Choose an AI stack around the work you need done

Modular workflow pieces, a laptop, and a local computer arranged on a warmly lit workbench.

Build with AI · Sources reviewed 2026-11-19

An AI stack is the set of components that gets a job from input to outcome. It may be one assistant and a folder of documents. It may be a workflow platform, a model API, a database, and several business applications. Start by deciding which pieces your job actually needs.

This guide is a decision framework and a dated product orientation, reviewed November 19, 2026. Features and entitlements change. Use the linked official documentation to confirm the exact account, region, plan, and deployment you intend to use.

On this page

Write a five-line build brief

Before choosing a product, record the outcome, trigger, information, permitted actions, and operating owner. For example:

Outcome: a review-ready quote draft
Trigger: a new website inquiry
Information: the inquiry and approved price book
Actions: save a draft; send only an approved version
Owner: office manager; integration support: named technician

This brief suggests a fixed workflow with one interpretation step. A free-roaming agent would add decisions that the process does not require. If the job instead involves investigating an unfamiliar issue across several permitted tools, agent-selected steps may help. Anthropic’s architectural guidance distinguishes these patterns; use the distinction as a design question, not a product ranking. Building effective agents.

Choose the smallest useful starting point

Approach A good starting job Main operating question
Existing assistant Analyze a bounded document packet and prepare a draft Can the person verify the evidence and manage the sources?
Workplace AI Work with approved information where employees already collaborate Which identity, source settings, and administrative controls apply?
Visual workflow Repeat a defined sequence across applications Are all required connectors, reviews, and recovery paths supported?
Custom application Provide a specific interface or enforce unusual business rules Who maintains the code, integrations, and deployment?
Local model service Run a bounded workload on controlled infrastructure Is the complete data path local, and can the hardware meet the job?

These approaches can be combined. An existing assistant may help design a workflow; the finished workflow can then run without that conversation.

Existing assistants: start with a supervised deliverable

ChatGPT and Claude can organize document context in projects. Their project features, sharing, and limits differ. OpenAI documents project files and instructions; Anthropic currently documents both established projects and a newer experience in staged beta. Choose the behavior available in your account, and test how source changes affect the result. ChatGPT projects · Claude projects.

This route suits research preparation, document comparison, content drafts, and a first prototype. Connected apps may add searches or actions, but capability is specific to the app and account. Check the permission and action you need instead of assuming that a visible connection can perform every operation. OpenAI’s connected-app guidance.

Use the site’s ChatGPT and Claude guides for platform orientation. Keep the original evidence and final deliverable in a format you can export.

Workplace ecosystems: follow the existing identity and data

For a Microsoft organization, separate the conversational Copilot experience, Power Automate flows, and Copilot Studio authoring. Copilot Studio provides a graphical environment for agents and workflows, while its app-creation experience is marked preview in the reviewed overview. Connector, publishing, and capacity entitlements require a plan-specific check. Copilot Studio overview · Access and licensing.

A useful pilot might retrieve an approved policy and prepare a service request. Confirm the caller identity, source permissions, and destination connection independently. Start with the site’s Microsoft Copilot and Power Automate guides.

For a Google organization, evaluate Gemini in the Workspace applications where the relevant documents live. Review source settings and available integrations. A pinned file does not by itself describe every source enabled for a conversation. Google’s source controls support that check; the site’s Gemini guide provides context.

Visual workflows: compare the whole run

Zapier, Make, and n8n are candidates when a process crosses several applications. Test your exact trigger, lookup, write action, and exception path before choosing between them.

Zapier documents review steps that pause a Zap. Make documents execution-order and incomplete-run settings. n8n documents review before selected agent tool calls. These are different controls; verify how the one you choose behaves when a reviewer rejects a request or a downstream service fails. Zapier review · Make scenario settings · n8n tool review.

Check the cost of a representative completed job, including branches and retries. Make’s credits vary with the feature and AI connection type; a custom provider connection can have separate model billing. Make’s credit guide.

n8n offers self-hosted deployment, which makes the hosting team responsible for operating that installation. Its source uses specified licensing terms, including Sustainable Use and separately licensed enterprise portions; self-hostable does not mean unrestricted reuse. Hosting documentation · Repository license.

Continue with the canonical Zapier, Make, or n8n guide when a candidate fits your brief.

Developer and open-source routes: own the application contract

Custom code is useful when you need a particular interface, unusual permissions, portable business logic, or an integration the visual platform does not support well. Start with a thin application and explicit model input and output. Add an orchestration framework when state and control flow justify it.

Claude’s API documentation distinguishes tools run by the application from tools run by Anthropic. That distinction matters when deciding where credentials and authorization checks belong. Claude tool-use documentation.

LangGraph offers checkpoint and store mechanisms for stateful applications; its core repository uses the MIT license. Hosted services and model providers can carry separate terms and charges. Persistence documentation · LangGraph license.

A coding assistant can help implement this application. The resulting code still needs a maintainer, meaningful tests, dependency management, and deployment credentials appropriate to its job.

Local models: inspect the full route the data takes

Ollama provides APIs for both local and cloud models. Selecting a local server address is only one part of checking whether the intended model and connected tools stay on your infrastructure. Ollama API documentation.

For a local pilot, document the model, its license, hardware memory, expected concurrency, and every outbound dependency. Benchmark with representative documents. Include electricity, hardware, updates, backups, and support in the budget. Local execution can change the deployment boundary; it does not automatically establish better answers or a complete privacy design.

Make a short, evidence-based comparison

Give each candidate the same sample records and expected outcomes. Record whether it can read the right sources, produce the required fields, reject an unsupported case, pause at the intended decision, recover from a duplicate event, and export useful logs.

Mark each requirement as confirmed, failed, or still untested. Price the actual plan and usage pattern needed to satisfy it. Include the person who will maintain the workflow in the decision.

Choose the candidate that meets the requirements with a support burden your team can carry. Then build one complete path: an individual research notebook, a small-business quote workflow, departmental onboarding, or an enterprise knowledge service.

Share this page