For construction finance managers and commercial teams who want to see a cash shortfall weeks before it arrives.
Key takeaways
- More than 3,400 Australian construction companies went into external administration in 2025-26, about one in four of all company insolvencies in the country.
- Agento reads the cost report you already keep in Excel, forecasts the next eight weeks of project cash with TimesFM 2.5 and writes the answer back into your workbook.
- In the demo project, the forecast gave a 73% chance of cash going below zero within eight weeks, with the shortfall expected in week 8.
- Power BI refreshed with no clicks, and Agento checked that the dashboard showed the same numbers as the forecast.
- A Forecast Analyst agent explains the result in plain words and recommends a next step. A drawdown request is only drafted and sent after a manager approves.
- The project and its numbers in this walkthrough are synthetic demo data.
The problem: project cash planning still starts with a hunch
Construction had more than 3,400 companies enter external administration in 2025-26. The next highest industry had 2,078. On most projects, cash planning still looks like this:
- Cost reports in Excel. A new workbook every week, each one a snapshot.
- Forecasts by gut feel. Someone experienced reads the numbers and makes a call.
- Drawdowns found too late. The shortfall shows up when there is little time left to arrange funding.
The question a finance manager needs answered every week is simple: will this project need a drawdown in the next eight weeks?
Why a forecasting model, not a chat model
We asked that same question on 97 past weeks and compared three ways of answering it. The Brier score measures how far off each one's chance of a shortfall was. Lower is better.
| Method | Brier score |
|---|---|
| TimesFM 2.5, a time-series forecasting model | 0.044 |
| Statistical baseline (damped trend) | 0.050 |
| General-purpose LLM (Claude Opus 5) | 0.337 |
TimesFM's score was about eight times lower than the general LLM's on the same weeks.
| Measure | TimesFM 2.5 | General LLM |
|---|---|---|
| Shortfalls caught | 94.8% | 62.5% |
| Valid, checkable answers | 100% | 87.5% |
| Compute cost | $0.007 for 97 forecasts | $0.12 for 8 calls |
| Hosting | $93 to $311 a month, always on | None, pay per call |
Two things are worth saying plainly. The simple statistical baseline came close to TimesFM, which is the point: this is a job for a forecasting method, and a general chat model is the wrong tool for it. And the test was a replay on a synthetic project ledger from 3 November 2024 to 6 September 2026, with the LLM capped at 8 calls. Hosting is priced for one always-on replica with 4 vCPUs and 8 GiB of memory.
Agento hosts TimesFM 2.5 for you in the Agento cloud, so there is no model for your team to run.
How it works
1. Connect your PC
Your Excel and Power BI stay on your desk. You add your finance PC to your workspace, and Agento works in both applications on that machine.
2. Say it, and Agento builds it
Describe the workflow in a sentence. The composer asks what it needs to know, finds the tools and lays out every step. The demo workflow has 15 steps and asks the same question on every run: will the M80 project need a drawdown in the next eight weeks?
One run, three results
The forecast: Excel in, TimesFM out
Agento reads the cost report, which was 439 rows in the demo, tidies it into a cash ledger and sends it to TimesFM. The eight-week forecast came back in 6.2 seconds. The answer lands in your own workbook:
| Result | Value |
|---|---|
| Chance cash goes below zero | 73% |
| First shortfall week | Week 8 |
| Low case (P10) | −$948,320 |
| Likely case (P50) | −$292,733 |
| High case (P90) | +$389,213 |
P10, P50 and P90 are the low, middle and high cases. The model expects cash to land above the P10 figure nine times in ten, and above the P90 figure one time in ten.
The dashboard: Power BI refreshes itself
Agento refreshed the Power BI report in 5.9 seconds with no clicks. It then read the dashboard back and checked that it showed 73% and week 8, so the report and the forecast cannot quietly disagree.
The advice: the analyst explains it
A Forecast Analyst agent turns the numbers into a next step. In the demo it explained that the forecast covers the eight weeks from 2 November to 27 December 2026 for the M80 Motorway Upgrade project, which opened with a cash balance of A$2.4 million and no undrawn facility headroom. It said a shortfall before Christmas was more likely than not, and recommended a drawdown before Christmas.
You can also ask the analyst yourself in chat. It has a job and read-only access to the forecasting tool, so it can explain a forecast but cannot change one.
Who makes the call? A manager.
When a shortfall is likely, the workflow stops and asks a manager for approval. Only after that approval does it draft the drawdown request, create the request document, add a Planner task for the finance manager and send the email.
When the forecast shows no shortfall, the drawdown branch is skipped.
Either way, every step of the run is recorded as evidence, so you can show how each cash decision was reached.
