Per-resolution SaaS (Agentforce, Intercom Fin, Zendesk Resolve) wins below ~20K resolutions/yr. Custom agents on Anthropic / OpenAI direct win above 100K. The middle band depends on workflow complexity, integrations, and your engineering bench. Fourteen questions to find the honest fit verdict.
Custom-agent build cost is derived from workflow shape (sales / support / ops), tool-integration count, judgement complexity, and compliance constraints. Monthly run cost is token spend (modelled per the AI API cost calculator) + infra ($1.5K-12K/mo) + minimal ops attention. Per-resolution SaaS comparison uses public list pricing for Agentforce ($2/conversation), Intercom Fin ($0.99/resolution), Zendesk Resolve ($1.50-3/resolution depending on tier).
Payback month assumes you migrate gradually — first 3 months parallel run, then ramp to 100% over months 4-9.
~20K resolutions/yr is where custom starts beating per-resolution SaaS on $/yr alone. Above 100K, custom is 5-10× cheaper. Below 20K, stick with the SaaS unless you have strategic reasons.
Sales/support agents: 8-12 weeks fixed-price typical. Ops agents (more workflow integration): 12-16 weeks. The memo includes a milestone breakdown.
Agentforce at $2/conversation gets expensive fast. The memo shows the inflection point and the exact prompt/eval architecture our customers use to migrate off.
Build-cost breakdown, model-mix recommendation, integration architecture sketch, ramp-down playbook for the SaaS being replaced, evaluation harness template. About 8 pages.