Stop asking "which AI is the best." Ask which option, if any, meets this task's acceptance criteria, data boundary, consequence level, integration needs, and review budget. Test the answer; do not assume it.
Smaller task scopes are easier to instruct, evaluate, and govern than an assistant responsible for "everything." Specialization is a testable design choice, not a guarantee of quality.
Start with one bounded task family. Compare candidate approaches on the same representative and adversarial test set, including review time and failure severity. Keep the simplest option that meets the documented criteria.
Build one bounded assistant at a time so ownership, permissions, failures, and regressions remain observable. Unlike a human hire, an AI assistant cannot own responsibility; the named human operator does.
These eight archetypes are a convenience, not an exhaustive taxonomy. Use them to form candidates for your top SOPs, then validate task fit with evidence.
Reads, summarises, cross-references. Pulls competitor info, reads long PDFs, writes briefs from raw inputs. Strong at synthesis.
Tool fit: Claude (long context), Perplexity (web grounding), ChatGPT with web tools.
Captions, ad copy, emails, blog posts, scripts. Anything where voice and tone matter as much as content.
Tool fit: Claude Projects with brand voice file, Custom GPTs, ChatGPT Projects.
Image generation, mockups, moodboards, simple layouts. Less "art director," more "production designer."
Tool fit: Midjourney, ChatGPT Images, Krea, Adobe Firefly. Sora for video.
Reads numbers, finds patterns, writes diagnoses. Great for performance reports, cohort analysis, A/B test reads.
Tool fit: Claude Code (with file access), ChatGPT with Advanced Data Analysis, Hex.
Uses authorized tools to propose or perform bounded actions. Start read-only; require least privilege, previews, approval, audit, idempotency, and rollback before any write.
Candidate tools to test: governed MCP integrations, automation platforms with approval steps, or a purpose-built service.
Writes / refactors / debugs code. Builds tools, scripts, integrations. Reads your repo and explains it.
Candidate tools to test: Claude Code, Cursor, GitHub Copilot, Codex, or another tool approved for the repository and data class.
Drafts replies, summarises threads, writes meeting notes, manages inbox triage. Knows your tone with people.
Tool fit: Claude Projects, ChatGPT Projects, Superhuman AI, Granola.
Reads SOPs, dispatches work to other assistants, tracks status, summarises what got done. The orchestrator.
Tool fit: Claude Code (subagents pattern), n8n, Make with branching, custom orchestration.
Open your SOPs file. For each SOP, write at the top which archetype it belongs to. Examples:
If an SOP truly needs two archetypes, that's a sign you should split it — or chain two assistants together (Phase 7 covers this). Don't try to make one assistant do two things at once.
You don't need 8 different tools. You need 1–3 platforms that cover your archetypes. Here's the honest comparison.
A terminal-based agent that can read permitted files, run approved commands, and use configured tools. Evaluate it in a sandbox with repository-specific permissions.
Candidate for technical workflows ★ Test before granting accessCandidate project workspace: verify current context, instruction, file, access, retention, regional, export, and deletion behavior in official documentation. Keep secrets and sensitive data out; select only after the common eval suite.
Best for content + researchSimilar to Claude Projects with a different vibe. Custom GPTs are great for sharing assistants with a small team. ChatGPT's image generation is built in.
Best if your team already uses GPTFor chained automation — when a trigger should fire an assistant, then pipe its output somewhere. The "no-code glue" between AI workers and your apps.
Best for automation chainsA code editor with AI assistance. Compare it with other approved options using the same repository tasks, security constraints, and review process.
Coders onlyVisual and video generation candidates. Evaluate rights, privacy, controllability, accessibility, cost, and output quality for your actual brief before investing.
For creative outputPlans, included usage, pricing, data terms, and capabilities change. Verify the current official product and security pages on the day you choose a tool; do not use this page as a price quote.
No stack covers every archetype or risk. Add a tool only when a documented task and evaluation show that it improves the workflow enough to justify its cost and exposure.
Don't build all assistants at once. Pick one frequent, low-consequence, testable task with permitted data and a clear baseline. High annoyance alone is not a safety or value criterion.
Screen your top 3 SOPs:
Choose the task that clears every screen. If none does, improve the SOP, data boundary, or evaluation plan before building.
"I need the most powerful AI for everything." No. Define representative, boundary, malformed and adversarial cases for the task; compare currently eligible models on quality, critical failures, abstention, latency, cost and reviewer effort. Record model/version, date and expiry, because a fixed family ranking or “just as good” claim goes stale.
Tool tourism. Evaluate a new tool only when a documented requirement or failure justifies it. Set a review trigger from model, policy, data or task change—not an arbitrary number of days.
Picking a tool because of marketing instead of the SOP. The right tool is the one that fits the archetype. Beautiful product pages and viral demos are not selection criteria. Your SOP is.
You now have a draft task-to-workspace decision record. Phase 5 builds one controlled, tenant-isolated prototype with synthetic inputs; no model is hired or given accountability.