← AI Lab learning hub

Learn to choose a useful setup

Choose tools that fit the work.

Start with your outcome, the systems you already use and the responsibilities you can support. Build a small comparison with evidence, costs and open questions you can take into a pilot.

Start with the work. Choose the tools second.

An assistant, an agent, an automation and a hosting platform solve different parts of a workflow. Compare the smallest setup that can produce your useful result. The tool you already use belongs in the comparison.

  1. Name the output and the person who will use it.
  2. Write the data, access, operating and cost constraints.
  3. Compare up to three options, then verify the exact plan and configuration.
  4. Save a provisional decision and test it in a bounded pilot.

Map the workflow first

Business size changes the context, not the standard of evidence.

Illustrative situation: A four-person shop wants a draft promotion using its approved product descriptions. Someone will check every claim before publishing.

How to reason about it: Begin with an approved work assistant and a draft-only workflow. Check account terms, sources and the publishing owner. A custom app or several agents may add work without improving this outcome.

These examples do not change your plan or claim measured business results.

1. State what the setup must satisfy.

Choose the checks that are non-negotiable, then make them specific. “Sensitive data” needs a concrete rule about which records, processing, retention and access are permitted.

Changing these requirements resets your assessments. Previous evidence notes remain available to review.

2. Build a small comparison.

These are teaching profiles, not a complete directory or a ranking. Features can cross categories. A profile’s sources describe a product boundary; they do not approve your data or verify your account.

12 profiles shown. 0 / 3 selected. Filtering keeps your selected options.

Work AI suites

Microsoft 365 Copilot

Organizational tenant; verify the exact license, model and enabled features

Consider an existing Microsoft environment for drafting, summarizing and work grounded in organizational content.

Fit, limits and what to verify

Consider it when

  • A team already works in Microsoft 365
  • A person will review the resulting document or decision brief

Execution and access

Cloud service. Existing identity and content permissions shape what the user can reach.

Microsoft documents inherited permissions, labels, retention and audit controls, with availability varying by subscription. Check agents and web-search boundaries separately.

Constraints

  • An existing permission mistake does not become safe because AI respects that permission.
  • Confirm the terms and controls for the actual agent and model, not just the Copilot name.

Cost components

  • Eligible base and AI licenses
  • Agent or related-service usage
  • Information cleanup, review and administration

Operating owner

Business workflow owner plus the Microsoft tenant administrator.

A useful pilot

Draft one brief from an approved document set. Test with an ordinary member and verify sources, access and the review handoff.

Primary documentation

Profile checked . Verify current eligibility and terms for your actual configuration.

Work AI suites

Google Workspace with Gemini

Eligible work account and Workspace edition; verify feature and language availability

Explore assistance around the documents, messages and meetings a Google Workspace team already uses.

Fit, limits and what to verify

Consider it when

  • Existing work in Gmail, Drive, Docs or Sheets
  • Drafts or summaries with a human source check

Execution and access

Cloud services under the applicable Workspace account and feature terms.

Google distinguishes work/school Workspace eligibility from personal Google AI plans and describes organizational protections and admin controls. Verify the account actually used.

Constraints

  • Do not transfer a personal-account assumption to a managed work account, or the reverse.
  • Check the specific feature, administrator setting and content permissions before connecting sensitive material.

Cost components

  • Workspace edition and optional features
  • Training and document organization
  • Review and administration time

Operating owner

Workflow owner and Workspace administrator.

A useful pilot

Use one approved source to draft a useful document. Check factual fidelity, sharing settings and whether a colleague can repeat the process.

Primary documentation

Profile checked . Verify current eligibility and terms for your actual configuration.

Work AI suites

Managed ChatGPT workspace

Business, Enterprise or Edu workspace; verify the selected feature and administrative controls

A shared organizational home for conversational work, research and reviewable artifacts.

Fit, limits and what to verify

Consider it when

  • Cross-functional drafting and research
  • A clear brief and approved sources that a reviewer can inspect

Execution and access

Cloud processing. A browser or desktop interface does not make the model local.

OpenAI documents no model training on Business, Enterprise and Edu workspace data by default. Retention, connected-app access and audit coverage still depend on the data category and configuration.

Constraints

  • No training does not mean no retention or no third-party processing.
  • Check the actual workspace and enabled capability before providing company data.

Cost components

  • Workspace and applicable usage charges
  • Connected services
  • Review, enablement and administration

Operating owner

Workspace owner and the person accountable for the business artifact.

A useful pilot

Create a sourced campaign brief from approved inputs. Record corrections and total effort, including human review.

Primary documentation

Profile checked . Verify current eligibility and terms for your actual configuration.

Agents and coding

ChatGPT Work Cloud

Cloud task with separately authorized apps or local computer access

A tool-using task workflow for producing an artifact across approved sources and applications.

Fit, limits and what to verify

Consider it when

  • Several dependent steps with a clear finished artifact
  • A bounded assignment whose results and actions can be checked

Execution and access

Cloud orchestration; optional approved tools on a connected computer. Local access does not mean local-only processing.

Cloud tasks do not automatically inherit device files or browser sessions. Workspace, app, browser and network permissions are distinct; review the controls that apply to this task.

Constraints

  • Local computer access with Work Cloud does not provide strict zero data retention.
  • A connected account's permissions and the agent's task instructions are separate boundaries.

Cost components

  • Workspace and task usage
  • Connected-service usage
  • Supervision, retries and coordination

Operating owner

Task owner, workspace administrator and owners of any connected systems.

A useful pilot

Prepare a draft performance brief from approved exports. Name one writer per output, disallow live campaign changes and verify the final numbers.

Primary documentation

Profile checked . Verify current eligibility and terms for your actual configuration.

Agents and coding

GitHub Copilot cloud agent

Repository-enabled cloud agent with branch rules and human review

Delegate a bounded repository change through a reviewable development workflow.

Fit, limits and what to verify

Consider it when

  • Code already lives in a maintained GitHub repository
  • Someone can review the diff and run meaningful checks

Execution and access

Hosted agent environment. Repository and network controls need explicit review.

GitHub documents restricted branch writes, required human review and guarded workflow execution. These are delivery controls, not proof that generated code is correct.

Constraints

  • Review code, dependencies, permissions and the combined result before release.
  • Check agent and integration settings; a firewall is not a complete defense against every data-exposure path.

Cost components

  • Copilot plan and usage
  • CI and connected-tool usage
  • Review, testing and maintenance

Operating owner

Repository maintainer with a release and recovery responsibility.

A useful pilot

Assign one small bug with reproduction steps. Inspect the diff, rerun the failure and test the actual preview before merging.

Primary documentation

Profile checked . Verify current eligibility and terms for your actual configuration.

Automation and integration

Microsoft Power Automate

Specific flow trigger, owner, connectors and licensing context

Build a repeatable handoff between systems when its inputs, actions and exception path are understood.

Fit, limits and what to verify

Consider it when

  • An existing Microsoft/Power Platform environment
  • Repeatable work with a durable owner and known source systems

Execution and access

Cloud-flow profile; desktop flows have a separate execution and licensing context to verify.

License context varies by flow/trigger. Advanced Connector Policies offer allowlists, but custom and HTTP connectors require separate consideration under classic policies.

Constraints

  • A Microsoft 365 subscription alone does not establish that a proposed flow is licensed.
  • Check the exact connector/action, service limits, failure notification and safe retries.

Cost components

  • Licenses and premium connectors
  • Requests, capacity and related services
  • Support and exception handling

Operating owner

A flow owner with a backup owner and an administrator for connection policy.

A useful pilot

Run a draft-only process manually, then force one failure and one duplicate input. Check notification, recovery and whether a second run repeats an external action.

Primary documentation

Profile checked . Verify current eligibility and terms for your actual configuration.

Automation and integration

n8n, self-hosted

Chosen edition, hosting environment, model endpoints and connector configuration

An integration workflow operated in infrastructure your team is responsible for maintaining.

Fit, limits and what to verify

Consider it when

  • Someone can own updates, monitoring and recovery
  • The team needs to control where the workflow runner operates

Execution and access

The runner is self-hosted. External models, connectors, telemetry and hosting choices must still be reviewed; local-only processing is unverified.

Operating the instance includes updates, monitoring and backup/recovery. Verify which controls are included in the chosen edition and how credentials are managed.

Constraints

  • Self-hosting is not an automatic privacy, compliance or cost guarantee.
  • Do not schedule a workflow without an error path, a responsible operator and a tested restore.

Cost components

  • Edition/license where applicable
  • Hosting, model and connector usage
  • Maintenance, backup, monitoring and operator time

Operating owner

A technical operator plus a business workflow owner; neither responsibility disappears with self-hosting.

A useful pilot

Run with sample records and inspect execution logs. Test failed credentials, a safe retry and restoring the workflow/configuration in a separate environment.

Primary documentation

Profile checked . Verify current eligibility and terms for your actual configuration.

Analytics and data

Microsoft Power BI

Service sharing, semantic model, workspace roles and recipient licensing

A governed reporting layer for shared definitions and repeatable business questions.

Fit, limits and what to verify

Consider it when

  • People need the same metric definitions
  • A data owner can reconcile and maintain the underlying model

Execution and access

This profile covers sharing through the Power BI service, not an isolated desktop-only file.

Row-level security applies to Viewer access; Admin, Member and Contributor roles have different access. Sharing also involves semantic-model access and licensing/capacity conditions.

Constraints

  • Hiding a page or field is not data authorization.
  • Test as the intended consumer and validate the model before asking AI to explain a dashboard.

Cost components

  • Author and recipient licenses or capacity
  • Refresh and source infrastructure
  • Model maintenance and data-quality review

Operating owner

Metric/model owner and workspace administrator.

A useful pilot

Reconcile one report to a source export, then test two users with different intended access. Confirm the recipient can use it under the actual license setup.

Primary documentation

Profile checked . Verify current eligibility and terms for your actual configuration.

Analytics and data

Google BigQuery

Project, dataset permissions, location and on-demand or capacity billing

An analytical data/query layer when approved exports and small spreadsheets no longer fit the work.

Fit, limits and what to verify

Consider it when

  • Repeatable analysis across larger datasets
  • An analyst or engineer can own queries, definitions and cost controls

Execution and access

Cloud warehouse. Check dataset location, access paths and every service that consumes its results.

Cost controls depend on the billing model. For on-demand queries, maximum bytes billed can prevent an oversized query from running; LIMIT alone does not cap bytes read from a non-clustered table.

Constraints

  • A warehouse does not reconcile inconsistent business definitions automatically.
  • A row limit on the result is not a reliable query-cost boundary.

Cost components

  • Storage and compute model
  • Refresh/query frequency and data movement
  • Pipeline and query maintenance

Operating owner

Data project owner and an accountable query/metric maintainer.

A useful pilot

Use a bounded sample and a documented query. Reconcile counts and totals, inspect estimated bytes and test the intended billing limit before scheduling refreshes.

Primary documentation

Profile checked . Verify current eligibility and terms for your actual configuration.

App and data hosting

Vercel

Specific project, environments, team plan and resource usage

An application hosting option when a maintained web project is the right way to deliver the outcome.

Fit, limits and what to verify

Consider it when

  • A web app has a repository and repeatable build
  • Someone owns deployment, configuration and production recovery

Execution and access

Hosted deployments. Review the project's runtime, integrations and data paths.

Spend Management's amount alone does not stop usage. Optional production pausing can be delayed, takes production offline and does not stop every charge.

Constraints

  • Do not treat an alert as a universal hard spending cap.
  • Validate deployment access and rollback separately from database recovery.

Cost components

  • Plan/seats and metered resources
  • Integrations and stored resources
  • Monitoring, support and maintenance

Operating owner

Application owner with access to deployment logs, billing and recovery.

A useful pilot

Deploy a small approved page, check its actual URL and rehearse restoring the previous build. Read what the chosen budget action pauses and what continues billing.

Primary documentation

Profile checked . Verify current eligibility and terms for your actual configuration.

App and data hosting

Railway

Project/services, environment access, selected plan and persistent resources

An application/service hosting option for a team prepared to operate its code and resources.

Fit, limits and what to verify

Consider it when

  • A service or application has a clear runtime and owner
  • The team can monitor failures and manage configuration

Execution and access

Cloud services. Service location, exposed endpoints and connected data stores need review.

Railway documents environment RBAC as an Enterprise feature. Verify the plan and role behavior when production access must be separated from development.

Constraints

  • Basic team membership does not establish the production isolation you require.
  • Check persistence, backup/restore, deployment health and cost controls for the actual service.

Cost components

  • Plan and running resources
  • Storage, backups and network use
  • Incident handling and technical ownership

Operating owner

Service maintainer and project administrator, with a documented fallback owner.

A useful pilot

Run a non-critical sample service. Test a failed deployment, restricted-user access and the recovery path before trusting it with business-critical work.

Primary documentation

Profile checked . Verify current eligibility and terms for your actual configuration.

App and data hosting

Supabase

Managed project; database grants, RLS, client/server keys and storage policies

An application-data layer when a custom app needs shared data and an explicit access model.

Fit, limits and what to verify

Consider it when

  • A real need for shared saved data
  • A developer can maintain schema, access policies and recovery

Execution and access

This profile covers a managed project. Self-hosted Supabase is a different operating decision to evaluate separately.

Row Level Security constrains row access. Login and table permissions are not enough to demonstrate ownership rules; privileged paths can bypass RLS.

Constraints

  • Keep privileged keys out of public clients.
  • An administrator's successful query does not prove an ordinary user's permissions work.

Cost components

  • Project compute, storage and usage
  • Backup/recovery features and related services
  • Schema, policy and application maintenance

Operating owner

Data/application owner who can test policies and restore the relevant data stores.

A useful pilot

Use two ordinary users and sample records. Try allowed and cross-user reads/writes through the real client path, then document recovery coverage.

Primary documentation

Profile checked . Verify current eligibility and terms for your actual configuration.

Include a tool your organization already approves

Approval for one use does not automatically cover this workflow. Record the actual tool and configuration, then verify your requirements below.

3. Check evidence, not marketing labels.

All checks begin unverified. Record what you checked for the exact configuration: a document, an administrator’s confirmation or a test, with its date and limits. This page does not connect to vendors or independently validate your evidence.

Add an option above to see its operating requirements and record your checks.

4. Save the decision and what remains open.

A choice with missing evidence is provisional. A known conflict with a required check must be resolved before proposing that option.

Loading your plan…

Open My Plan

Loading your plan from this browser. This workspace runs in your browser and makes no AI or vendor-account calls. Use My Plan’s export to keep a portable copy.

Reference review: . Requirements, sources and user evidence remain separate. No universal ranking, exact price quote or compliance certification is produced.