Indian treasury teams are already pasting collections histories, payable ladders and covenant calendars into generative tools to produce a thirteen-week cash-flow forecast. The output arrives faster than a linked Excel model and looks more confident. That confidence is the problem. A generative model does not maintain a cash-flow waterfall; it produces a plausible continuation of the pattern it was shown. When that number reaches a board pack without a stress test that a non-specialist director can follow, the company has substituted narrative smoothness for treasury judgement. This briefing is a practical sequence for controllers, treasury heads and FP&A leads who must decide whether a GenAI cash-flow forecast is fit to leave the building.
The setting is Indian practice: listed companies whose treasury sits on a shared ERP extract, NBFCs whose collections still live in branch Excel files, MSME groups whose working-capital cycle is seasonal and informal, and campus or hospital finance cells that forecast fee or billing receipts on the same laptop used for payroll. Enterprise-grade forecasting engines exist for a few large groups. They are not the median desk. The stress test has to work where the median desk is — one analyst, a model the company does not host, and a board meeting on Friday.
Further reading on DrDPKlass: how finance controllers should contain shadow GenAI, and how CFOs should govern sanctioned GenAI in finance operations. Today’s note sits downstream of both: once a forecast is produced, test it before the board sees the number.
1. A GenAI cash-flow number is not a spreadsheet bridge
A conventional cash-flow forecast is a bridge. Opening cash, receipts by source, payments by type, net movement, closing cash. Each line can be traced to a source file, a collection assumption or a payment schedule. A reviewer can change one driver and see the effect. A generative model does not produce that bridge. It produces a sequence of numbers that is statistically consistent with the sequences it has seen. The consistency can be useful. It is not a bridge.
Three differences matter for Indian treasury. First, the model has no notion of a covenant test date. A number that looks smooth can breach a minimum cash or current-ratio covenant on the day the bank tests it. Second, the model does not distinguish a one-off advance from a recurring collection. An analyst who pastes a history that includes a large inter-company receipt will often receive a forecast that treats the receipt as recurring. Third, the model will fill gaps. If the payable ladder is incomplete, the model will invent a plausible continuation rather than flag the missing schedule.
Indian cash cycles add specific failure modes. Agricultural and commodity businesses see collections that move with mandi prices and monsoon timing; a model trained on the last twelve months will not invent the next monsoon. Exporters face realisation lags that move with shipping and bank documentation; a smooth forecast will miss a delay that is already visible in the bill-of-lading register. NBFCs and housing-finance companies have SMA and NPA movements that change collections non-linearly; pasting a collections series without the SMA ladder produces an optimistic continuation. None of these are exotic. They are the ordinary drivers a treasury head already tracks. The model does not track them unless the human puts them in front of it and then checks the output against them.
2. Map what the model was actually shown
Before any stress, write down what entered the prompt. Indian teams lose time later because the analyst cannot reconstruct the paste. A one-page input map takes fifteen minutes and prevents an hour of reconstruction in the audit committee.
| Input class | Typical content | What the model will do with it | What the human must do |
|---|---|---|---|
| Collections history | Daily or weekly receipts, 12–24 months | Continue the recent slope or seasonality | Flag one-offs and known future delays |
| Payable ladder | Vendor dues by week, incomplete | Smooth the gaps | Replace invented weeks with the actual ladder |
| Covenant calendar | Test dates and minimum cash | Ignore unless explicitly asked | Force the model to flag the test date |
| Management commentary | Last quarter’s narrative | Echo the tone | Remove tone; keep only drivers |
Ask the analyst three questions and write the answers on the map. Which file was pasted, and was it the live file or a dummy? Which periods were included, and were any large receipts or payments manually removed? Which future events were described in the prompt in words rather than in numbers? If the answer to the third question is “none,” the model has no information about a planned capex drawdown, a tax payment, or a large customer’s known delay. Those events must be added by hand after the model returns its sequence.
For listed issuers, note whether any unpublished numbers entered the prompt. A collections series that includes the current month’s still-open days is price-sensitive if the company is in the silent period. The input map is also a disclosure control. If the map cannot be produced, the forecast should not leave treasury.
3. Run three stress tests the board can read
A stress test that requires the director to understand a model architecture will not be read. Three simple stresses, each expressed as a change in a driver the board already knows, are enough to show whether the GenAI number is fragile.
Collections delay
Shift the largest three customers’ receipts by two weeks. Show the new closing cash on the covenant test date. If the model cannot be re-run, apply the shift by hand to the output sequence and note that the model itself was not re-tested.Payable acceleration
Bring forward the next month’s largest vendor payments by ten days, or apply a known early-payment discount request. Recalculate the low point. Many Indian vendors demand faster payment when their own working capital tightens; the forecast should survive that request.The third stress is the one Indian boards most often miss: a one-off that the history treated as recurring. Identify the largest single receipt in the pasted history that was an inter-company transfer, a government refund, or a one-time customer advance. Remove it from the history, re-run or adjust, and show the difference. If the difference is material on the covenant date, the original number was carrying an assumption the model never stated.
Seasonal businesses need a fourth, optional stress: a monsoon or mandi-price scenario that the last twelve months do not contain. Write the scenario in words, adjust the collections line by a percentage the commercial team can defend, and show the effect. Do not ask the model to invent the monsoon. Ask it only after the human has named the percentage.
Keep the arithmetic simple enough that an audit-committee member can check it with a calculator. A stress that lives only inside a prompt history is not a stress the board can rely on. The table is the deliverable; the model output is the working paper.
4. Separate the point forecast from the scenario narrative
Generative models produce language as easily as numbers. The language is often more dangerous. A sentence that says “cash remains comfortable through the covenant date” can survive even when the stress table shows a breach under a two-week delay. Separate the two.
The point forecast is the single number or the weekly sequence that treasury is prepared to defend as its base case. It must rest on named sources for every material driver. The scenario narrative is the set of sentences that describe what would happen if one of the three stresses materialised. Those sentences should appear under a heading that the board can see is not the base case. Mixing them produces a pack in which the smooth sentence is remembered and the table is forgotten.
A useful test: read the board paragraph aloud without the table. If a director would believe the company is comfortable, and the table shows otherwise under a realistic delay, rewrite the paragraph until the discomfort is visible in the words.
Indian packs often carry both a management commentary and a treasury appendix. Put the stress table in the appendix and one sentence in the commentary that points to it: “Under a two-week delay in the top three customers, closing cash on the test date falls below the covenant minimum; see appendix.” Do not bury the sentence. Directors read the first page.
For NBFCs and housing-finance companies, add the collections impact of a one-notch deterioration in the SMA book. The commercial team can supply the percentage; treasury applies it. The model is not required to understand SMA classification. The human is.
5. Document residual model risk before the pack leaves treasury
Even after the stresses, some risk remains. The model may have been shown incomplete data. It may have been run on a consumer account whose terms allow training. It may have produced a number that looks precise to the rupee when the underlying drivers are known only to the nearest week. Write the residual down.
- Input completeness. Which schedules were missing, and what did the model invent in their place? Name the invention.
- Tool status. Was the run on a sanctioned tenant with training switched off, or on a personal account? A personal account means the residual includes a confidentiality risk already covered in earlier notes.
- Precision claim. The model may return a number to the nearest thousand. Treasury should round to the level the drivers support — often the nearest lakh or crore for a group forecast.
- Reviewer. Who compared the output sequence to the source files and signed the stress table? A name and a date.
If the residual includes a material unknown — for example, a large customer whose payment behaviour has changed and whose new pattern is not yet in the history — say so in the board pack. A forecast that hides the unknown is a forecast that will surprise the board later. Indian directors have seen enough working-capital surprises; they will not thank a pack that smoothed them away.
6. What the audit committee and the external auditor will ask
Expect three questions. First: “Who owns this number?” The answer must be a named treasury person, not “the model.” Second: “What happens if the largest customer pays late?” The stress table is the answer. Third: “Did any unpublished or client data leave the company?” The input map and the tool-status line are the answer.
External auditors in India are beginning to ask how management produced forecasts that support going-concern or covenant representations. A GenAI draft that was not stress-tested and not minuted will be treated as an unsupported assertion. A draft that was mapped, stressed and minuted can be treated as one input among others. The difference is documentation, not sophistication.
For companies under limited assurance on BRSR or under enhanced review of related-party cash movements, the same file helps. Inter-company receipts that the model treated as recurring are visible in the one-off stress. The auditor does not need to understand the model; the auditor needs to see that management did not treat a one-off as recurring.
Campus and hospital finance teams face a simpler version of the same questions from their governing bodies. A fee-receipt forecast that a generative tool smoothed across a known admission delay will be challenged when the actual collection misses the payroll date. The same three stresses — delay, acceleration of payments, removal of one-offs — work without any change in language.
7. Field checklist and closing brief
Use this list before any GenAI cash-flow number is copied into a board pack, a bank covenant certificate, or a management representation.
- Input map exists: file, periods, one-offs removed, future events described in words.
- Three stresses calculated and shown in one table: collections delay, payable acceleration, one-off removed.
- Point forecast separated from scenario narrative; the first page of the pack does not claim comfort the table contradicts.
- Residual model risk written in one paragraph: completeness, tool status, precision, reviewer name and date.
- No Class A data (identifiers, unpublished price-sensitive series, client ledgers) entered a personal account.
- The number that will be spoken in the meeting is the base column after stresses, not the raw model output.
A GenAI cash-flow forecast can save an analyst an evening. It cannot replace the treasury head’s knowledge of which customer is slow, which vendor will demand early payment, and which receipt was a one-off. The stress test is the mechanism that forces that knowledge back into the number before the board sees it. Indian boards do not need a more confident sequence. They need a sequence whose assumptions a director can challenge in the meeting. The table and the minute give them that sequence. The raw model output does not.
AI & Business Series · 23 October 2026