Module 8 · 95 min
AI Strategy, Portfolio & Economics
Where to Place the Bets, and Why. Opportunity and use-case prioritisation, portfolio choices, buy/build/partner decisions, vendor strategy and due diligence, total cost of ownership, ROI and value realisation, funding and operating-model choices. The central question: what should we pursue, and why?
- Frame AI strategy as a set of choices under scarcity — capital, leadership attention, integration capacity, clinical change capacity and data access — rather than as a list of technologies to adopt.
- Turn an AI idea into a strategic opportunity by attaching it to a service problem, a value hypothesis and a named route to realising the value.
- Compare candidate use cases on value potential, strategic fit and reuse, feasibility, evidence confidence and delivery burden, without collapsing them into a single misleading score.
- Use 'high value but not ready' as a deliberate portfolio position with a stated readiness gap, instead of a rejection or an indefinite pilot.
- Build a portfolio that balances near-term productivity, clinical value bets, enabling capabilities and longer-horizon options, and identify the shared capabilities that unlock several use cases at once.
- Name the full cost stack of an AI deployment — acquisition, integration, data preparation, evidence, redesign, training, oversight, infrastructure, compliance, monitoring and exit — and separate fixed from variable and one-off from recurring.
- Distinguish cashable from non-cashable benefit, and trace a claimed benefit through operational mechanism, measurable KPI, financial or clinical consequence and named owner.
- Recognise misalignment between who pays and who benefits, and say what that implies for funding and sequencing.
- Express an economic case as scenarios with stated assumptions and uncertainty rather than as a single-point return figure.
- Choose between buy, build and partner on capability, differentiation and maintenance appetite, and run a vendor diligence conversation covering lock-in, data rights, version control and exit.
- Make and defend fund, fund-conditionally, hold and stop decisions across a constrained portfolio.
The economics: total cost of ownership
The licence fee is the part of the cost that arrives with an invoice, which is why it dominates the conversation and why business cases built on it are optimistic in a predictable direction. Most of the cost of a healthcare AI deployment is internal effort, and internal effort is easy to leave out of a paper because no one raises a purchase order for it.
Total cost of ownership is simply the discipline of listing everything the deployment will consume over its life, including the parts paid for in clinical and technical time. The point is not accounting precision. It is to prevent the systematic error of comparing a fully costed internal option against a vendor's headline price.
The cost stack
Acquisition, subscription and vendor fees
Licence or per-user, per-encounter or per-call pricing, plus the professional services in the first contract. Check how the price moves with volume, users and future modules.
Implementation and integration
Interfaces to the record, single sign-on, workflow configuration, test environments and the internal engineering time to build and assure them. Usually the largest one-off internal cost.
Data preparation and mapping
Extraction, terminology mapping, quality remediation and the legal work to establish an access route. Cheap when a shared pipeline already exists; a project in itself when it does not.
Validation and evidence generation
Local validation, subgroup analysis, prospective evaluation and the analyst and clinical time to run and interpret them. Belongs in the cost of the investment, not treated as free research.
Clinical and operational redesign
Rewriting the pathway, agreeing who acts on what, updating standard operating procedures and absorbing the productivity dip while people learn. Real and routinely omitted.
Training, change and adoption
Initial training, refreshers, materials, super-users, and the ongoing cost of onboarding rotating staff. In a service with high turnover this is a permanent recurring line, not a launch cost.
Review and oversight burden
The clinician seconds spent checking each output, multiplied by volume. Ninety seconds of review on three hundred encounters a day is roughly a full-time equivalent of clinical attention — the most expensive resource in the stack.
Infrastructure, inference and usage
Compute, storage or per-call API charges where relevant. Variable with volume, and the line most likely to be underestimated when usage grows faster than the pilot assumed.
Security, privacy and compliance effort
Assessments, information-governance review, clinical safety documentation, penetration testing and periodic reassessment. Reassessment recurs on every material version change.
Monitoring, support and version change
Running the monitoring, triaging incidents, and re-validating after each model or product update. A vendor releasing a materially improved model is a cost event as well as a benefit.
Switching, exit and decommissioning
Exporting data and configuration, migrating or archiving records, retraining staff and running two systems in parallel during transition. Almost never in the original case, and it is what determines how free the organisation is later.
Three axes that change the answer
Fixed vs variable
Fixed costs are insensitive to volume: integration, initial validation, governance documentation. Variable costs scale with use: per-call inference, review time, per-encounter fees. A pilot dominated by fixed cost looks expensive per patient and gets cheaper with scale; one dominated by review time does not improve with scale at all.
One-off vs recurring
One-off costs are easier to fund and easier to over-weight. Recurring costs — subscription, oversight, monitoring, retraining, reassessment — determine whether the organisation can still afford the system in year four. Ask for a five-year run cost, not a first-year total.
Cash vs opportunity cost
The integration engineer working on this is not working on something else; the clinical lead's sponsorship is not available elsewhere. Opportunity cost never appears in a finance system and is frequently the real reason a portfolio underdelivers.
Fictional figures
Illustrative arithmetic with invented figures for a fictional 300-clinician deployment. The numbers demonstrate the shape of a cost stack; they are not benchmarks and should not be reused as estimates.
Worked example: a fictional documentation assistant, first two years
- Subscription, 300 clinicians × £600/year × 2 years£360,000Recurring, variable with users
- Integration and single sign-on (internal engineering, 90 days)£54,000One-off, fixed
- Local validation and note-quality audit (analyst + clinical time)£40,000One-off, fixed
- Training and onboarding, including rotating staff in year 2£70,000Part one-off, part recurring
- Information governance, clinical safety and reassessment after one version change£30,000Recurring on change
- Monitoring and support (0.4 FTE across two years)£64,000Recurring, fixed
- Clinician review time: 90 seconds × 120,000 encounters£150,000Recurring, variable with volume
- Two-year total of the lines above
- £768,000
- Share represented by the subscription
- ≈ 47%
- Share that is internal effort
- ≈ 53%
Two readings matter more than the total. First, the subscription is under half the cost, so a negotiation that wins ten per cent off the licence moves the case by less than five per cent — while a design change that halves review time moves it by ten.
Second, review time and subscription both scale with volume, so this deployment does not get dramatically cheaper per encounter as it grows. A deployment dominated by integration and validation would. That distinction should shape how ambitious the rollout plan is, and it is invisible if the costs are presented as one number.
Sources & evidence · 4 sources
This module cites public or consensus guidance, scholarly literature, technical documentation.
Content reviewed: September 2026. Publication dates of the individual sources are shown in each citation.
World Health Organization. Health technology assessment of medical devices, 2nd ed. Geneva: WHO; 30 May 2025. ISBN 9789240110878.
WHO's practical guidance on health technology assessment as a multidisciplinary evaluation of clinical, economic, organisational, ethical and social implications, intended to support efficient allocation of resources. It frames the discipline of assessing value before adoption; it is guidance on process, not evidence about any particular technology, and it does not prescribe the prioritisation categories used in this module.
Open sourceNational Institute for Health and Care Excellence. NICE HealthTech programme manual (PMG48): early-use HealthTech guidance assessments. Published 14 July 2025; updated 17 December 2025.
Describes how NICE assesses promising health technologies where evidence is still developing — weighing unmet need, potential value for money, the nature of remaining uncertainty and what further evidence generation is required, including workforce and system efficiency impacts. Directly relevant to staged commitment and to treating a weak evidence base as a reason to size a first investment as learning. It is a UK process manual, not a general economic method.
Open sourceNational Institute for Health and Care Excellence. Introducing CHEERS-AI: improving health economic evaluation reporting for AI technologies. NICE blog, 31 October 2024.
Introduces an AI-specific extension to health economic evaluation reporting, covering transparency about input data, model functionality, conflicts of interest and the assumptions behind claimed economic effects. A reporting standard: it improves how an economic case is described and audited, and following it does not make the underlying case correct.
Open sourceBinkley CE, Bouslov D, Zaidi A, et al. An early pipeline framework for assessing vendor AI solutions to support return on investment. npj Digit Med. 2025;8:368. doi:10.1038/s41746-025-01767-z
Proposes an early-stage framework for triaging vendor AI proposals around strategic alignment, shared accountability, measurable value and risk, with an impact/value case covering the problem, the solution landscape, quantifiable objectives and costs. Supports the case for structured early triage before procurement; it is a proposed framework rather than validated evidence that using it improves investment outcomes.
Open source
Capstone Board Pack
Add what you just learned to your own strategy document while it is fresh.