How to choose the right AI model and tools for the task

The selection problem has changed

Choosing AI used to mean choosing a chatbot. You opened a conversation, supplied context, and judged the answer. That remains sufficient for many small jobs, but it is no longer the whole market. A modern agent can search the web, inspect a drive, edit local files, query a CRM, run code, create an artifact, and sometimes execute the final action. The model is now one component in a working system.

Ethan Mollick makes this shift the center of his Summer 2026 guide to using AI. His low-stakes advice is deliberately simple: “pick the one you like.” When the work is consequential or extended, however, his recommendations move toward stronger models, higher thinking settings, agentic work surfaces, and carefully chosen connections. That distinction is more useful than a universal model ranking.

A powerful model with no access to the required evidence can only improvise. A broad tool catalog attached to a weakly governed agent can act quickly for the wrong reasons. The practical unit of choice is therefore the whole task system: model, reasoning effort, work surface, tools, permissions, and review.

Write a six-line task brief first

Do not begin with a product comparison. Begin with the work. A compact task brief exposes what the AI must reason about, what it must reach, and what it must never do. It also prevents tool shopping from becoming a substitute for process design.

Use six fields. Outcome describes the finished deliverable or changed state. Stakes describe the cost of a wrong answer or action. Evidence names the facts and sources required. Systems name where that evidence and the destination live. Action boundary says whether the AI may read, draft, recommend, or execute. Sensitivity covers private, customer, financial, legal, health, security, or regulated material.

• Outcome: “Prepare a renewal brief with cited risks,” not “look at this account.”
• Stakes: low, material, or high consequence—and who remains accountable.
• Evidence: current records, documents, public sources, calculations, or media.
• Systems: web, local files, email, CRM, analytics, design software, or codebase.
• Action boundary: answer, create an artifact, prepare a draft, or make a change.
• Sensitivity: what may enter the model context, logs, exports, and connected services.

Choose the model by reasoning difficulty and consequence

Model choice matters most when the task is ambiguous, unfamiliar, long, adversarial, or expensive to get wrong. For routine rewriting, brainstorming, classification, or low-stakes questions, a current default model is often enough. Favor the interface you will actually use, then check the result. Paying extra for maximum reasoning on every small task adds latency and cost without automatically improving the workflow.

Move to a frontier model and a higher reasoning or thinking setting when the job requires reconciling conflicting evidence, following many constraints, planning across several stages, auditing an important document, diagnosing a difficult technical problem, or handling high-stakes subject matter. Stronger performance reduces error; it does not remove the need for authoritative sources or a qualified human decision-maker.

When the task depends on a particular medium, capability fit can matter more than the general leaderboard. Native image, audio, video, long-document, or code handling may determine which system can see and manipulate the work properly. Test the exact input and output you care about. A benchmark score does not prove that a product preserves spreadsheet formulas, edits video coherently, or returns traceable citations.

Choose where the agent should work

A chatbot is an answer surface. An agent surface is a work environment. Hosted agents are useful for delegating research and artifact creation that can continue after you close the app. They are convenient on a phone and easier to isolate from the rest of your computer, but they can only reach the files and applications exposed by the provider.

A local agent is better when the job spans a codebase, a collection of private files, command-line tools, or desktop software already on your machine. It can inspect the actual working state, make changes, run checks, and leave a detailed record. That extra reach also increases the consequence of a bad instruction, unsafe dependency, or malicious text encountered along the way.

Use a conversational surface when you need an answer. Use a hosted work surface when you need a bounded deliverable assembled from approved connections. Use a local work surface when the source material and tools live on your computer or when you need to inspect the work as it happens.

Add tools only for capabilities the task is missing

Tools do not make a model generally smarter. They give it a specific ability: search current information, retrieve a record, manipulate a file, run code, create a design, or change a system. Select them by tracing the task from required evidence to finished outcome. If a capability does not serve that path, leave it disconnected.

Use web research when facts are current, external, or need citations. Use an MCP server or first-party connector when the authoritative evidence lives in a business system. Use a filesystem or coding tool when the agent must inspect and change many local files. Use a specialist media tool when the output must be rendered or edited in its native form. Use an Agent Skill when the missing ingredient is a repeatable method, standard, or review checklist rather than access to new data.

A plugin may package several of these pieces together. Evaluate the pieces, not the label. Ask which tools become available, which instructions are loaded, who publishes them, what identity they use, what data they receive, and which actions they can take. Ten narrow, well-described tools are usually easier to govern than a marketplace-sized bundle loaded for every request.

• Fresh public facts → web search or research with source links.
• Internal system-of-record facts → a narrow first-party connector or reviewed MCP server.
• Repeatable process → an Agent Skill with explicit inputs, steps, checks, and output shape.
• Many local files or code → a local agent with version history and tests.
• Images, audio, video, slides, or spreadsheets → a tool that handles the native medium.
• Predictable high-volume transformation → conventional code or automation may be safer than an open-ended agent.

A practical routing matrix

The following combinations are starting points, not permanent product rankings. Product names, limits, and model tiers change quickly. The durable question is whether the selected system can reason at the required level, reach the minimum evidence, produce the right artifact, and stay inside an acceptable action boundary.

Treat permissions as part of model selection

Mollick’s blunt observation is that “permissions matter a lot.” The same model can be a helpful researcher when limited to read access and a costly operator when it can send messages or change records without review. Capability and authority should be selected together.

Begin with read access. If the workflow proves useful, allow the agent to draft the intended change. Only then consider a narrow write operation with argument validation, visible confirmation, logging, rate or volume limits, and a recovery path. Keep high-impact tools separate from broad browsing when possible because untrusted web pages, emails, and documents can contain prompt injection attempts.

Use approval gates for anything that “sends, spends, or deletes,” and extend the rule to publishing, purchasing, permission changes, customer communications, bulk updates, and exports of sensitive data. A request to be careful is not a control. The control is the absence of authority until a person reviews the exact action.

Prefer the smallest complete stack

The goal is not the fewest tools at any price. It is the smallest set that can complete the task with evidence and control. Start with one model, one work surface, one authoritative evidence connection, and one output destination. Add a skill if the process needs consistency. Add another connector only when a named handoff remains.

Tool sprawl consumes context, increases cost, complicates authentication, and makes it harder for the model to select the correct operation. It also enlarges the security and maintenance surface. Record an owner, purpose, connected identity, enabled actions, data class, last test, and removal procedure for every connection.

More expensive plans often buy more agent time and usage rather than a different level of intelligence. Estimate the hours of AI work the task needs, but measure the result as a reviewed deliverable. A long agent run that produces unsupported conclusions is not productivity.

Use this five-minute routing check

Before starting a meaningful task, answer the following questions. If you cannot answer them, the job is not ready for autonomous execution. It may still be ready for a read-only investigation that helps define the work.

• Is the desired result an answer, an artifact, a draft, or an external action?
• What is the cost of a plausible mistake, and who is accountable for the final decision?
• Which facts must be current, authoritative, or cited?
• Where do those facts and the output live?
• Does the model need stronger reasoning, a specialist modality, or simply better context?
• What is the minimum tool set and narrowest identity that can complete the path?
• Which actions require approval, and how will success or partial failure be verified?
• Would deterministic code, a template, or an ordinary automation solve the task more reliably?

Evaluate the system on one real task

The fastest useful comparison is not a generic prompt contest. Give two candidate systems the same real task, source access, constraints, and output standard. Record setup time, unsupported steps, factual corrections, review time, permission friction, and whether the final artifact can be used. Run at least one messy case, not only a polished demo.

Inspect the work itself. Open the cited sources, review changed files, reconcile totals, check the final canvas, and verify the external system after an action. Agent completion text is a status claim, not evidence that a write, upload, deployment, or send actually succeeded.

Then ask for a revision rather than treating the first answer as a verdict. Working with an agent is closer to managing a capable colleague than operating a search box: set the outcome, provide constraints, review the artifact, correct direction, and keep responsibility for the result.

The bottom line

Choose the task before the model. Choose the evidence before the connector. Choose the permission boundary before the automation. For low-stakes work, a convenient default model may be all you need. For difficult or consequential work, use stronger reasoning and a work surface built for sustained execution. Add only the tools that close the gap between required evidence and the finished result.

The best AI system is not the one with the longest feature list. It is the one that produces a useful, inspectable outcome with the least unnecessary access—and makes it obvious where human judgment still belongs.