Value
Evidence That Changes an Investment Decision
When a survey, operating records and a business case disagree, identify which evidence could change the commitment—not merely strengthen its presentation.
Ronny BoesingPublished 8 min read
Share this perspective

An investment sponsor can receive three credible-looking documents and still lack a reason to sign. An external survey suggests that similar organisations are seeing returns. An internal operating record shows work being completed more quickly. Finance says the proposed commitment does not yet meet its purpose. Which document should win?
That is the wrong starting question. First ask whether the documents address the same decision. A survey can describe other organisations. An operating record can describe a limited local activity. A business case can express assumptions about a future commitment. They need not contradict one another—and none becomes decisive merely because it is the most polished.
The buyer's task is to identify what could change the choice now before them. Not to collect every available fact. Not to stage another demonstration. And not to describe a decision already made as if it were still open to evidence.
Start with the commitment, not the technology
Consider a hypothetical operations sponsor in a healthcare organisation. The sponsor is considering a multi-year purchase for an administrative document workflow across several sites. It does not make clinical decisions. A demonstration was persuasive, staff report a useful improvement in one setting, and an industry report is encouraging. Finance disputes whether the evidence justifies the full commitment. This is an invented scenario, not a customer case or a claim of measured performance.
The immediate question is not whether AI can be useful. It is whether to authorise this particular purchase, at this breadth and duration, for this purpose. Alternatives include retaining the current process, repairing it with existing tools, negotiating a narrower commitment if genuinely available, or declining the proposal. A limited purchase is not automatically feasible: the supplier may not offer it, and implementation work may still be difficult to reverse.
Write the choice as an action with a boundary. State what would be committed, which sites and work it covers, which future choices it closes, and the deadline that actually matters. Distinguish a real contractual or operational deadline from a date chosen to keep a presentation moving. If the boundary changes during discussion, the evidence must be reconsidered against the revised choice.
Do not turn different evidence into a vote
The Bessemer Venture Partners / Bain & Company 2026 Healthcare AI ROI Scorecard, introduced by Sofia Guerra and Steve Kraus on 30 September 2026, reports healthcare survey findings about adoption and returns. Its linked report identifies the survey source on its ROI presentation; another chart excludes “I don't know” responses and flags low respondent counts. Its respondent breakdown groups some small categories. Those details matter when interpreting the findings. They do not establish audited returns, causal effects, or what a particular buyer will achieve.
Use that report as external context, not as a forecast inserted into the hypothetical buyer's business case. No survey return multiple or percentage is transferred here. The published material reviewed for this draft does not establish representative sampling; raw responses were not reviewed. Healthcare context also cannot simply be extended to every industry.
In the scenario, the local operating record is closer to the proposed work, but closeness is not completeness. Ask what it records, what it leaves out, and whether the proposed sites and activities fall within its scope. A credible record from one setting can support a limited proposition without supporting a broad purchase. Finance's objection may concern that gap rather than dispute the record itself.
Do not count a supplier deck, a press summary and an industry presentation quoting the same survey as three independent confirmations. For each consequential proposition, trace the evidence back to its origin and state the question it can answer. A helpful distinction is between observations, interpretations and preferences: what happened, what someone thinks it means, and what the sponsor values. These are different contributions, not competing scores in a confidence contest.
Resolve disagreements about purpose before requesting more data
Suppose the sponsor wants a more dependable service while Finance is assessing a cost-reduction case. The first disagreement is about the objective. An additional timing measurement will not decide whether service quality deserves investment when cash savings remain uncertain.
Separate three kinds of disagreement. An evidence disagreement asks whether a record is reliable or applicable. An assumption disagreement asks whether an observed change supports the forecast attached to it. A value disagreement asks which outcome is worth pursuing and what trade-off is acceptable. More observation may help the first two. The third requires a decision by the person responsible for the objective; it cannot be delegated to a metric.
HM Treasury's Green Book offers a useful, bounded reference. Its 2026 guidance concerns appraisal for government objectives and public money, not private AI procurement. It says a single summary metric is insufficient for a balanced value-for-money judgement and that uncertainty and considerations that cannot be monetised also matter. That does not prescribe this buyer's financial hurdle or turn an unpriced benefit into a licence to spend. The editorial lesson is narrower: specify the objective before treating a financial calculation as the complete answer.
Ask for the observation that could change the choice
For each unresolved premise that matters to the current recommendation, complete this sentence: “If we learned ___, we would change from ___ to ___.” Name both the observation and the alternative. “We need more confidence” is not a usable request.
In the hypothetical case, the sponsor might prefer a full purchase only if evidence supports the intended work across the proposed settings. Evidence of an important unsupported setting could change the choice to a narrower purchase—if that option exists. Finance might support the service objective without relying on an unsubstantiated savings forecast. That would change the stated rationale, not prove the original forecast correct. If no affordable and timely observation can resolve the relevant premise, the sponsor must decide whether to narrow, decline or explicitly accept the remaining uncertainty.
The Green Book distinguishes sensitivity analysis, which tests changes in assumptions, from a switching value: the value at which a key variable changes whether an option represents value for money. Those are formal appraisal concepts. The questions above borrow the discipline of identifying what would alter a choice; they are not calculated switching values, a statistical test or a validated decision method. Do not invent probabilities or a numerical certainty score to make a qualitative judgement look calculated.
This question also exposes an important boundary. If nothing could change the recommendation, ask why. A mandatory requirement or an explicitly chosen strategic purpose may legitimately determine the action. Say so. If the rationale is merely that money has already been spent, distinguish that history from consequences still avoidable. The Green Book excludes sunk costs from forward appraisal while retaining consideration of alternative uses of resources already paid for. Future cancellation charges, remaining obligations and usable assets are not all erased by calling past spending “sunk”. This article does not interpret any actual contract.
Buy information only when it can arrive before the decision closes
An evidence request has its own commitment: staff effort, specialist cost, delay, and sometimes a dependency created while gathering it. Prefer the least burdensome credible way to answer the named question. Existing operating records, a clarification of a contract provision by the appropriate specialist, or a focused internal review may be enough. Another pilot is not the default answer.
Before commissioning work, identify who will obtain the information, what access is authorised, when it will be available, and what decision it could alter. Use existing authorised records where possible; do not collect identifiable patient or customer material for an illustrative business case. Agree an end to the enquiry. If the requested evidence will arrive only after the purchase is effectively irreversible, it may support later operating choices but cannot justify presenting today's purchase as conditional on its result.
The Green Book says appraisal effort should be proportionate to scale, cost, complexity and risk. It also discusses sequential decisions and flexibility, while warning that estimating scenario probabilities can introduce spurious accuracy. Here, proportionality is a management judgement, not a monetary value-of-information calculation. Check the practical burden of waiting alongside the burden of being wrong; do not assert that further research always pays for itself.
Keep narrowing, waiting and declining as real alternatives
Waiting is not free. A reversible improvement might be worth making despite incomplete evidence, especially if the current process is causing a known operational problem. Equally, a proposed “small first step” can create integration work or contractual dependence that makes the larger commitment difficult to avoid. Assess the action's actual reversibility, not its label.
Some constraints should exclude an option rather than become a low-weight item in an average score. A buyer should not trade away a necessary safety or legal boundary merely because another benefit looks large. Those boundaries require the appropriate responsible people and, where needed, specialist advice; this article does not determine them.
Strategic learning, service quality and rare but consequential risks need not be forced into invented cash equivalents. But an explicit willingness to invest for learning is different from claiming a proven financial return. State which choice is being made and what uncertainty remains. Retaining the current process can be defensible; it should not escape scrutiny simply because it is familiar. A short internal memo may be sufficient for a bounded, low-risk decision. There is no presumption that the buyer needs an external review.
Close with a decision record, not a claim of certainty
Before the sponsor signs or declines, use these gates as an original editorial checklist—not an audit, certification or empirically validated framework:
- Choice gate: Is there one stated commitment with feasible alternatives, a purpose and an actual deadline?
- Relevance gate: Does each decisive record address that choice, with its origin, scope and limits visible? Which disagreements concern objectives rather than missing facts?
- Change gate: Which observation could reverse or narrow the preferred choice? If none could, is the determining rationale stated honestly?
- Acquisition gate: Can that information be obtained with authorised access, proportionate effort and enough time to matter? What is the consequence of waiting?
- Commitment gate: What boundary can the available evidence justify, who makes the decision, and which unresolved disagreement or assumption is being accepted?
Record the selected action, the rejected alternative and why the difference matters. Attach the unresolved premise and the circumstance that would reopen the decision. Preserve a material objection even when the sponsor chooses to proceed; recording disagreement is not the same as requiring unanimity.
The result may be a narrower commitment, a justified full purchase, a short evidence request, or a refusal. The point is not to eliminate uncertainty. It is to avoid confusing a persuasive collection of material with evidence that addresses the commitment actually being made.
A proportionate next step
If the workflow problem itself is still unclear, start the free Workflow Fit Diagnostic. It can help frame the problem and a possible next step; it does not validate investment returns, audit evidence, review contracts or replace specialist advice. The free result may be enough.
Sources & perspective
The recommendations in this article are Boesing’s analysis. These publications provide context, not independent validation of business results. Reading an original source does not verify its claims.
- Sofia Guerra and Steve Kraus: The 2026 Healthcare AI ROI Scorecard — Bessemer Venture Partners / Bain & Company, 30 September 2026
Healthcare executive survey context · Original page and linked 35-page report reviewed 22 November 2026: pp. 9, 11 and 31. Survey-reported findings, not audited returns, causal effects or a buyer forecast. Item-level exclusions and grouped small categories are not universal denominators. Raw responses and representative sampling were not established.
Original publication reviewed; reported results are not independently verified.
- Healthcare AI ROI Scorecard 2026 — linked original BVP / Bain report
Original survey report; same study as the preceding source · Printed pp. 9, 11 and 31 checked directly. This is the same study, not independent corroboration; no return multiple or percentage is adopted as a buyer forecast. Introduced by the publisher on 30 September 2026.
Original publication reviewed; reported results are not independently verified.
- HM Treasury: The Green Book — 2026 edition, issued 5 February; publication page updated 28 October 2026
UK government public-money appraisal guidance · Page dates and current original PDF checked 22 November 2026. Paragraphs 1.1–1.8, 6.32, 6.73–6.75 and 6.95–6.103 support the attributed guidance. Not a private procurement obligation, investment advice or empirical validation of the author’s checklist. The original page’s 2013 date is not this edition’s date.
Original publication reviewed; reported results are not independently verified.
- The Green Book 2026 — current original HM Treasury PDF
Original normative guidance; same publication as the preceding source · Printed pp. 1–2, 37, 43 and 47–48 reviewed. Public-money scope, proportionality, sunk costs, sequential choices, false precision and switching values. No switching value or probability has been calculated for this article. This PDF is not another independent source.
Original publication reviewed; reported results are not independently verified.
Share this perspective
Reader discussion
Share a considered perspective. Every comment is reviewed by a moderator before it appears; nothing is published automatically.
Related insights

Value
From AI Pilot to Proven Business Value
Pilots are easy to start and hard to finish. The difference between an interesting experiment and real value is decided before the pilot begins.
Ronny BoesingPublicly available since at least 4 min read

Economics
The Hidden Cost of AI Automation
The licence fee is the visible part. The real cost of automation sits in integration, exceptions, oversight and the work of keeping it right.
Ronny BoesingPublicly available since at least 4 min read