Build, Buy, or Leave It Manual
Outcome: every step in your workflow assigned to build, buy or manual, with the arithmetic behind each call written down.
- Surface
- App and MCP server
- Level
- Beginner
- Uses
- None — this is a decision lesson
- Credits
- 0
- Prerequisite
- Lesson 08's cost model
The default is wrong in both directions
Two failure modes, equally common.
- Automating everything. A step that runs eleven times a year gets a week of build time and a maintenance burden forever. The build never pays back.
- Automating nothing. A rep spends fifteen minutes per account on research that an action does in four seconds, and it is invisible because it is spread across a team.
The rule below separates them without much ceremony.
The decision rule
Compute annual manual cost
minutes per run × runs per year × loaded hourly rate ÷ 60.
A step taking 20 minutes, run weekly, at $60/hour loaded: 20 × 52 × 60 ÷ 60 = $1,040 a year.
Estimate build cost, then double it
Your first estimate excludes debugging, edge cases, the week it breaks, and the handover when you leave. Doubling is not pessimism; it is the observed multiplier.
Compare over two years, not one
Automation is an asset with maintenance attached. If it does not pay back inside two years, it is a hobby.
Check frequency independently
Under about 12 runs a year, automate only if the step is also error-prone or blocking. Frequency, not effort, is what makes automation compound.
When to buy instead
Buy when the work is a solved commodity and the vendor’s marginal cost is below your build-and-maintain cost. Three tests:
| Test | Buy if |
|---|---|
| Is it a data problem or a logic problem? | Data. Nobody should be building an email-finding waterfall in 2026 — the coverage is the product, and you cannot replicate it. |
| Does it need to keep working while you do other things? | Yes. Maintenance is the real cost of building, and it never appears in the original estimate. |
| Is it your differentiator? | No. If the step is your edge, build it. If it is table stakes, buy it. |
The corollary is the useful part: the glue is where you build. Providers sell you lookups. Nobody sells you your scoring model, your gate threshold, your suppression logic or your routing rules — those encode how your business sells, and they are exactly what a GTM engineer owns.
What should stay manual
Four categories, and they do not shrink as your tooling improves.
Judgement on small numbers
Choosing which 10 of 80 qualified accounts get a founder’s personal outreach. The list should be automated; the pick should not.
Anything irreversible with a person on the other end
Sending to a customer, contacting an executive at a strategic account, anything where a wrong send costs a relationship. Automate the preparation, keep a human on the send.
The first three runs of anything new
Run it by hand until you know its failure modes. Automating an unproven process just makes the mistakes faster and harder to see.
Work that is really thinking
Positioning, message strategy, deciding what “qualified” means. AI can draft; it cannot decide what your company should say.
A useful middle setting: automate to a draft, keep the human on approval. A queue of pre-written, evidence-backed messages that a rep approves in bulk gets most of the leverage with none of the reputational risk. It is also the honest answer for most AI steps.
Worked example
The developer-tool build, decided:
| Step | Call | Reasoning |
|---|---|---|
| Source companies | Buy | Firmographic data is a commodity |
| Tech stack detection | Buy | Impossible to replicate at any reasonable cost |
| Job listing check | Buy | Same, and it is the timing signal |
| Fit and timing score | Build | Encodes how we sell; no vendor has it |
| Gate threshold | Build | A capacity decision, revisited monthly |
| Email waterfall | Buy | Coverage is the product |
| Opening line | Build the prompt, buy the model | The prompt is the asset |
| Choosing top-10 for founder outreach | Manual | Judgement on small numbers |
| Sequencer handoff | Buy | Sequencers exist |
| Reply triage | Manual for now | Under three runs old — see rule three |
Do this now
- List every step in your workflow, including the ones a person currently does in a spreadsheet.
- For each, compute annual manual cost with the formula above.
- Mark each buy, build or manual using the tests.
- For anything marked build, write the doubled build estimate next to the two-year manual cost. If the build is larger, change the mark.
- Circle every step marked manual that runs more than weekly. Those are your next quarter’s automation backlog, in order of annual cost.
Check your work
- Every step has a call and one line of reasoning
- Build decisions have arithmetic next to them, not a preference
- At least one step is deliberately left manual
- Nothing brand new is scheduled to run unattended
Where this breaks
The failure that hurts most is automating a process nobody has validated. The workflow runs perfectly and produces the wrong output at scale for a month, because the definition of “qualified” was never tested against a real reply. Run anything new by hand three times, look at the output row by row, and only then schedule it. The three runs cost you an hour; skipping them can cost a sending domain.
Further automation
Revisit these calls quarterly. Steps move categories as volume grows — a manual step at 20 accounts a month is an obvious build at 200, and a build you maintain alone becomes a buy the moment a vendor covers it. The list you just wrote is a living document, not a one-time exercise.
Next lesson
10 — Design your own GTM system, where the ten lessons become one page you take into every other course.
Reference for this lesson: Actions, AI agents, Integrations, Credits.