Stop asking if AI can do this. Start asking what shape the work is.

Executive Briefing by Nate Jones — the Build, Buy, Hire, or Wait matrix. Adapted today as the business-side lens.

The Build, Buy, Hire, or Wait Matrix Prompt Kit

Four prompts by Nate Jones: Workflow Decomposer, Six-Dimension Scorer, Build-vs-Buy Pressure Test, Describability Test. Available to Executive Circle subscribers.

Over 40% of Agentic AI Projects Will Be Canceled by End of 2027

Press release, June 25, 2025. Source of the stat opening Slide 03 — cost, unclear business value, and inadequate risk controls cited as the drivers.

Giving AI the right information to do the right work.

The shift from "prompt engineering" to "context engineering" recognizes that output quality depends on what's in the model's context window — not just how cleverly you ask. That includes structured inputs (schemas, examples), constraints (output format, persona, domain rules), retrieval (relevant documents pulled in at runtime), and an explicit definition of "good." Without it, the model improvises off training-data averages. With it, it executes against your standard.

Connecting steps so AI works end-to-end.

An "agent" is shorthand for an LLM loop that can call tools, decide next steps, and run until a goal is reached. Orchestration is the design of that loop — how steps connect, when control passes from human to model and back, how errors propagate, what state persists between calls. Trade-off: linear workflows are more debuggable; more agentic workflows (the model picks next steps) are more flexible but harder to evaluate.

Knowing when the output is good enough to trust.

The thing that separates one-off prompts from production AI. An eval is a test — sometimes a regex, sometimes a model-as-judge, sometimes a human reviewer — that scores outputs against a written standard. Build evals before you optimize prompts, not after. Start with 30 examples, log every output, score them by hand, find the patterns of failure.

Connects apps, moves data between steps.

Visual workflow builders that link triggers, data sources, AI calls, and output destinations without you writing infrastructure code. Make is the most visual and flexible; Zapier is fastest for simple connections; n8n is self-hosted and code-extensible. Pick by which integrations are mature for the apps you actually use AND how much branching/transformation logic your workflow needs.

Examples: Make · Zapier · n8n · Google Apps Script.

Reads, writes, reasons, classifies.

Direct API access (Claude API, OpenAI API) is what you call from an orchestration platform — you send a prompt + context, you get a response back to use in the next step. Hosted assistants (Claude.ai, ChatGPT) are for ad-hoc use, not production workflows. For most operators, start with one API as your default and only branch out when you have a specific reason.

Examples: Claude API (Anthropic) · ChatGPT API (OpenAI) · Gemini API (Google).

Fires the automation when something happens.

Three categories: time-based (run every Monday at 8am), event-based (run when a form is submitted, a new row appears, a webhook fires), or manual (run when I push a button). Most "automation" failures are actually trigger failures — the workflow never fires, or fires too often, or fires on the wrong condition. Triggers are also where most permissions and authentication problems show up.

Examples: Google Apps Script time-based triggers · Webhooks · Cron · Native triggers (HubSpot workflows, Airtable automations).

Holds inputs, outputs, and records.

Where data lives between steps and after the workflow runs. Most automations write somewhere — a log, a record, a notification destination. Choose by who already uses what: if your team lives in Notion, store there; if your team lives in Sheets, store there. The cost of asking the team to learn a new tool is almost always higher than the benefit of the "better" tool.

Examples: Supabase (relational + auth) · Google Sheets (simplest) · Notion (human-readable records) · Airtable (structured + view-flexible).