You will lock ChatGPT so your close data cannot train a model, then write a one-page reconcile-first checklist that stops any AI number from shipping unchecked.
In finance the fastest way to lose trust is one AI figure that turns out wrong, or one confidential number that leaks into a training set. This lesson sets both guardrails before you paste a single actual into a model: your ChatGPT work stops feeding anyone's training set, and every AI number in the rest of this path, whatever tool produced it, must tie to a total you can defend to your controller.
This is the free first module of urfired.ai's Finance & Analytics path. No signup, no card. The steps note what ChatGPT access you need.
Open Settings and check your plan. On ChatGPT Business (renamed from Team on 29 Aug 2025) and Enterprise, your workspace conversations are excluded from model training by default, with no toggle to flip.
On Free, Plus, or Pro, training is on by default and you have to turn it off yourself. Know which side you are on before you paste anything from the close.
On Free, Plus, or Pro go to Settings, then Data Controls, and switch off Improve the model for everyone. New conversations then stop being used for training, and the setting syncs across your signed-in devices.
If you are on Business or Enterprise this is already handled at the workspace level, so just confirm the plan and move on.
Create the one-page doc that will become your reconcile-first checklist and write this line at the top: the opt-out is not retroactive and prompts still transmit to OpenAI servers. Anything already used in a training run cannot be pulled back out, and your prompts still travel to OpenAI to be processed and are retained under your plan's policy.
That one sentence is why the paste rule two steps down exists, and it stays on the page so nobody on your team mistakes the toggle for a vault.
Start a new chat and click the Temporary Chat toggle at the top of the chat before you type anything. Temporary chats are not used to train models and do not show up in your history, and OpenAI retains them for up to 30 days.
There is no settings default for this: the toggle resets, so you must re-select it at the start of every new chat. Make it your working mode for actuals so a stray conversation never lingers in your history.
Decide the line before you are mid-close and tempted. Name what you will never paste even with training off, for example employee-level comp, customer names tied to revenue, or unreleased results before they are public.
Keep those in Excel and paste only the aggregated or anonymized figures you actually need reasoning over.
Start a Temporary Chat and give it your real close steps in a sentence, then ask it to draft a one-page checklist that forces every AI-produced number to be reconciled before it ships. Feed it your actual review points, do not accept a generic list.
The rule at the top: no number leaves ChatGPT until it ties to a total you already trust.
Edit the draft so it names the specific tie-outs from your own books, not placeholders. For example: cash on the balance sheet ties to the bank rec, net income ties from the P&L to the change in retained earnings, revenue ties to the GL control total.
Each AI figure must map to one of these before it can go in a schedule or deck.
Build in one step that makes you ask ChatGPT to expose how it got each figure: the formula, the source cells, or the method. You reconcile that against your own total, not against the model's confidence.
Save the finished page and pin it where you start the close, so it is the first thing you open next month.
Take one real number from your most recent close that clears the paste rule you just wrote, aggregated or anonymized, nothing market-sensitive, for example a variance figure or a quick total you would normally eyeball. Open Settings and either confirm your Business or Enterprise plan or switch off Improve the model for everyone under Data Controls, then start a new chat and click the Temporary Chat toggle at the top.
Ask ChatGPT to draft a one-page reconcile-first checklist from your actual close steps, edit it to name three real control totals from your own books, and run that single number through the checklist end to end, including asking the model to show how it produced it and reconciling it to your known total.
A one-page reconcile-first checklist anchored to your own control totals, plus a confirmed privacy setup (Business or Enterprise plan verified, or training toggled off on a personal plan) so no AI number ships unchecked and no close data trains a model.
Module 02: Audit a three-statement model with Claude for Excel (Claude for Excel, 22 min)
Free gets you this starter lesson today. Pro unlocks the full 12-module path when checkout opens, no card and no charge until then.
New here? The free 2-minute assessment reads your role and gaps, then matches you to the right path. Get your score