The AI Toolkit in Sync GTM
Outcome: each AI step in your own workflow assigned to one of five tool types, with the cheaper non-AI alternative named next to it.
- Surface
- App and MCP server
- Level
- Intermediate
- Uses
- AI agent columns · MCP server
- Credits
- 0 — this lesson is a map
- Prerequisite
- Lesson 01
Five tools, five jobs
| Tool | Job | Cost per row | Cheaper alternative |
|---|---|---|---|
| Research agent | Answer an open question by reading sources | ~1–2 credits | A scrape, if you know the URL |
| Classifier | Put a row into one of a fixed set of buckets | ~0.5 credit | A keyword rule, for the predictable 80% |
| Scorer | Produce a number from evidence | ~0.5 credit, often free | A formula — usually the right answer |
| Writer | Generate copy from row data | ~1 credit | A template, if the row has no distinguishing evidence |
| MCP chain | Assemble many tools in one pass | Cost of the tools it calls | A table, when the run repeats weekly |
Lesson 01 gave you the test. This lesson is the map you apply it against.
Research agent
Use when the answer requires reading and judging. “Do they sell to enterprise?” “What does their pricing model imply about their buyer?”
Do not use it for facts a database holds — headcount, funding, tech stack, emails. It is slower, costlier and less accurate than the enrichment that exists for exactly that.
Full treatment: Research & Web Scraping lessons 05–07. Reference: AI agents.
Classifier
Use to collapse free text into a value you can filter on: a job title into a function, a description into a segment, a research answer into a bucket.
Rules first
A keyword formula handles most rows for free and is instantly debuggable. “VP Engineering”, “CTO”, “Head of Platform” → engineering.
AI on the remainder
Run the model only where the rule returned unclassified, constrained to the same closed value set.
Never let the value set drift
Written down, closed, with Unknown and at most one Other.
This split is the single largest AI cost saving available in a normal table, because classification is the most-run AI column and rules cover most of it.
Scorer
Use AI here rarely. A score is arithmetic over columns you already have, and a formula does it for free, deterministically, and explains itself.
The exception: scoring something genuinely qualitative that has no column — “how closely does this company’s stated problem match what we solve?” Even then, have the model return one of five levels rather than a number out of 100. A model’s 73 is not meaningfully different from its 68, and treating it as though it were produces a gate nobody trusts.
Lesson 12 builds the scorer properly.
Writer
Use for the one sentence that must differ per row. Not for the pitch, not for the CTA, not for the subject line — those are written once by a human and reused, because they should be the same for everyone.
The writer’s constraint set is the whole craft: one sentence, banned phrases, must reference the evidence, abstain when there is none. Lesson 14 covers it, and SyncGTM 101 lesson 15 has the base version.
MCP chain
The other surface entirely. Instead of columns running down a table, tools run in sequence inside a conversation — search, filter, enrich, verify, push, in one pass.
| Use a table when | Use an MCP chain when |
|---|---|
| The run repeats on a schedule | It is ad hoc or exploratory |
| Volume is hundreds or thousands | Volume is tens |
| Several people need to see the result | You need the answer now |
| Each step needs inspecting | You trust the chain end to end |
Lesson 04 is the mental model; the MCP prompting guide has the phrasing rules.
The two surfaces are not competitors. The common pattern is to prototype a workflow as an MCP chain — fast, cheap, no setup — and move it into a table once it is worth running every week.
Do this now
- List every step in your workflow where a model runs, or where you are considering one.
- Assign each to one of the five tool types.
- Next to each, write the cheaper non-AI alternative — a filter, a formula, a database enrichment, a scrape.
- Delete any AI step whose alternative would work. Most tables have at least one.
- For anything left, note whether it belongs in a table or an MCP chain using the comparison above.
Check your work
- Every AI step has a type and a named alternative
- No AI step is doing something a database enrichment does
- Your scorer is a formula unless you can justify otherwise
- Classification runs rules first, AI second
Where this breaks
Using a research agent to fetch a fact a database already holds is the most common and most expensive misuse in this course. It costs several times more per row, it is less accurate, and it fails silently — the model returns a plausible headcount rather than saying it does not know. Before adding any research column, check the enrichment catalogue for a tool that answers it directly.
Further automation
Once each step is typed, the cost model from lesson 03 falls out of it: rows entering each type, multiplied by cost per type. That number is what tells you whether the workflow scales, and it is much easier to compute once the steps are labelled.
Next lesson
03 — Credit discipline, the ordering rules that keep an AI-heavy workflow affordable.
Reference for this lesson: AI agents, Actions, MCP tools, AI integrations.