Which AI Do You Actually Need?

General AI, specialists, coding agents, local models — or possibly none of them.

Choosing an AI tool often starts with a list of product names.

That is convenient for product pages. It is less convenient for deciding what you actually need.

You probably do not want an AI subscription as an end state. You want a researched answer, a finished document, working code, less admin, a private processing path, or a task that was previously too expensive to bother doing.

So the order here is deliberately less exciting:

Purpose → Means → Tool.

AI tools are multiplying faster than most reasons to use them.

1. Can the work be finished without AI?

Before comparing models, ask whether the task needs one at all.

A short factual edit, an existing spreadsheet formula, a five-minute call with the person who actually knows the answer — these can all beat opening a new category of software.

Manual and human are not fallback failures.

Doing nothing new is also a valid technical architecture. It is unusually easy to maintain.

2. Can one general AI finish the job?

If AI is useful, start with the smallest broad tool that can complete the recurring work.

We use general AI as an editorial category for products that cover several kinds of knowledge work: writing, questions, files, analysis and similar tasks. It is not a formal industry standard.

The first comparison is not which underlying model wins a benchmark.

It is whether the product fits the work surface you return to every week:

• can it work with the information you already have?
• does it produce something you can actually use next?
• does it fit the ecosystem you already operate?
• will you use the extra capability often enough to notice it?

For ordinary writing, synthesis, file analysis and everyday research, one general AI may be enough.

If it is, stop there.

3. Check the specialist functions you already have

General AI products increasingly contain narrower workflows inside the broader product.

OpenAI documents deep research in ChatGPT. Anthropic documents Research in Claude. Google documents Deep Research in Gemini Apps. The exact plans, limits and regional availability change, so those details are treated as volatile facts rather than permanent prose.

The useful question is:

Is the missing capability already included in the product you are paying for?

If yes, try to finish the purpose there before buying a second research subscription.

An included specialist is not automatically as good as a dedicated product. It simply deserves to be checked before another recurring charge is added to the experiment.

4. Name the missing thing before naming another product

When one general tool is not enough, classify the gap.

Quality

The output itself is not good enough.

Workflow

Several steps still need to be stitched together by hand.

Completion

The AI gives material; you need the finished presentation, video, application or other deliverable.

Context

You keep re-supplying the same files, history or working assumptions.

Capability

The tool could make a previously uneconomic task practical.

If the gap is still difficult to describe after this, the case for another subscription is probably not yet mature.

5. Add a specialist only when the gap survives the check

A specialist is not a higher tier of intelligence by definition.

It is a separate product that earns its place by being materially better at one relevant job: the output, the workflow, the context, the completion step, or a new capability.

If removing it tomorrow would not meaningfully break the workflow, the subscription may be optional in the most literal sense.

P04 goes deeper into that test.

6. Coding AI is a different decision surface

Coding is where product categories become especially easy to blur.

A model, a coding agent, an IDE assistant and an application-building platform can all be described as an AI coding tool, but they change different parts of the development process.

OpenAI currently describes Codex as a dedicated coding agent. Anthropic documents Claude Code as a coding-focused development surface. That is enough for the P01 decision: coding agents should not be treated as merely stronger versions of general chat.

Before choosing one, decide whether you want to stay in the IDE, delegate a bounded repository task, choose your own model, or obtain a finished application rather than code.

P05 is the coding decision hub.

7. Local and API are constraint decisions, not prestige levels

Local AI is not the graduate programme after cloud AI.

It is a hosting decision.

A local or hybrid route becomes rational when there is a real constraint: confidential data, offline operation, infrastructure control, reproducibility, a predictable workload, or experimentation with model/runtime choices.

Local runtimes such as LM Studio can run downloaded models on the user's machine and can operate offline after the necessary assets are available. That does not make local = private a universal rule. The application, network, logs, model licence and surrounding stack still matter.

API access is similarly about control and integration. If the purpose is to call models from software, route work programmatically or build a repeatable system, an API may be the right surface.

If the main result is merely more configuration, the configuration has at least delivered on schedule.

8. A practical decision map

Writing, ordinary analysis, file work, everyday research

→ Start with one general AI.

Repeated deep research

→ Check the research workflow already included in your current general AI.

One output or workflow is clearly weak

→ Consider a specialist.

Repository work with commands, tests and review

→ Move to the coding-agent / IDE decision.

Confidentiality, offline use or hosting control is the actual constraint

→ Evaluate local or hybrid.

The capability needs to live inside your software

→ Evaluate API / routing.

None of the above

→ Keep what you have, use a free tier, ask a person, or add nothing.

The last answer is not an edge case. It is one of the reasons to have a decision tree in the first place.

9. What would change our mind?

This map should change when the products change structurally.

If general AI products absorb more useful specialist workflows, the need for separate specialist subscriptions falls. If dedicated products create a much larger workflow or completion advantage, the case for adding them gets stronger.

Cheaper secure cloud options, easier local runtimes, different data-handling terms, or coding surfaces that converge would all move the boundaries again.

That is why the durable part of this article is the decision structure rather than the current model names.

The product teams will continue to help by changing those.