What it does
Every week the finance team spends time rebuilding the cash flow forecast: exporting ledger entries, consolidating them in Excel, recalculating for “what if” cases.
The automation removes this routine and moves the forecast from “a snapshot once a month” to “an up-to-date picture every day”.
The AI agent takes data from the accounting system and the data warehouse / BI, builds a 30/60/90-day forecast and adds a written explanation. The finance manager doesn't get a bare table — they get a short report explaining what has changed and why.
What the automation does
- Pulls actual data from accounting (receipts, payments, liabilities) and from the data warehouse (contract schedules, recurring subscriptions, sales forecast).
- Categorizes movements: operating receipts, taxes, payroll, rent, one-off payments.
- Builds the forecast for three horizons: 30, 60 and 90 days.
- Produces three scenarios: base (actuals and confirmed commitments), optimistic (factoring in the sales pipeline), pessimistic (with key receipts delayed).
- Compares it with the previous version of the forecast: what has changed, which new risks have appeared, where the cash position is trending.
- Prepares a written summary — a short explanation of the figures in a format an owner without a finance background can understand.
- Delivers the report on a schedule (for example, every Monday morning) or on request — by email, Slack or Notion.
What the automation doesn't do
- It doesn't replace the treasurer or the CFO. Decisions on payments, negotiations with the bank and raising financing stay with a person.
- It doesn't predict revenue that isn't in the system. If the sales pipeline lives in a manager's head rather than in the CRM, the forecast will be incomplete.
- It doesn't fix data at the source. If the books aren't closed or entries are delayed, the forecast will reflect that but won't correct it.
How it works
Under the hood is a combination of connectors to the sources, an ETL layer for data normalization, a forecasting model and an LLM that generates the text commentary. It is built as custom Python code: off-the-shelf SaaS solutions rarely cover the specifics of a particular company's accounts, counterparties and tax schedule.
Architecture flow
Raw data comes from two categories of sources. Accounting (QuickBooks, Xero, NetSuite — whichever is in use) provides actual balances, ledger entries, counterparties and open invoices. The data warehouse / BI (BigQuery, Snowflake, Metabase) provides product metrics, contract schedules and an extended sales pipeline. If there is no warehouse, the AI agent works directly with the CRM and spreadsheets.
Next, the ETL layer brings the data into a single model: accounts are rolled up into movement categories, recurring payments (rent, subscriptions, salaries) are flagged as known commitments, and one-off payments as scenario items. The forecasting model builds three trajectories and calculates the balance for each day of the horizon.
The LLM (an AI model or equivalent) takes the figures and context as input and generates a written summary: what's in focus this week, which risk matters most, what has changed since the previous forecast.
Implementation steps
- Source audit. Describe where the actual data lives (accounting), where the planned data lives (CRM, contracts), and which expense categories need tracking.
- Account mapping. Consolidate the chart of accounts and CRM categories into a single taxonomy for the forecast (operations, payroll, taxes, investments, one-off).
- Connector setup. APIs for accounting and the data warehouse / BI; if the API is limited, data comes in through a regular export.
- Forecasting model build. The core logic is recurring + scheduled + scenario adjustments. No ML in the first iteration, so the result stays explainable.
- Scenario setup. Parameters for the optimistic and pessimistic scenarios are agreed with the CFO: % of delayed receipts, % of the sales pipeline that falls through, accelerated payments.
- Connecting the LLM layer for the written summary. A prompt with the company context, the latest forecast and the changes. The tone and format of the commentary are agreed over several iterations.
- Delivery channel. Email with a PDF, a Slack message, a Notion page, a BI dashboard — chosen based on how the team already works with financial reporting.
- Launch in shadow mode for 2-4 weeks. The AI agent builds the forecast in parallel with the manual version; discrepancies are reviewed and calibrated.
Components
Layer | What it does | Technology |
|---|---|---|
Connectors | Pull data from accounting and BI | Python + API clients |
ETL | Normalization and categorization | Python (pandas) |
Model | Builds 30/60/90 scenarios | Custom code |
LLM summary | Text commentary | language model |
Delivery | Report to the chosen channel | Email / Slack / Notion API |
Orchestration is a low-code platform or code in Airflow / cron, depending on what the team already uses. For SMBs it's more often a workflow engine: a visual workflow is easier to maintain without dedicated DevOps.
Prerequisites
Implementation requires access to the data sources and a little preparatory work from the finance team.
Access and data
- Access to the accounting system (QuickBooks, Xero, NetSuite or another) — ideally via API, exports as a last resort.
- Access to the data warehouse / BI or to the sources the sales pipeline is built from (CRM, billing, contract spreadsheets).
- An up-to-date chart of accounts and movement taxonomy.
- Information on recurring commitments: rent, subscriptions, payroll, taxes.
Team readiness
- A CFO or finance manager who signs off on the taxonomy and scenario parameters. This isn't a one-off task: the first iteration needs 4-6 working sessions.
- Willingness to run in shadow mode for 2-4 weeks: manual and AI forecasts in parallel for calibration.
- Access to someone who understands the data structure in the sources (accountant, BI owner).
Timeline
A basic implementation takes 6-10 weeks:
- Source audit and mapping — 2 weeks.
- Model and integration build — 3-4 weeks.
- Shadow period and calibration — 2-4 weeks.
The timeline grows if data in the sources is fragmented or if non-standard scenarios are needed (multi-currency forecasting, factoring, handling advance payments from large clients).
Pain points
- Poor Forecasting (cashflow/sales/stock)
- Time on Manual Reports
FAQ
How long does implementation take?
A basic launch takes 6-10 weeks. Of that, 2 weeks go to the source audit and chart-of-accounts mapping, 3-4 weeks to building the model and integrations, and 2-4 weeks to shadow mode, when the AI forecast runs alongside the manual one for calibration. The timeline grows if the books close late or multi-currency logic is needed.
What if we don't have a data warehouse / BI?
You can work without a data warehouse — the AI agent pulls data directly from accounting and the CRM. In that case, forecast accuracy is limited by how accurately the sales pipeline is reflected in the CRM: if the sales team doesn't log deals in the system, or logs them haphazardly, the receipts forecast will be partial. In the first iteration this doesn't block the launch — even a basic auto-updating forecast gives the CFO more predictability than a manual monthly report.
What are the risks and what can break?
The main risk is garbage data going in. If the books close 2-3 weeks late, the forecast will run on stale balances. Technical failures happen on the connector side: a provider's API change, token rotation, a change in export structure. Plan for monitoring and a monthly reconciliation against actuals — especially in the first 2-3 months after launch.
Is it a fit for our industry?
The AI agent suits Professional Services / Consulting (where the cash balance depends on the contract and payment schedule), SaaS / Tech (where there is a recurring and a variable part), and horizontal cases at any SMB. For companies with a heavy inventory cycle (e-commerce, manufacturing), the forecasting logic needs additions: inventory, payment deferrals, seasonality.
How often is it recalculated?
By default, an updated forecast every week on Monday, plus on request. Technically the model recalculates daily, but sending the report once a week is more convenient for the finance team and the owner. Triggered recalculations kick in on major events: a key client's payment, a one-off expense, a change in the sales pipeline.
How is this different from an Excel export from the accounting system?
An Excel export from your accounting system or QuickBooks gives a point-in-time snapshot and quickly goes stale. The AI agent ties the forecast to actuals, adds scenarios, updates automatically and accompanies the figures with a text commentary. The finance team stops assembling the report by hand and moves on to analysis: what to do with the cash flow picture it now sees.
Want this in your business?
Book a free audit — we'll show how this automation will work for you.