A forecast used to live in a spreadsheet where anyone could trace a number to a cell. Plain-English scenarios open a new question: what stands behind each one?
Finance ran on a forecast that people built and people argued over. A model sat in a spreadsheet, the assumptions sat in cells, and the debate in the room was the control.
The old Container of Trust was the visible model. Anyone with access could trace a number back to an assumption and an author, and an auditor tested the numbers against the ledger.
One forecast could be examined line by line, so the room knew what it was betting on.
One forecast, debated in the room.
Fifty scenarios on demand, asked for in plain English.
Forecasts, variance commentary and investor updates now auto-draft. Cash flow scenarios are requested in plain English, and portfolio memos are summarised from filings.
The tools can be good. In one study, a language model predicted the direction of future earnings changes from anonymised financial statements with about 60 percent accuracy, against about 53 percent for analysts.
They can also be wrong in ways that read as right. On a 2023 benchmark of questions about public company filings, a leading model with a retrieval system answered incorrectly or refused to answer 81 percent of the time.
Finance now holds fifty answers and one board. Each scenario is cheap, so the work moves from producing numbers to knowing which model, with which assumptions, produced them.
The old Container of Trust assumed the model was visible. When a scenario comes from a request typed in plain English, the assumptions live inside the tool, and the auditor needs another way to see them.
A spreadsheet showed its assumptions in cells. A generated scenario carries them inside the prose and inside the model.
Shift 01 · Assumption Extract
One forecast could be debated. Fifty cannot, and the room either picks the first or averages them without asking why they differ.
Shift 02 · Forecast Reconciliation
A filing summary used to be checked by whoever wrote it. A survey of hallucination describes how summaries can include content their source never held.
Shift 03 · Filing Summary Check
A confident answer invites acceptance. A field experiment with consultants found that on a task outside the AI's strengths, those using it were less likely to reach the correct answer.
Shift 04 · Confidence Band Record
An analyst's forecast was challenged by peers. A 2026 review of automation bias found that people over-rely on automated outputs, and that explanation tools often fail to reduce it.
Shift 05 · Analyst Challenge Log
Banking supervisors' long-standing guidance on model risk management asks for validation, documentation and governance of models. Generative tools are models, and few finance teams treat a prompt as one.
Shift 06 · Prompt Model Inventory
A model's answer to the same question can change from one run to the next, so a scenario that cannot be reproduced is hard to audit.
Shift 07 · Run Reproducibility Record
A vendor's assurance is an incomplete answer for a supervisor. A generative tool used for a stated purpose needs an independent record of how it was tested and where it stops being reliable.
Shift 08 · Model Validation Certificate
The Container of Trust has moved from the spreadsheet to the prompt. The analyst who once owned the assumptions now owns the request, and the model holds the assumptions it chose to apply.
That leaves room to define accountability. If a board acts on a scenario that rested on an assumption no one chose, the model supplied it, the analyst accepted it and the auditor tested the arithmetic. A US standards body's generative AI profile lists confabulation among its twelve risks, and a forecast is where it becomes a number.
Every earlier era that met an abundance went on to build a Container of Trust, and finance has room to build one of its own. In The Series, its working name is the Assumption Registry.
The name is a working label, and a better one may arrive. What matters here is what it would hold: the assumptions behind a number, the model that produced it and the proof that the answer can be reproduced.
The eight shifts for Finance in Market Signals come in two documents. See what is changing is the brief profile, a short map with one entry per shift. See what can be built is the detailed profile, the longer working record behind each shift.
Market Disruptions records what is happening in the information market. Market Signals shows the potential ahead, and for finance it comes as two documents. Both cover the same eight shifts, at different depth.
The pattern repeats across all sixteen functions. More output arrives, trust in it is tested, and someone builds the Container of Trust that lets people believe the output again. Start with the brief for the overview, or open the detailed profile for the working record.
The external sources behind this article, reviewed in October 2026. Links open in a new tab.
Tests a language model on predicting earnings changes from financial statements, against analysts.
Tests language models on questions about public company filings.
Reviews how language generation produces text unsupported by its source, and how researchers measure it.
Tests 758 consultants with and without AI, and shows where AI raises quality and where it lowers accuracy.
Reviews why people over-rely on AI recommendations, and why explanation tools often fail to reduce it.
Sets out how banks should validate, document and govern models, and defines model risk.
Names twelve risks specific to or made worse by generative AI.
Isyamba, V. (2026). Finance: Fifty Scenarios on Demand and a Record of What Stands Behind Each Number. The Trust Ledger Project Series. GreenDeveX.com. https://greendevex.com/trust-ledger/market-disruptions/finance/
Isyamba, Victor. “Finance: Fifty Scenarios on Demand and a Record of What Stands Behind Each Number.” The Trust Ledger Project Series, GreenDeveX.com, 2026, https://greendevex.com/trust-ledger/market-disruptions/finance/
Isyamba, Victor. “Finance: Fifty Scenarios on Demand and a Record of What Stands Behind Each Number.” The Trust Ledger Project Series. GreenDeveX.com, 2026. https://greendevex.com/trust-ledger/market-disruptions/finance/
@misc{isyamba2026finance,
author = {Isyamba, Victor},
title = {Finance: Fifty Scenarios on Demand and a Record of What Stands Behind Each Number},
year = {2026},
publisher = {GreenDeveX.com},
series = {The Trust Ledger Project Series},
url = {https://greendevex.com/trust-ledger/market-disruptions/finance/}
}
Note: The perspectives shared on GreenDeveX.com are the author's independent analytical interpretations of trust as a historical medium of exchange. This commentary is intended for conceptual exploration and does not constitute peer-reviewed academic research or commissioned institutional findings.