Credit Discipline
Outcome: your workflow reordered cheapest-first with gates between the tiers, and a pre-run checklist you apply before anything expensive fires.
- Surface
- App and MCP server
- Level
- Intermediate
- Uses
- check_credits · formula columns
- Credits
- 0
- Prerequisite
- Lesson 02
The cost ladder
Order every step by cost per row. This ladder is the backbone of the whole course.
| Tier | Steps | Cost per row |
|---|---|---|
| Free | Filters, formulas, status columns, keyword rules | 0 |
| Cheap | Company enrichment, tech stack, job listings, signals | ~0.3–0.5 |
| Moderate | Person enrichment, verification, single page scrape | ~0.5–1 |
| Expensive | Research agents, AI classification at volume | ~1–2 |
| Most expensive | Generated copy, multi-source agentic chains | ~1–4 |
The rule: each tier gates the next. A row only reaches the expensive tier if the cheap tiers said it was worth it.
Where the gates go
Gate 1 — after free filters
Geography, size band, obvious exclusions, suppression. Removes the rows that never belonged. Costs nothing, so it runs first, always.
Gate 2 — after cheap enrichment
Firmographics, stack, hiring. This is where the real cut happens, usually 60–80% of rows.
Gate 3 — after research
The research verdict qualifies or disqualifies. Rows that came back Unknown should be penalized, not passed through.
Gate 4 — before generated copy
Only rows with usable evidence get a written line. Everything else abstains, per lesson 16.
Four gates sounds like a lot until you compute a run without them.
The arithmetic
2,000 sourced companies, three contacts each at the person stage.
Ungated:
| Step | Rows | Cost | Total |
|---|---|---|---|
| Company enrich | 2,000 | 0.5 | 1,000 |
| Research agent | 2,000 | 1.5 | 3,000 |
| Find people | 6,000 | 0.3 | 1,800 |
| Email + verify | 6,000 | 1.1 | 6,600 |
| AI copy | 6,000 | 1.0 | 6,000 |
| 18,400 |
Gated:
| Step | Rows | Cost | Total |
|---|---|---|---|
| Company enrich | 2,000 | 0.5 | 1,000 |
| Gate → 600 | — | 0 | 0 |
| Research agent | 600 | 1.5 | 900 |
| Gate → 350 | — | 0 | 0 |
| Find people | 1,050 | 0.3 | 315 |
| Email + verify | 1,050 | 1.1 | 1,155 |
| Gate → 800 with evidence | — | 0 | 0 |
| AI copy | 800 | 1.0 | 800 |
| 4,170 |
Same pipeline, 77% cheaper, and the 800 people who get a message are better qualified than any 800 of the ungated 6,000.
Batching versus per-row
A subtle one. Some AI work can process several rows in one call, some cannot.
- Per-row is required when the output must differ per row: copy, per-account research, classification of that row’s text.
- Batching helps when you are deriving a shared rule — for example, asking a model once to produce the keyword rules for your classifier, then applying them as a free formula to 5,000 rows.
The second pattern is underused. Using AI to write the rule and a formula to run the rule gives you model-quality logic at formula cost.
The pre-run checklist
Before any run above a few hundred rows:
- Every free filter runs before every paid step
- A gate sits between each cost tier
- The most expensive column runs on the smallest row count in the table
- Each new column has been tested on five rows and read
- Company-level work runs on the company table, not the person table
- Refresh cadences are set — nothing expensive re-runs weekly by default
-
check_creditsbalance covers the estimate plus 20%
The company-versus-person line matters more in AI workflows than anywhere else. A research agent on a person-grain table with three contacts per account runs — and bills — three times per company for an identical answer. Check every AI column’s input: if it is the domain, the column is on the wrong table.
Do this now
Tier every step
Free, cheap, moderate, expensive, most expensive.
Reorder
Cheapest first. Any expensive step above a cheap one moves down.
Insert the four gates
Even where they pass everything today.
Compute both numbers
Run cost ungated and gated, using the table above as the model.
Check for the person-table mistake
Any AI column keyed on domain moves to the company table.
Run the checklist
Then run.
Check your work
- Steps are in ascending cost order
- Four gates exist, with written conditions
- You can state the gated and ungated cost of one run
- No company-level AI column sits on a person table
- Balance covers the estimate with headroom
Where this breaks
Scheduled AI workflows are where budgets disappear quietly. An ungated research column on a growing table costs more every week, nobody is watching the consumption between runs, and the first signal is an empty balance mid-run — which leaves the table in a half-processed state that is tedious to untangle. Check consumption specifically on the second and fourth run of any new schedule.
Further automation
Log credits per run against qualified rows produced. That ratio — cost per contactable, qualified row — is the number that tells you whether an AI step is earning its place. GTM Engineering lesson 08 works it through in full.
Next lesson
04 — How to think with MCP, the surface where the same discipline moves from column order into the prompt.
Reference for this lesson: Credits, check_credits, Actions, AI agents.