# I Built an AI CFO That Monitors The Business 24/7 in Claude

https://www.youtube.com/watch?v=idcAjzDCfEA

[00:00] Most finance teams are sitting on everything they need to monitor the business continuously.
[00:05] The data is already there.
[00:07] The accounting system is already live and the employees are already updating it every day, but none of that translates into real-time financial awareness because the analysis layer does not exist.
[00:18] Someone still has to go in, they've got to pull the data, they've got to build the model and then write the commentary before leadership can see what's happening.
[00:26] Now, to solve this, as a quant and finance automation developer, I've created an AI CFO continuously monitors the business, automatically identifies overdue receivables, customer concentration risk, collection issues, and cash flow concerns, and then it turns those findings into a complete executive review with prioritized recommendations for management.
[00:48] Instead of manually reviewing reports every month, leadership gets an on-demand CFO briefing built directly from the latest financial data.
[00:57] I'll show you exactly how to build it by creating the AI CFO.
[01:00] Framework first, then connecting it to live financial data, and then building the analyst layer that continuously monitors the business and feeds those findings into the review process.
[01:11] Now, if you want the prompts and the data sets from this video, they're in my free community.
[01:15] Link is in the description below, but for now, let's get into it.
[01:19] Building the AI CFO framework in Claude.
[01:22] First things first, I open a fresh Claude session and I connect it to a dedicated skill development folder.
[01:29] Now, we're building the framework first before connecting any data because connecting data and asking questions right away tends to give inconsistent answers because the AI has no clearly defined role yet.
[01:42] And I want this system to behave like an actual Chief Financial Officer, so I have to define that role first.
[01:47] I have to save it as a reusable skill for future use.
[01:51] And this folder is where Claude generates the skill so I can save it and call it later on.
[01:56] Now, the prompt asks Claude to build a complete skill called AI CFO.
[02:03] Important part is what the skill is not allowed to do.
[02:08] It does not run accounting calculations.
[02:10] It doesn't do any bookkeeping.
[02:11] It doesn't do reconciliations or any invoice processing.
[02:16] Those would be handled in other parts of the workflow.
[02:19] This skill only starts working after the analysis is already done.
[02:25] So, its job is to read finance outputs like KPI summaries or receivables dashboards, risk rankings, and alerts, and then turn those into executive decisions.
[02:37] Now, that separation matters because it mirrors how a real finance team works.
[02:42] Analysts, they gather the data and then they build the reports.
[02:47] The CFO then reads those reports and decides what matters, what creates risk, and what the business should do next.
[02:56] And that skill is built to do the second part.
[02:58] The review method inside the skill follows the same path a CFO uses.
[03:01] It starts by understanding the.
[03:05] Overall situation, finding the material movements, connecting the related issues into one story, and then prioritize the biggest risks and recommends actions.
[03:16] Now, that is what keeps the output decision focused rather than just another list of metrics.
[03:24] The prompt also defines how every finding has to be expressed.
[03:29] The skill is not allowed to just repeat numbers.
[03:31] For each finding, it has to answer three questions.
[03:34] Why does this matter?
[03:36] What risk does it create?
[03:39] And what should management do next?
[03:41] Now, those three questions shape every recommendation that the system produces later on.
[03:45] Now, the last requirement is the output format.
[03:47] The skill produces a self-contained HTML CFO review dashboard.
[03:53] Now, the final deliverable needs to feel like something a CFO would actually put in front of leadership.
[04:01] Claude first checks the skill authoring best practices to make sure the structure will be valid.
[04:06] Then it builds the package, which is a skill file with the role and instructions, supporting reference documents, and the HTML dashboard template.
[04:16] Now, during the build, it measures the description field against the length limit.
[04:19] It finds it's too long, and then it trims it automatically until it fits.
[04:23] Before installing, I read through what Claude has produced.
[04:27] The skill is built to consume analytical outputs from a connected workbook without being hardcoded to a specific file or tab structure, which means I can then point it at any similar data set later and reuse the same logic.
[04:46] The decision framework focuses on business impact and risk instead of calculations.
[04:51] Now, the dashboard template is built to open with alerts and an executive summary because those are what a CFO would look at first, then move into the KPI summary, the aging analysis, and customer risk ranking, then the collection forecast, and close with the business risks and
[05:08] Recommended actions.
[05:11] Now, that order follows how a real executive review actually unfolds, starting with what's urgent and ending with what to do about it.
[05:23] Once I confirm all of that, I click save skill and Claude installs it as AICFO, and it's available for any future review.
[05:32] Setting up QuickBooks and Quadratic.
[05:34] The system needs an operational source of truth, and for that, I use QuickBooks.
[05:39] It holds the customer records, the invoices, the payments, and the rest of the day-to-day accounting activity.
[05:44] Now, the most important piece of this workflow is the invoice and payment data because that is what tells us who owes money and what has actually been collected.
[05:54] What makes QuickBooks the right starting point is that it is already live.
[05:59] Employees create invoices, record payments, and update customer records as part of their normal day-to-day work.
[06:06] The data keeps evolving on its own, which means the monitoring layer always has current
[06:10] Information to work from without anyone preparing a thing.
[06:15] The layer that turns that raw activity into analysis is Quadratic because QuickBooks captures the business as it happens, Quadratic monitors and analyzes it every day, and then Claude reviews the findings and produces the CFO briefing on demand.
[06:31] At first, Quadratic looks like another spreadsheet, but it's built differently.
[06:35] It's an AI spreadsheet that connects directly to your business system and lets the AI do the analysis inside the workbook itself.
[06:45] Now, a few things about it matter specifically for finance.
[06:48] And the first is that it connects straight to your data sources like QuickBooks, databases, and maybe other accounting tools.
[06:56] And since it pulls from the source instead of a manual export, there is less room for the errors that creep in when you move files around by hand.
[07:05] Now, the second is auditability, and it's the most important one for finance because when the AI works inside
[07:12] Quadratic, every formula, every transformation, and every step it takes stays visible in the workbook.
[07:19] And that gives you a real verification layer.
[07:21] You can trace any number back to the exact logic that produced it instead of just trusting an answer that you can't see behind.
[07:30] Now, the third is that it keeps itself current.
[07:33] Once your data sources are connected, then the workbook refreshes on its own.
[07:37] And when QuickBooks updates, Quadratic pulls the new data and recalculates.
[07:42] Quadratic also holds enterprise agreements with Claude, so the data you work with carries the same protections as using Claude directly, and it's not used to train external models.
[07:54] Connecting the two takes a minute.
[07:56] Inside Quadratic, I go to the connection section in the left navigation.
[07:59] I click new connection, find QuickBooks in the list, and click connect.
[08:05] That sends me through the QuickBooks authorization flow where I sign in and I grant access.
[08:09] And once that's done, Quadratic can pull company data directly from QuickBooks whenever.
[08:13] It needs it.
[08:16] Then I connect Quadratic to Claude so that the AI CFO can read those outputs later.
[08:21] I open Claude co-work, I go to the customize in the sidebar, and I open the connectors section.
[08:27] Then I click the plus icon, choose add custom connector, and enter Quadratic as the name along with the Quadratic MCP server URL.
[08:37] I click add, and then Claude registers it as a connector.
[08:40] I've reviewed the available permissions, which covered reading files, accessing spreadsheet data, authenticating with the Quadratic environment, and writing data when required.
[08:51] And with the connection in place, Claude can now open AI CFO workbook, read any analytical tab, and pull the financial intelligence generated by the finance analyst layer into its own review process.
[09:06] Building the AI finance analyst inside Quadratic.
[09:09] Now it's time to turn Quadratic into the finance analyst.
[09:12] I go back to the
[09:15] Quadratic homepage, and I create a new workbook, and I name it AI CFO Command Center.
[09:20] Now this becomes the analytical workspace that sits between the QuickBooks data and the CFO review in Claude.
[09:27] Before building, I check the model in the selector inside the chat, and I choose Opus because this build needs heavy reasoning across the full data set.
[09:40] And the prompt tells Claude to act as a senior analyst and a finance systems architect at the same time.
[09:46] Now, this instruction is to build dynamic analytical tabs that recalculate every time the QuickBooks data refreshes or the file is rerun.
[09:57] It defines six tabs, each covering one area of financial oversight.
[10:02] Now, the first is an executive KPI summary, the one-page view of revenue, collections, outstanding receivables, and top customers.
[10:11] The second is an accounts receivable aging dashboard that sorts every outstanding
[10:16] Balance into current, 30, 60, 90, and 90-plus day buckets.
[10:22] The third is a customer risk ranking that scores each customer on balance, overdue invoices, and concentration.
[10:31] Now, the fourth is a collection forecast that estimates expected cash inflow over the next 7, 30, and 60 days.
[10:39] And the fifth is an overdue invoice monitor that lists the specific invoices needing action.
[10:43] The sixth is a CFO alerts engine that fires when conditions like customer concentration or aging cross a set threshold with the severity, the business impact, and the recommended action attached to each individual alert.
[11:00] Claude does not jump straight into dashboards.
[11:03] It first inspects the connected QuickBooks data to see what tables and what fields exist.
[11:08] And then it lays down a raw data tab that pulls live from QuickBooks and acts as the single source inside the workbook.
[11:17] Every dashboard references that layer, so the whole workbook stays in sync with the accounting data.
[11:24] And from there, it builds the six tabs in order, confirming the invoice and payment fields are correct before constructing each one.
[11:31] And it finishes by applying the charts, currency formatting, and risk highlights.
[11:37] The alerts engine is the piece that changes what the workbook actually is.
[11:42] So, where the other five tabs describe the state of the business, the alerts tab keeps checking it against set thresholds on every single run.
[11:50] So, when concentration climbs too high or invoices age past the limit, it raises the flag on its own.
[11:58] And that is what turns the workbook from a passive report into an analyst that monitors the business.
[12:06] Now, the last piece is keeping it current.
[12:08] So, I go to the run section in the left side bar.
[12:10] And this is where the workbook's calculations, queries, and charts get executed.
[12:15] I set the workbook to refresh daily, so it automatically stays up to date.
[12:19] Run executes, then Quadratic first refreshes the connected data tables and pulls in any new data since the previous run.
[12:27] Then, it automatically updates all downstream analysis across the workbook.
[12:33] To check everything works before I trust the schedule, I select all cells in file, which then forces every query, every calculation, every chart, and every alert to run.
[12:45] And I watch the aging charts redraw, the forecasts regenerate, the risk scores recalculate, and the alerts re-evaluate against the latest date.
[12:55] And that means whenever I open the workbook, I'm looking at fresh data and updated insights rather than stale analysis.
[13:04] Now, that full run is also my verification.
[13:07] I confirm all six tabs built and populated, that they reference the raw data layer rather than the hard-coded numbers, and that the charts and alerts refresh cleanly when every cell runs.
[13:17] Now, once that passes,
[13:20] The workbook is a working finance analyst that monitors the business on its own.
[13:27] Now, if you want to try Quadratic yourself, the link is in the description below.
[13:30] Generating the CFO review with Claude.
[13:33] With the analyst up and running, the last step is the review itself.
[13:37] So, I open a fresh Claude session because I want it pulling the latest outputs directly from Quadratic rather than learning on earlier context.
[13:44] Now, the prompt here is short.
[13:46] I tell Claude to review the AI CFO Command Center workbook, ignore the raw data tab, use the six analytical tabs, and run the review with /aicfo to invoke the skill.
[14:02] Now, it can be this short because all the thinking already lives in the skill.
[14:08] And that prompt only has to point at the workbook and then trigger the framework.
[14:13] So, Claude loads its tools and connects to Quadratic.
[14:16] And since it is a fresh session, it needs authorization first.
[14:20] So, it gives me a verification link and a code.
[14:20] I open the
[14:22] Link, sign in to Quadratic, approve the access, and Claude is connected.
[14:27] It finds the workbook, confirms the six analytical tabs are there, and leaves raw data out as instructed.
[14:34] Then, it reads each tab and pulls the metrics, the rankings, the forecasts, and the alerts that the analyst generates.
[14:41] Now, before writing anything, it cross-checks the figures against the tab so that the review ties back to the source.
[14:49] Now, once it has the full picture, it loads the AI CFO skill, and then the work shifts from analyst to judgment.
[14:57] Claude reads the receivables, the concentration risk, the forecasts, and the alerts through the CFO framework.
[15:04] It weighs the business impact of each, and then it turns them into executive recommendations.
[15:10] And the output is a self-contained HTML CFO review dashboard.
[15:15] So, I open it in the browser, and it reads like a management reporting package rather than an AI answer.
[15:20] At the top is a CFO section that puts the
[15:23] Urgent issues first.
[15:26] Over 70% of outstanding receivables have aged past 90 days.
[15:29] Every open invoice is overdue.
[15:32] Several customers carry enough balance to be a concentration risk.
[15:36] And for each one, the dashboard says why it matters and what to do, not just that it exists.
[15:42] The executive summary is where the framework really shows its value.
[15:47] The main conclusion is that the company has a collections problem.
[15:52] So, revenue is healthy, the money has been invoiced, but it is not yet turned into cash.
[15:57] And that distinction changes the whole management response.
[16:02] It points the team toward credit control and collection efforts instead of chasing more revenue.
[16:08] And from there, the KPI summary gives the snapshot.
[16:10] It shows invoice revenue, collection rate, outstanding and overdue balances, and average days outstanding.
[16:19] And each metric is labeled as a strength or a risk.
[16:22] Now, the aging analysis pairs
[16:26] A table with a chart, and it shows most of the receivables sitting in the oldest, hardest-to-collect bucket.
[16:33] The customer risk ranking takes balance size, days overdue, payment behavior, and concentration, and turns the customer list into a priority order based on exposure.
[16:45] And the collection priority section answers a practical question, which is where the team should spend its time this week.
[16:51] The cash collection forecast is a very good example of the CFO layer doing its job.
[16:57] The underlying forecast suggested the full balance could be collected quickly, but Claude flagged that as being too optimistic given how old the receivables are, and it called for more realistic recovery expectations.
[17:10] So, the dashboard then consolidates the major threats into a key business risks view covering collection risk, liquidity risk, concentration risk, and the forecasting risk is just flagged.
[17:23] They each have their own consequences if it goes unaddressed, and then it closes with the
[17:27] Recommended actions carrying priorities, ownership, and next steps.
[17:33] So, that's the full system.
[17:36] QuickBooks captures the business as it happens, Quadratic monitors and analyzes it every day, and then Claude reviews the findings and produces the CFO briefing on demand.
[17:46] Now, remember, if you want the prompts, the skills, and the data sets from this video, they're all in the free community, which is linked in the description below, and I'll see you next time.