Cost the scripted run before busy season
Lesson 3 of 5 · The practitioner's week
1 · Learn the move · Constraints + negative instructions
Having a model draft one notice translation is free, near enough. Drafting one for every notice the firm receives, for every client, through a season, is a line in the firm's books, and under the OPR's reading of Circular 230 the savings belong to the client, not to the fee [P]. So before the first full run, do the arithmetic on paper: what one run uses, times the number of runs, times the price. The price is the one number the model must never supply. Prices change, and a model will state an old one with full confidence. Copy it from the provider's pricing page, write down the date you read it, and hand it over as an input. Then ask for the arithmetic line by line, a monthly figure at your real volume, and a cap: the count or spend at which the script stops and asks. Missing input means a missing line, never a guessed one. And one constraint this door adds: the run is costed on de-identified inputs only, because a scripted run that pastes return information is a scripted disclosure, and no price makes that cheaper.
I am costing a scripted run that drafts [the task] from de-identified inputs. Inputs I will give you: tokens per run from my test [paste], the provider's price per token from its pricing page, copied today with the date [paste], the number of runs per month at my real volume [paste]. Constraints: never supply a price, a token count or a volume I did not give you; a missing input is a missing line marked [NEEDS]; show the arithmetic line by line; give the monthly figure and a cap at which the script stops and asks; say nothing about hours saved or fees. Assume every input is de-identified and say in one line what happens to the cost estimate if it is not: the run does not happen.
Vendor notes, dated
Prices and plan names move. Each note says when it was looked up, whether a person has confirmed it, and the day it comes due. Past that day, treat it as a question.
- Claude. Claude Pro is $20 a month billed monthly, or $17 a month billed annually. Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
- ChatGPT. ChatGPT Plus is $20 a month. Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
- Google Gemini. Google Gemini's paid consumer plan is Google AI Pro at $19.99 a month; Google AI Ultra starts at $99.99 a month. Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
- Anthropic. Anthropic's consumer plans (Free, Pro, Max) may use chats to train models under a setting the user controls, with five-year retention when training is on and 30 days when it is off; Claude for Work, Education, Government and the API are excluded. [P] Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
- Anthropic. Anthropic's usage policy, effective 2025-09-15, requires a qualified human reviewer and disclosure of AI use for housing, legal and financial decisions. [P] Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
- OpenAI. OpenAI's consumer plans train on chats unless the Improve the model setting is turned off, and a Temporary Chat is not used for training; Enterprise, Business, Edu and API data are not used by default. [S] Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
- Google Gemini. Google Gemini Apps keeps activity 18 months by default, keeps human-reviewed chats up to three years, and tells users not to enter confidential information. [P] Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
- Microsoft Copilot. Microsoft Copilot on a personal account may use chats to train models unless the user opts out; Entra and Microsoft 365 work accounts are excluded. [S] Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
- Claude. Claude Team, Enterprise and API inputs and outputs are not used for training by default; a BAA is available only on Enterprise and the first-party API after admin action. Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
- Claude. Claude Free, Pro and Max inputs can be used for training if the user allows model improvement, with retention up to five years if allowed and 30-day deletion otherwise (pages dated 2026-03-16 and 2026-07-01). Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
- OpenAI. OpenAI API inputs are not used for training by default; abuse-monitoring logs are kept up to 30 days; zero data retention is available by approval. Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
- OpenAI. OpenAI Business, Enterprise and Edu no-training defaults are reported by secondary sources only; the vendor trust pages returned 403 on the check date. Read the vendor page yourself. Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
- Microsoft. Microsoft Copilot and Copilot Chat with Enterprise Data Protection: prompts and responses are not used to train foundation models, the DPA applies, and Microsoft acts as processor (Learn page updated 2026-08-18). Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
- Google. Gemini in Google Workspace content is not human-reviewed or used for training outside the customer's domain; default retention is 18 months. Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
2 · Your turn. You write the prompt
Harlow & Vance wants a script that drafts the plain-language translation for every IRS notice the firm receives, from the notice type and a de-identified summary. Last season the firm received about three hundred notices. The partner wants the monthly cost at that volume before anyone writes a line of code.
Write the prompt for this scenario as given, or swap in a real one from your own work. The score works the same either way.
Remember: the AI sees only your prompt, not this page. If the situation isn't in your prompt, it doesn't exist.
Optional. These shape the output when you run your prompt below, not your score.