five different jobs hidden behind “AI”
A real system can pass through every lane below. This is not a regulatory or scientific taxonomy. It is an M4TIC task-fit map for identifying which component produces each result and which control question follows. A product name does not prove which lane a workflow step occupies: a chatbot may generate a response, retrieve data and call tools in the same conversation.
1. Generate
Output: a draft, summary, translation, image, code sample or transformation.
Control question: can the result be evaluated before it creates a consequence?
Typical failure: plausible content that is false, incomplete, off-brand or too similar to existing material.
2. Predict or classify
Output: a score, probability, label or category.
Control question: against which samples, thresholds and costly error types has the result been validated?
Typical failure: a wrong classification that looks precise because it is expressed as a number or decisive label.
3. Retrieve
Output: an existing record, document, source or indexed item.
Control question: is the source correct, current and authorized for this person and purpose?
Typical failure: missing, stale, irrelevant or wrongly sourced information.
4. Apply rules
Output: a repeatable result derived from explicit conditions.
Control question: is the rule complete, tested, versioned and able to handle exceptions?
Typical failure: incomplete or incorrectly implemented logic—not generative variability.
5. Act
Output: a modified record, sent message, initiated payment or triggered process.
Control question: who authorized the action, what limits apply, and how can it be stopped or reversed?
Typical failure: an unauthorized, duplicated or hard-to-reverse effect.
Governing a system that combines lanes means separating those moments instead of treating them as one “AI answer.”
Where generative AI fits well
Generation creates value when a task permits more than one acceptable result and a person or control can assess the output without cancelling the benefit.
Suitable tasks include:
- preparing a first draft from approved information;
- summarising a document while retaining links to its sources;
- adapting tone, format or level of detail;
- translating with terminology and context review;
- explaining one concept to different audiences;
- proposing alternatives for a prototype;
- assisting with code when tests and review remain part of the process.
The advantage is not that the model “understands like a person.” It can rapidly produce variations that respond to instructions and examples. Variation is valuable when exploring a solution space; it becomes a liability when a process requires the same exact answer every time.
“New content” also needs careful wording. It means the output is produced at that moment. It does not guarantee truth, uniqueness, ownership or an absence of similarity to existing material. Those questions depend on inputs, product terms, use and jurisdiction and need their own assessment.
Classification through a generative model trades consistency for flexibility
A language model can return a label or extract fields from unstructured text. For example, it can read requests written in many different styles and propose a category, urgency level and key details. This can be useful when the schema changes, examples are limited or the language is too varied for simple rules.
The generative interface does not remove the need for validation. When volume is high, categories are stable and a wrong classification has material consequences, a specialised classifier, explicit rules or a combination may provide better consistency, more predictable cost and a clearer audit trail.
Compare the approaches on the real work:
- quality by category, not only the overall average;
- the most expensive false positives and false negatives;
- stability across wording and languages;
- treatment of uncertain cases;
- review and correction cost;
- behaviour when the input format changes.
Using a generative model to classify is an architectural choice. It does not turn classification into content generation.
When retrieval, rules or conventional software are better
If a correct answer already exists in an authorized source, the main job is to retrieve it. If a result follows a stable rule, the job is to apply that rule. If an exact calculation is required, software designed for calculation should perform it. Generation may improve the interface around these steps; it should not replace the mechanism that produces the operational truth.
Prefer a deterministic path when:
- the same condition must produce the same outcome;
- the logic is explicit and changes infrequently;
- the organization must reconstruct exactly why a decision occurred;
- an error has a high or hard-to-correct consequence;
- variable wording adds no value;
- a query, formula or rule already solves the task more simply.
This does not require removing the model from the experience. A system can retrieve a record with a query, apply a rule in code and use a model only to explain the result in plain language. Strong systems often combine components rather than asking generation to do everything.
Automation and agents begin where action appears
Generating the phrase “update the ticket” changes nothing. To create an effect, the system needs a tool or integration, the correct identity, sufficient permission and a policy that determines whether the action may proceed.
That separation matters:
- the model can propose an action;
- an application can validate arguments and policy;
- a person or rule can approve it;
- a tool can execute it;
- logs and monitoring must record the outcome.
When a system selects among tools or paths within goals and limits, it enters the territory of AI agents. When several applications need a shared contract for exposed capabilities, MCP may become relevant. Neither capability is implied by the fact that a model generates text.
The limits that change the buying decision
Limits are not a footnote after a list of benefits. They determine whether the use case fits.
- Variability: the same input can produce different outputs. Decide how much variation the process can tolerate.
- Confabulation: the model can produce false or unsupported claims. The broader AI failure map helps identify where the problem enters.
- Context: the model does not automatically possess current company data. The system must provide necessary context and control its provenance and access.
- Prompt sensitivity: wording, order, examples and retrieved content can change the result.
- Evaluation: “looks good” is not a criterion. Use representative samples, edge cases and thresholds linked to consequences.
- Data, privacy and security: transmission, retention, use and permission depend on product and configuration. Private AI for business covers the data boundary.
- Intellectual property: do not assume originality or ownership; assess the specific use.
- Operating cost: include review, integration, monitoring, correction and exceptions—not only the price per call.
A use case remains attractive when its value survives these controls. If it works only by assuming every output is correct, every source is current and no review is needed, the project is not yet defined.
A task-fit table for the first use case
Scroll sideways inside the table to compare every column
| Question | Signal for generation | Signal for another component |
|---|---|---|
| What is the output? | Draft, summary, explanation, transformation or variation | Exact record, calculation, rule, score or action |
| How many results are acceptable? | Several formulations can satisfy the need | The same condition requires the same result |
| How will an error be detected? | Review or testing is available before use | The error is hard to see or appears after the effect |
| What is the consequence? | Low, reversible and containable | High, irreversible or subject to specific obligations |
| Which sources are required? | Approved context can be supplied and cited | An authoritative, current source must be queried directly |
| Must the system act? | The model prepares a proposal | Permissions, approvals, logs and revocation are required |
The table does not produce a universal score. It makes conditions visible before products are compared.
Choose a proportionate first use
Select a bounded, reversible task that is already understood well enough to define an acceptable result. Collect representative examples. State what must remain unchanged, who checks the output, which data may not enter the system and when the case must move to a person.
Then split the workflow into five questions:
- What must be generated?
- What must be predicted or classified?
- Which data must be retrieved, and from where?
- Which rules must remain deterministic?
- Which actions are allowed, approved and reversible?
This decomposition avoids two opposite mistakes: using a generative model where a rule is enough, and building a rigid sequence where language variation is the actual problem. The technology follows the task boundary.
Sources and further reading
Primary guidance behind this article.
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1)National Institute of Standards and Technology
Risk profile covering confabulation, evaluation, privacy, security and intellectual property in generative AI.
- Generative AI Technologies and Their Commercial Applications (GAO-24-106946)U.S. Government Accountability Office
Authoritative overview of generative technologies, commercial applications and principal limitations.
- Explanatory memorandum on the updated OECD definition of an AI systemOECD.AI
Updated definition distinguishing outputs such as predictions, content, recommendations and decisions.
- Framework for the Classification of AI SystemsOECD.AI
Framework for describing AI systems and contexts without reducing every capability to generation.
- Attention Is All You NeedVaswani et al., NeurIPS 2017
Original transformer architecture paper; technical context rather than business-selection evidence.
