What kind of AI does your business actually need?

Which kind of AI fits the work?

Start with the business question, then choose the least complex capability that can answer it reliably.

Artificial intelligence is often discussed as though it were one product. It is not. A demand forecast, an invoice reader, a policy assistant, a writing tool, and an autonomous service agent can all use AI, but they do fundamentally different jobs.

That distinction matters because many disappointing AI projects begin with the wrong question: “Which model should we buy?” A stronger starting point is, “What result must improve, what information is available, and how much variation or risk can the process tolerate?”

Microsoft’s current Cloud Adoption Framework makes the same strategic point. It advises organisations to begin with business problems, separate generative from nongenerative workloads, assess data and skills, and choose the simplest adoption path that meets the need. It also cautions that an AI agent adds unnecessary complexity when ordinary software or a straightforward knowledge assistant can do the job.

For a business leader, the useful taxonomy is therefore not a list of algorithms. It is a list of jobs that AI can perform. The following six kinds cover most practical business opportunities.

1. Predictive AI: What is likely to happen?

Predictive AI finds patterns in historical information and estimates a future outcome. The output is usually a number, probability, forecast, or risk score rather than a paragraph of newly written content.

A retailer might forecast demand by store and product. A maintenance team might estimate which machine has the highest probability of failure. A subscription business might identify customers at risk of leaving. A finance team might forecast cash flow under several assumptions.

Predictive AI is most useful when three conditions are present:

• The organisation has enough relevant historical data.
• The outcome can be defined clearly, such as a late payment, equipment failure, or product return.
• The prediction leads to a practical decision.

The third condition is easy to overlook. A churn score is not valuable by itself. It becomes valuable when the company has a sensible retention action, a team that can take it, and a way to measure whether that intervention worked.

Prediction is also not certainty. A model can say that a delivery is at high risk of delay, but a person or business rule still decides whether to reroute it, contact the customer, or accept the risk. Leaders should ask how often the model is wrong, which type of mistake is more expensive, and whether performance changes for different products, regions, or customer groups.

2. Detection and classification AI: What is this, and is it unusual?

Detection and classification AI sorts information into categories or flags activity that differs from an expected pattern. It is the quiet engine behind many fraud, quality, security, routing, and compliance workflows.

Examples include identifying a suspicious transaction, classifying an incoming service request, detecting a manufacturing defect, labelling an email as phishing, or flagging unusual access to a cloud system.

The business value is usually prioritisation. Instead of asking a person to inspect every transaction, alert, or request, the system helps direct attention toward the items most likely to matter.

This kind of AI requires careful threshold design. A very sensitive fraud system may catch more genuine fraud while also blocking more legitimate customers. A less sensitive system may create a smoother experience but miss costly events. There is no universal “accurate enough” setting. The threshold should reflect the consequence of each error and the organisation’s appetite for risk.

For high-impact decisions, classification should be an input to judgment, not a concealed final verdict. If a system affects employment, credit, healthcare, insurance, legal rights, or access to essential services, leaders need documented review, appeal, fairness testing, and named accountability.

3. Perception AI: Can the business read, hear, or see at scale?

Perception AI turns unstructured input into usable information. It includes document intelligence, speech recognition, language processing, computer vision, and video analysis.

An accounts-payable team can extract supplier names, dates, totals, and line items from invoices. A contact center can transcribe calls and identify recurring topics. A logistics business can read delivery documents. A factory can inspect product images. A property company can classify photographs from site inspections.

The most important distinction is between reading and understanding. Basic optical character recognition converts visible characters into text. Document intelligence goes further by identifying tables, layout, labels, values, and document types. That structured result can enter a workflow instead of remaining a block of text.

Perception AI often creates the input for another system. An invoice reader may pass extracted fields into validation rules. A call transcription service may feed a summarisation tool. An image classifier may send possible defects into a quality-review queue.

The safe operating model is “extract, validate, then act.” Confidence scores help route uncertain results, but confidence is not approval. Critical values such as bank details, identity information, contractual dates, and payment totals should be checked against business rules or reviewed by a person before a consequential action occurs.

4. Generative AI: Can it help create, explain, or summarise?

Generative AI produces new material in response to an instruction. It can draft text, summarise documents, explain complex information, generate images, translate content, or assist with code.

This makes it well suited to work where a useful first draft has value and some variation is acceptable. Common business uses include drafting proposals, preparing meeting summaries, simplifying policy language, producing campaign alternatives, writing product descriptions, and turning technical findings into an executive brief.

Its fluency can be deceptive. A well-written answer can still be incomplete, outdated, or wrong. Generative AI should therefore be treated like a fast junior assistant: useful for acceleration, but not an unquestioned authority.

A good use case has a visible reviewer, clear source material, and an output that can be checked before it causes harm. A poor early use case is one where the model independently makes a binding decision, moves money, changes a customer entitlement, or publishes regulated advice.

Leaders should also distinguish between productivity and business value. Generating more words is not automatically valuable. Useful measures include time saved to an approved output, reduction in rework, consistency with the brand or policy, and whether the user completes the task more successfully.

5. Knowledge-grounded AI: Can it answer from what the business actually knows?

A general-purpose model does not automatically know an organisation’s latest policies, contracts, product catalogue, service history, or permissions. Knowledge-grounded AI addresses that gap through retrieval-augmented generation, commonly called RAG.

RAG works like an open-book assistant. When a user asks a question, the system searches approved sources, retrieves the most relevant passages, and provides those passages to the language model. The model then composes an answer based on that evidence.

This pattern supports HR assistants, service-manual search, procurement guidance, compliance support, product knowledge, and internal research. It is especially useful when information changes frequently or must remain under the control of a content owner.

The hard part is not the chat box. It is the knowledge discipline behind it. Documents need owners, effective dates, permissions, sensible structure, and a refresh process. The assistant must not retrieve material the user is not entitled to see. It should cite its sources and decline to answer when evidence is missing, stale, or contradictory.

The right quality questions are straightforward:

• Did the system retrieve the correct source?
• Did the answer stay within that source?
• Did it answer the actual question?
• Did it include all important conditions?
• Did it correctly abstain when it lacked evidence?

Microsoft Foundry provides evaluation concepts for retrieval, groundedness, relevance, and response completeness. Even without discussing those tools with end users, the business should use the same principles when deciding whether a knowledge assistant is ready.

6. Agentic AI: Can it work through a changing sequence of steps?

An AI agent goes beyond producing an answer. It can reason about a goal, choose among available tools, inspect the result of an action, and decide what to do next.

Consider a support request that says, “My order has not arrived, I am travelling tomorrow, and the delivery address may be wrong.” A simple workflow struggles because the next step depends on what it discovers. An agent could retrieve the order, check delivery status, consult the change-of-address policy, ask for missing information, draft options, and route an exception to the right person.

Agents are valuable when work is genuinely variable, spans several systems, and requires decisions along the way. They are not the best answer for a stable process with fixed steps. Ordinary software is usually cheaper, faster, easier to test, and more predictable for that kind of work.

The more an agent can do, the more important identity and permission become. An agent should have its own governed identity or operate within clearly defined user permissions. It should receive only the data and actions required for its task. Irreversible or high-impact actions should require human approval, and every material action should be logged.

A worked example: one business, six different jobs

Imagine a regional distributor with thousands of products, multiple warehouses, and a busy service team.

Predictive AI forecasts demand by product and location. Detection AI flags unusual order patterns and possible duplicate invoices. Perception AI extracts details from supplier documents. Generative AI drafts product explanations and customer updates. A RAG assistant answers staff questions from current product, warranty, and returns policies. An agent handles complex order exceptions by checking several systems and preparing a recommended resolution for approval.

The company does not force every problem into one platform or model. Each capability has a different owner, control level, and success measure. The forecasting model is judged on forecast error and stock outcomes. The document system is judged on field accuracy and exception rates. The knowledge assistant is judged on grounded answers and correct abstention. The agent is judged on completed tasks, safe escalations, and action accuracy.

That is what a coherent AI portfolio looks like: shared governance, but use-case-specific design.

A practical selection method

Before approving an AI project, ask the team to complete one page with the following information:

• Business problem: What observable friction or missed opportunity exists today?
• Desired outcome: Which cost, speed, quality, risk, revenue, or experience measure should move?
• Decision type: Is the task prediction, detection, perception, creation, knowledge retrieval, or multistep action?
• Data: What information is needed, who owns it, and can the intended users legally and appropriately access it?
• Tolerance for variation: Must the same input always create the same result?
• Consequence of error: What happens if the system is wrong, incomplete, or unavailable?
• Human role: Who reviews, approves, overrides, or receives an escalation?
• Simpler alternative: Could rules, search, process redesign, or ordinary automation solve the problem?

Microsoft’s AI adoption guidance recommends prioritising use cases by strategic value, feasibility, available skills, data readiness, resources, and user desirability. This prevents a technically impressive demonstration from outranking a smaller project with clearer business value.

Implementation steps that work across all six kinds

Start with a baseline. Measure the current cycle time, error rate, cost, backlog, satisfaction level, or other relevant outcome before adding AI. Without a baseline, the organisation can report activity but cannot prove improvement.

Choose a narrow pilot. One document type, one team, one knowledge collection, or one decision is easier to evaluate than an enterprise-wide launch. Microsoft’s adoption framework recommends beginning with internal, non-customer-facing proofs of concept where possible.

Design controls early. Decide data boundaries, permissions, prohibited uses, approval points, logging, escalation, and shutdown authority before the architecture becomes difficult to change.

Test realistic exceptions. Include incomplete requests, poor scans, conflicting documents, unusual customers, sensitive data, adversarial instructions, system outages, and questions with no answer.

Train the people doing the work. Users need to know what the system can do, what it cannot do, when to challenge it, and how to report a problem. Adoption is not achieved by making a tool available.

Review outcomes, not enthusiasm. A positive demonstration is not evidence of repeatable value. Compare the pilot against the baseline, review errors, calculate operating cost, and decide whether to stop, refine, or scale.

Guardrails every leader should require

Microsoft’s responsible AI principles cover fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. Those principles become practical through a small set of requirements:

• A named business owner remains accountable for the outcome.
• Users know when they are interacting with AI.
• Access follows least privilege.
• Sensitive data has an approved purpose and retention treatment.
• High-impact and hard-to-reverse actions require human approval.
• Generated answers show evidence where evidence matters.
• The system can refuse or escalate rather than invent.
• Quality, safety, security, cost, and user feedback are monitored after launch.
• The organisation can pause the system and investigate an incident.

Measures that reveal real value

Use a balanced scorecard rather than a single productivity claim.

Business outcome: revenue protected or created, cost per completed task, cycle time, backlog, service level, or risk reduction.

Quality: field accuracy, forecast error, groundedness, rework, false positives, false negatives, and correct escalation.

Adoption: active use by the intended team, completion rate, override rate, accepted suggestions, and reasons for abandonment.

Risk: privacy events, unauthorised retrieval, harmful output, consequential errors, policy violations, and unresolved complaints.

Operations: availability, latency, unit cost, incident volume, model or data drift, and time to restore a safe service.

The best AI is not the system that appears most intelligent. It is the system that improves a real outcome, behaves within clear boundaries, and leaves a person accountable for the result.

• https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai/strategy
• https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai/plan
• https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/business-strategy-plan
• https://learn.microsoft.com/en-us/azure/architecture/ai-ml/guide/choose-ai-model
• https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview
• https://learn.microsoft.com/en-us/azure/foundry/concepts/evaluation-evaluators/rag-evaluators
• https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/overview?view=doc-intel-4.0.0
• https://learn.microsoft.com/en-us/agents/center-of-excellence/responsible-ai

This article is original editorial work for a business audience. Product facts were checked against the official sources above. Any performance figures are targets to validate against your own baseline, not vendor guarantees.

Bring us the operating need, risk, or opportunity. We will connect the people, security, AI and enterprise systems required to act.