Which assumptions hold up your AI strategy?

A hospital group likes an AI company’s demonstration. The clinical team sees a useful application. The company’s expansion plan treats this interest as evidence that a new market is opening. A budget is proposed for additional sales staff and a broader product. Several decisions have already been compressed into that short account: whether the problem matters enough, whether the right organisation will pay, and whether deployment can happen within the period the company can finance.

This is a constructed example. I would want those links visible before committing the expansion budget. A spreadsheet can calculate a coherent answer while the commercial explanation behind it remains incomplete. The useful question is which assumption makes the proposed allocation sensible, and what evidence would change it.

Give each link a subject and a verb

Aswath Damodaran’s June 2023 solo presentation on NVIDIA prompted this reflection on translating an AI opportunity into examinable assumptions. 1

Start with one sentence describing how the intended benefit becomes a viable business. In our example: a clinical department experiences a recurring problem, its staff adopt an application, the hospital can authorise and support its use, and a budget holder buys continued access at a price that sustains delivery. Every transition deserves its own explanation.

I would ask the team to replace broad nouns with observable actions. “Demand” becomes a named group choosing to address a defined problem within a stated period. “Adoption” becomes people repeatedly completing the intended work under agreed conditions. “Scalability” becomes the next customer receiving a service without requiring an entirely new delivery arrangement. These descriptions remain hypotheses until the relevant observations exist.

This makes disagreement more useful. A clinician can agree that the problem matters while an operations lead doubts that deployment will fit the available capacity. A finance colleague can accept the operational benefit while questioning where the purchase will sit in the budget. Their answers concern different links in the same proposal.

Keep benefit, purchasing and capture visible

In a healthcare strategy, I would write three separate entries: who receives the benefit, who controls the resources needed to obtain it, and how the company earns enough to continue providing it. Sometimes the same organisation appears in all three. The responsible people, budget periods and conditions can still differ.

Suppose the fictional application reduces interruptions for a clinical team. That is an outcome worth investigating. The expansion plan must also explain what the hospital can do with that improvement. It might improve working conditions, release usable capacity or support a different service arrangement. Each possibility needs its own evidence. A claim about minutes saved alone leaves the budget decision open.

The company must then explain its proposed price and the work required to deliver that benefit reliably. A useful product can face a purchasing process that moves more slowly than its funding plan allows. Enthusiastic users and an identified budget holder are valuable observations, but neither establishes the timing of a contract. This is where strategy connects to the separate calculation of the full cost of buying or building AI.

Mark how much each assumption has earned

On the NYU page introducing his book, Damodaran connects company narratives with valuation inputs and invites revision when events challenge the story. The resulting valuation reflects the author’s judgements. 2

For an operating decision, I would add a small evidence record to each important assumption: the observation, its date, the population or customer segment it concerns, and what remains inferred. A demonstration, a discussion with a prospective buyer, a signed agreement and repeated use answer different questions. They should retain their own descriptions when the strategy reaches the board.

A forecast may reasonably depend on assumptions. The record should make their status legible. If deployment timing comes from one unusually supportive hospital, label that origin. If willingness to pay comes from an interview, record who answered and whether they control the relevant resources. If a rate is an internal planning estimate, preserve that label when it enters a chart.

I would also record the strongest alternative explanation. Positive feedback might reflect courtesy, interest in research or enthusiasm for a feature outside the planned product. Asking which observation would separate these explanations helps determine what the next conversation or test must establish. It can also reveal that the available evidence is being used to answer a question it never addressed.

Build scenarios whose assumptions belong together

An optimistic forecast can quietly combine the easiest customer to reach, the fastest implementation and the highest price. Those conditions may describe different customers. A scenario needs a plausible account of how its assumptions coexist.

Imagine two routes for the fictional company. One large hospital group offers access to several sites, but expansion depends on a shared technical arrangement and a central decision. Independent hospitals could decide separately, while requiring more varied implementation work. Neither route is automatically preferable. Their dependencies differ, so I would keep the corresponding timings, delivery requirements and commercial terms together.

The same discipline applies to a downside scenario. Ask which event would move several assumptions at once. Losing an internal sponsor could affect access to users, purchasing progress and deployment scheduling. Treating those as unrelated risks can hide the concentration of the plan’s dependence on one relationship. Numerical probabilities would require a defensible basis; a clearly described dependency can already improve the decision without invented precision.

Give the next budget a question to answer

In four randomised trials from 2016 to 2019 involving 759 firms in Italy and the UK, Camuffo and colleagues found that scientific-approach training increased idea termination relative to other business training. These entrepreneurial samples provide no evidence of clinical effectiveness. 3

My practical inference is to connect the next resource commitment to a consequential uncertainty. For the fictional company, the issue might be whether the proposed customer segment has a workable purchasing route. Further product development would then need a separate justification. A bounded investigation with relevant budget holders could provide information that changes which expansion route deserves investment.

Before spending, write down the possible decisions after that investigation. Specify what evidence would support the current route, what would favour a narrower segment, and what would justify deferring expansion. Include the cost and time of obtaining the evidence. Some uncertainties cannot be resolved cheaply; identifying them can still change the size or reversibility of the commitment.

A disappointing result also needs interpretation. An unanswered invitation says little about a product’s value if the intended decision-maker never received it. Conversely, repeatedly redefining the target customer after every rejection can make the strategy impossible to challenge. Preserve the original question, examine whether the observation addressed it, and explain any revision. A formal pilot can then use a decision agreed in advance.

Return to the allocation

The resulting board note can be short: the proposed allocation, the causal chain behind it, the most consequential uncertain link, and the next observation that could alter the choice. It should name the person responsible for returning with that evidence and the date of the next decision.

As a physician and the founder of aiomics, I find this a useful way to connect a meaningful clinical problem with the obligations of building a company around it. The method proposed here has not been evaluated as a package. Its immediate purpose is practical: colleagues should be able to see exactly what they are being asked to believe before they commit people, time and capital.