How to audit an AI agent managing your money

Treat an AI agent like a new hire on your finance team. Give it rules that bound what it can do, have it preview every change before it touches your data, then audit what it did. Let it handle the routine 90% and bring the hard 10% back to you. The tool keeps changing. The controls don't.

That's the approach Mike Dion takes. Mike reconciles other people's numbers for a living, yet their own money still took a couple of hours a month by hand. So Mike connected 34 real accounts to Era, let Claude work on them for 30 days, and kept a controller's grip on every step. This guide walks through the method, with the prompts Mike used, so you can run the same checks on your own accounts.

What are the three controls for an AI agent managing your money?

Bound it, preview it, audit it. Rules set the limits of what Claude does on its own. A preview shows you exactly what a rule would change before it runs. An audit tells you which rules fired, which never did, and where Claude was guessing.

The order matters. Each control catches what the one before it missed.

How do you give Claude access to your accounts?

Link your accounts in Era, then add Era Context to Claude as a connector and approve access. Era holds the context about your finances: your transactions, categories, rules and tags. Claude does the work. Nothing connects until you approve it.

1. Link your banks and cards in Era. Era brings the transactions from every connection together in one place.
2. Add Era Context to Claude. Connect Era Context to Claude opens Era Context in Claude's connector directory.
3. Sign in and approve. Sign in to Era and approve the consent screen.

For screenshots and the other ways to connect, follow how to connect Claude to your bank account.

Claude asks your permission before it acts, and it can't change anything at your bank. It works on Era's copy of your data, with your approval.

In Era, the Refine section is where you see what Claude has shaped: your categories, rules and tags. The work happens in your AI. The context lives in Era.

Can you do the same in ChatGPT?

Yes. Era Context is also an official plugin in ChatGPT and Codex. Your rules, categories, tags and memory live in Era, not inside one assistant, so ChatGPT works from the same context Claude does. Mike makes exactly this point in the video: switch to ChatGPT, ask the same question, and you should get the same answer.

1. Add the plugin. Open Era Context in ChatGPT and add it.
2. Sign in and approve. Sign in to Era and approve the consent screen, as you did for Claude.

Era works with any MCP-compatible client — Claude, ChatGPT, OpenClaw, and dozens more — so you set up your controls once and every assistant you connect follows them. For the full setup, follow how to connect ChatGPT to your bank account.

Add to ClaudeOfficialAdd to ChatGPT & CodexOfficial

How do you find out where your categories are wrong?

Ask Claude to audit the last 30 days before you ask it to fix anything. You want a count, the merchants behind it, and the places where a wrong category distorts a real budget line. Then ask for rules, and tell Claude not to create them yet.

Here's the prompt Mike used:

"Pull my last 30 days of transactions and audit the categorization. I want three things: how many transactions landed in a category I would not have picked; which merchants are driving most of that; where the money is showing up in a category that overstates or understates a real budget line. Show me the count and the merchant. Take the top three problems and write me an automation rule for each. For every rule, give me the rule name, the exact filter logic in plain English, the action, and the priority. Do not create anything yet."

In Mike's account, Claude reviewed 343 transactions and found 12% of them in a category Mike wouldn't have picked. It proposed three rules:

1. Treat every variant of a card autopay as a transfer, not spending.
2. Treat HSA reimbursements into chequing as transfers.
3. Mark marketplace orders for review, because one order can span several categories.

"Do not create anything yet" is the control. You get the rule on paper first, in words you can check, before it touches a single transaction.

How do you preview a rule before it changes anything?

Ask Claude to test the rule against your history and show you the matches. A preview saves nothing and changes nothing. If the matches look right, tell Claude to create the rule and apply it to your existing transactions as well as new ones.

Mike's preview prompt:

"Preview rule one against my last 90 days. I want the match count and 10 sample transactions before it touches anything."

In Mike's account, the first rule matched two transactions. They looked right, so Mike asked:

"Create these rules and apply them retroactively."

The rule went live and changed three transactions, from June, July and August. From then on, new transactions that match arrive already sorted.

Two details help here. Each rule has a priority, and when two rules could set the same category, the higher-priority rule wins. And if you delete a rule later, you can choose to undo every change it made, so a rule is never a one-way door. Rules need edit access, which starts on the Organize tier. See pricing for what each tier includes.

For more on writing rules in plain English, see how to create AI rules for your money.

Which built-in detectors should you turn on?

Era's rules library ships ready-made pattern detectors alongside the rules you write, and most of them start switched off. Ask Claude to list both groups, then to explain what each detector that's off would catch. Turn on the ones that match a question you actually care about.

Start with the inventory:

"List every automation rule I have, split into two groups: the pattern detectors Era ships and the transaction rules I wrote. For the detectors, show me which are on and which are off."

In Mike's account, that came to 75 rules. 21 came from Era's library, with 5 on and 16 off. The other 54 were rules Claude had written on Mike's behalf, and some of them had never matched a transaction.

Then ask for the trade-off, one line per detector:

"For each detector that is currently off, tell me in one line what it would flag and roughly how often it would have fired."

The library covers patterns like these:

• Duplicate charges — the same amount charged twice.
• Subscription price increases — a recurring charge that goes up.
• Buy now, pay later and cash advances — charges from those lenders.
• Weekend spending spikes and weekend impulse spending.
• Charitable donations and medical expenses — tagged for tax time.
• Dining out frequency and coffee shop habits.

Mike turned on dining out frequency, coffee shop habit detection and medical expense tagging. Pick for your own life, not for coverage. A detector you'll never act on is noise.

How do you audit what your rules actually did?

Ask Claude for a rule hit report, then ask where it was least confident. The report shows which rules pull their weight, which overlap, and which never fire. The confidence question shows you where Claude filled a gap with a guess.

Mike's audit prompt:

"Give me a rule hit report for the last 30 days. Every rule that fired, how many transactions it touched, ranked by hit count. Flag anything that fired zero times. And of everything you categorized, what are you least confident about, and tell me what you were guessing at."

Here's what Mike's audit turned up, and what each finding tells you to do:

1. A rule that duplicated another. Merge them or delete one.
2. A rule whose results Mike corrected by hand 16 times. The rule is close but wrong. Refine its filter.
3. Rules that overlapped. Check the priorities so the right one wins.
4. Rules that fired zero times. Keep them if they guard against something rare. Delete them if they don't.
5. Low-confidence calls. For Mike, these were park souvenir shops, resort snack stands, local food and portrait studios. Nothing a rule could have known without asking.

Run this every month. It takes one prompt, and it's the difference between trusting your categories and hoping they're right.

Why aim for 90% instead of 100%?

Because the last 10% is where the judgement lives. A souvenir from a family trip might be travel, gifts or entertainment, and only you know which. Let Claude sort the routine transactions and save the ambiguous ones for your weekly or monthly review.

Chasing 100% means writing ever-narrower rules that break the first time a merchant changes its name. Aiming for 90% keeps your rules simple and puts you in the loop exactly where you add value. The low-confidence question in your audit is how the hard 10% finds its way back to you.

If you want ideas for what to review together, see what to ask your AI about your money.

What can Claude build once your categories are clean?

Once your data is sorted, Claude can build views that fit how your household actually runs. Mike asked Claude for a household money view, and Claude built it as an artifact in under two minutes.

The view showed:

• Spending by category.
• Sinking funds.
• Debt, excluding mortgages.
• How well the rules were sorting transactions.
• What the view couldn't see.

That last item is the controller's instinct at work. A good report tells you its own blind spots. And because your context lives in Era rather than inside one finance app, you shape these views your way, in whichever AI tool you prefer.

Watch Mike audit an AI agent on real accounts

Mike Dion of AI For Finance walks through all three controls on their own accounts, 30 days after connecting them to Era: I Gave an AI Agent Access to My Real Bank Accounts for 30 Days. The full video plays below. If Claude's screens look different from the video, follow the steps in this guide.