As granular as you can defend in a procurement conversation. Most teams settle on six to eight line items: license, implementation, infra, support, training, switching cost, and (optionally) labour overhead and integration cost. More than that and the math becomes opaque, which undercuts the chatbot's main advantage: a number the buyer can audit.

From a custom table or ACF field set you maintain. Cloud deployment costs scale with seats; on-prem and hybrid follow different curves. Seed assumptions from your reference architecture docs or from public cloud calculators, then refine as customer data accumulates. Updates take effect on the next conversation, no redeploy needed.

Yes. Deployment type is one of the structured inputs. The system prompt asks for it early, and the formula picks the right infra coefficient. On-prem typically front-loads cost in year one, cloud spreads more evenly, and the breakdown surfaces that distribution visibly in the reply.

Implementation cost rarely lands on a single number. The formula can return a range (e.g. $40k to $60k) by reading a low and high coefficient per company-size bucket. The bot shows the range, then quotes the mid-point in the total. Buyers appreciate the honesty, and the mid-point holds up well in a procurement memo.

Yes. The visitor's input on current annualised tool cost gets multiplied across the chosen horizon and subtracted from the gross TCO. The reply shows both the gross figure and the net, so the buyer can choose which one to present internally. Some CFOs want gross, others want net plus rationale.

If you publish the competitor's list pricing and reasonable infra assumptions in a separate ACF or custom-table field, the bot can produce a side-by-side. Be careful: competitive claims should be auditable. The system prompt should refuse to invent competitor numbers, and only quote from the stored assumptions you control.

Three options. In-chat: the bot shows each line item in its reply. Email: the visitor confirms an address and the bot triggers a wp_mail with a one-page PDF. CRM: the lead-capture hook pushes the inputs, line items, and total into the lead record. Most teams enable all three. The PDF is the one that gets forwarded internally.

A reasoning-capable model handles multi-step arithmetic more reliably than a small chat model. GPT-4.1, Claude Sonnet, or similar are common choices. The horizon is usually three years, matching procurement convention; some teams add a five-year option for deals where customer commitment is longer. Bring your own API key, so cost per conversation stays predictable.