AI agents for GTM

Good GTM jobs for a model

LLMs are strong at reading messy text and answering a narrow question about it. Good fits:

• Answering yes/no/unknown questions about a company from its website and job posts.
• Summarising a company's offering in one sentence a rep can scan.
• Classifying replies: interested, not now, wrong person, unsubscribe, out of office.
• Drafting one personalised line from research, for a human or rule to accept.
• Turning call notes into CRM fields and next steps.
• Poor fits: exact numbers (headcount, revenue) a data provider sells, and anything final and irreversible without review.

Design a research pipeline, not a mega-prompt

Break the job into steps: filter with cheap data first (industry, size), fetch the specific pages that matter, ask one clear question per step, and return structured JSON with the answer, a confidence level and the source URL. One prompt that tries to research, score and write all at once is hard to debug and hard to improve.

Run each step on 50 hand-labelled rows before scaling. If agreement with your labels is under the bar you agreed with sales (say 85–90%), improve inputs and instructions before adding more rows.

Know your cost per row

Multiply it out before you run it. If a research step uses about 8,000 input tokens and 300 output tokens per company and runs on 10,000 companies, that's around 80 million input tokens. Depending on the model, that ranges from a few dollars to a few hundred. Web search and browsing add per-call fees on top. Use smaller models for classification, cache results so re-runs don't pay twice, and only run deep research on accounts that passed earlier filters.

Autonomy in steps

Increase what the AI may do only as measured accuracy and trust allow:

• Suggest: the model's output is shown to a rep, who decides.
• Draft: the model fills in a message or field; a human approves.
• Auto with sampling: low-risk actions run automatically and a person reviews a random sample weekly.
• Never automated: messages to existing customers, anything involving pricing commitments, deleting records.