Every week someone asks Apex Automation Team for "an AI agent that finds leads and books meetings". Parts of that are ready today; parts are not. Here is the honest split from the agents we have shipped.
What works well now
- Research and enrichment: given a company or profile, an agent can browse the site, LinkedIn-style public pages and news, then fill a structured record (what they sell, size, tech stack, recent hiring) with source links.
- Qualification: scoring leads against written criteria with a short justification — consistent and auditable.
- Personalised first drafts: a two-line opener referencing something real about the prospect, generated from the research record.
- Classification and routing: tagging inbound replies (interested / not now / wrong person) and routing them.
What still needs guardrails
- Hallucinated facts: agents must cite the page every claim came from, or the claim is dropped.
- Sending autonomously: we keep a human approval gate before anything goes out, at least until the templates have proven themselves.
- Cost drift: per-lead token costs are capped and logged; retries are bounded.
- Compliance: data sources, consent and opt-outs are handled in code, not left to the model.
Architecture we use
Scraper/API layer for reliable data → LLM steps for reading and writing → rules for decisions → CRM/Sheets/Slack for humans. The model is one component, not the whole system. Prototypes start from $200; production agents $500–$2,000+. See custom AI agents or scope one with us.
Quick checklist
- Write the qualification criteria in plain language first
- Require a source link for every fact the agent records
- Keep a human approval gate before sends
- Cap cost per lead and log every run
- Measure reply rate against your current process
Frequently asked
Which models do you use?
OpenAI, Anthropic, Gemini or open models depending on cost, privacy and quality needs; the architecture is model-agnostic.
Can the agent use my CRM data?
Yes — HubSpot, ClickUp, Google Sheets and custom databases are common inputs and outputs.