CFOs should communicate an AI investment forecast as a conditional estimate, with a clear basis, range and funding decision. Separate what the company has observed from what it expects to happen, explain which assumptions could change the answer and keep the original estimate available when results arrive.
A revised forecast should improve the next decision without erasing the earlier expectation. The board needs to understand both what management now believes and what the company has learned about its ability to deliver the proposed benefit.
This concerns forecasts of value from an AI investment. If you are preparing the initial funding proposal, the AI adoption business-case guide explains alternatives, full costs and benefit ownership. Here, the task is to communicate uncertain expectations and reconcile them with later evidence.
Name the benefit and its evidence status
Start with the outcome being forecast, its period and the comparison. “AI will save money” conceals too many different claims. A reduction in a supplier charge, usable staff capacity and a better customer experience need different evidence.
Suppose a 260-person exhibition-services company uses AI to draft stand-installation completion reports. The drafts draw on crew checklists, job photographs and approved stand plans. Sam Taylor, the operations lead, keeps a site coordinator responsible for checking what is complete and what remains unresolved before a report goes to the client.
Claire Bennett, the CFO, is considering wider use because the company pays an external service to help prepare reports during busy periods. Faster drafting might reduce that charge. It will do so only if the relevant work can be removed from the service and the supplier's charges actually change. The coordinator's quality check still matters.
Claire should distinguish these evidence states:
- Observed work. What happened in the reviewed sample, including checking, corrections and the kinds of reports included.
- Expected benefit. What management believes could follow at the proposed volume, with the necessary operating and commercial conditions.
- Realized result. What the company can substantiate for the completed period on the agreed financial or service basis.
An observed reduction in drafting effort can support a forecast. It does not establish the later invoice reduction. State the missing connection beside the number, where a decision-maker can see it.
In an October 2026 Reddit discussion, a Reddit user described a sales team's vendor-reported time savings without additional deals or revenue. The account is unverified, with no stated employer or company size. Its useful warning is the gap between the reported intermediate measure and the business outcome the reader expected.
Show what makes the range move
A range becomes useful when people can explain its conditions. Widening a single estimate by an arbitrary percentage does not establish how uncertainty was assessed.
The UK government's April 2026 Digital and Data Benefits framework recommends testing assumptions such as uptake, productivity and costs through sensitivity analysis. It concerns public-sector economic cases. Its illustrative percentages are not forecasts for a private company's workflow.
For Claire's report-preparation proposal, separate the questions that would move the result:
| Uncertainty | Evidence to seek | Implication for the funding discussion |
|---|---|---|
How much work is suitable? | Report types, demand and the records available | A small tested population may not represent peak-period work |
How much complete effort changes? | Preparation, checking and correction across comparable reports | Faster drafting may leave total effort unchanged |
Which expenditure can change? | The service scope, charges and commercial owner's confirmation | Released staff time does not automatically reduce the invoice |
What else affects the outcome? | Other process, price and demand changes | A smaller bill may have several explanations |
Keep the assumptions coherent. A scenario with more eligible work may also require more support and checking capacity. Combining every favorable input can produce a mathematically correct result for an arrangement the company cannot operate.
Describe scenarios in words before showing their financial consequences. For example, one case may retain the existing external service because report quality or commercial flexibility is insufficient. Another may remove a defined part of its workload after acceptance. Explain what happens to costs and useful work in each case.
The authors of Forecasting Principles and Practice describe scenarios as forecasts under plausible conditions, rather than necessarily likely outcomes. Calling a selected low-to-high range a “95% prediction interval” requires a statistical basis that choosing scenarios alone does not provide. Ask the analyst to state the method and limitations; otherwise label the range as conditional scenarios.
Keep the original estimate beside the current outlook
The current forecast should use the best available information. The original forecast remains useful for understanding what management expected when it committed resources.
Suppose Claire initially forecasts US$12,000 of gross reduction in report-preparation charges for one quarter. Six weeks into that quarter, a narrower eligible workload and later-than-expected acceptance lead her to forecast US$6,000 for the whole quarter. At quarter-end, finance substantiates US$4,000 of reduction against the agreed comparison.
Those are three views of the same quarter. The US$6,000 outlook already covers the full quarter, including the elapsed weeks. It is not an amount to add to the eventual result.
Three views of the same quarter
Claire keeps the gross reduction in external report-preparation charges on the same quarterly basis.
US$12,000 expected before the quarter
Retain the original date, eligible report volume and acceptance assumptions. This estimate is the starting expectation, not money already saved.
US$6,000 expected for the whole quarter
Explain the narrower eligible workload and later acceptance. This replaces the current outlook, not the original decision record.
US$4,000 substantiated against the comparison
Keep the evidence and any attribution limits. Show the quarter’s incremental costs separately before claiming a net financial return.
Preserve the date, scope and assumptions attached to each view. If the team expands the report types or changes the supplier arrangement, identify the changed scope before comparing totals. Otherwise a favorable variance can reflect a different job rather than better delivery.
HM Treasury's Budget holder forecasting handbook recommends communicating possible outcomes and their drivers, recording assumptions and updating them as information changes. It is guidance for public budget holders. The reporting discipline is useful here too: an updated expectation needs an explanation, not a rewritten history.
Keep costs visible on their own basis. If the example's complete incremental cash costs for that quarter are US$7,000, the US$4,000 gross benefit leaves a US$3,000 net cash cost. It would be misleading to report US$4,000 as the net return. Existing staff effort and service quality still need to be considered alongside the cash comparison.
Use the full-cost budgeting guide to distinguish authorized spending, the latest spending forecast and actual commitments. A benefit forecast cannot supply missing permission to spend.
Explain the difference without inventing attribution
A variance tells you that the result differs from an expectation. It does not, by itself, explain why.
In Claire's case, the original estimate could have overstated the eligible report volume. Acceptance could have taken longer. The supplier's price could have changed independently. Each explanation needs evidence, and some of the difference may remain unresolved.
For the review, take these steps:
- Reconcile the basis. Check the period, report population, supplier charge and treatment of costs against the original case.
- Identify supported changes. Show which assumptions changed, the evidence and who confirmed them.
- Keep unexplained differences visible. Do not allocate the remainder to AI performance merely to make the explanation add up.
- Revise the next decision. Change scope, timing, support or the funding request where the new evidence warrants it.
In a July 2026 LinkedIn article, Cardaxia founder Jon Dobell argues for acknowledging attribution limits when AI arrives alongside process and organizational changes. This is a consultant's perspective, not an independent estimate of AI's financial effect. It is a useful challenge to a neat percentage that cannot survive questions about the other changes.
If a new template and the AI drafting service jointly reduce the external workload, retain both contributions. Use the guide to overlapping benefits and AI ROI before assigning separate returns to the initiatives.
End the update with a decision someone can make
Communicating uncertainty should help the board choose what to do next. It should not leave readers with a cautious paragraph followed by an unconditional spending request.
Explain the commitment being requested, the exposure if the benefit arrives late or fails, the evidence that would change the recommendation and the owner of the next review. The relevant review point depends on the work and commitment. It could be acceptance of another report type or the next supplier decision, rather than a universal number of months.
The decision may still favor further investment. A service improvement or newly possible work can matter even when no cash saving has been established. Define that outcome and how it will be evaluated instead of inventing a monetary benefit to complete the slide.
Keep the update connected to the wider evidence of AI value. Readers should leave knowing what is supported, what remains conditional and what management is asking them to authorize.
Questions about uncertainty in AI forecasts
What if the board wants one forecast number?
Provide a central estimate if the organization's planning process requires one, but keep its basis, plausible scenarios and decision-sensitive assumptions alongside it. State whether it is a management estimate or the output of a statistical method. A single planning figure does not remove uncertainty or authorize spending. Use the range discussion to explain which conditions would change the recommendation.
Are low and high scenarios a confidence interval?
Not automatically. Low and high scenarios describe outcomes under chosen assumptions. A probability-based interval needs an appropriate statistical method, data and interpretation. Ask the analyst what the range represents before attaching a probability. When no such basis exists, present the scenarios and their conditions plainly.
Does missing the forecast mean the AI investment failed?
A missed forecast is a reason to examine the evidence and next commitment. Check whether scope, timing, demand, costs or other changes explain the difference, and whether the intended quality and business outcome were achieved. Preserve unexplained variance rather than assigning a convenient cause. The reconciliation steps help turn the miss into a reviewable decision.
How should we report a capability we did not have before?
Describe the new work, its quality, delivery cost and intended use. A historic time-saving baseline may not exist. Agree how the capability will support a decision or service and what observations would justify continued funding. Keep that evidence separate from cash savings that require a spending change; do not invent a manual process or hypothetical salary saving solely to calculate ROI.
Can we use a vendor's savings estimate in the board pack?
You can identify it as a vendor estimate with its scope and assumptions. Check it against your own work, complete costs and the route to a business outcome before using it as a forecast. Do not present reported hours or an illustrative financial conversion as a realized saving. The evidence-status distinction makes the missing verification visible.
