For the past few years, an AI product could attract attention by doing something that previously seemed difficult. It could produce a report, summarise a document, generate code or answer a question convincingly. The demonstration itself often carried the story.
At Web Summit 2026, the tone was different. During a panel with Cloudflare co-founder and CEO Matthew Prince, Lovable co-founder and CEO Anton Osika, and Recursive and You.com founder Richard Socher, speakers returned repeatedly to economic discipline. Prince argued that real demand for AI could coexist with questionable financial structures and overspending. Osika said the conversation was shifting towards what customers can actually measure after adopting the technology.
That is a useful signal for HealthTech founders. Hospitals have never been able to buy a demo and declare victory. Even so, the sector has seen plenty of products marketed through model performance, efficiency claims and conference-stage excitement without equally clear evidence that the full service improves.
The next commercial conversation should be more demanding. It should ask what happens after installation.
Accuracy is not return on investment
Suppose an AI system drafts discharge letters with impressive linguistic quality. The vendor reports that its model achieves a high score against reference text. That is interesting, but it does not answer the buyer's most important questions.
How much staff time does it actually save? How often must a clinician correct it? Does it create new review work? Does it introduce documentation errors? Does the average benefit hold for difficult cases? Does anyone have to spend additional time integrating it into existing electronic records?
A model can be technically excellent and commercially weak if staff cannot use it in their normal workflow. It can also save administrative time yet create problems elsewhere, such as additional patient queries or follow-up work. A serious evaluation has to consider both sides of the ledger.
There is no single ROI formula for all healthcare technology. A clinical diagnostic tool may need to show outcomes related to detection, decision quality and harm. A scheduling product may be judged on access, cancellations and utilisation. A documentation assistant may be evaluated using time, accuracy and clinician workload. Different intended uses create different evidence requirements.
The mistake is to choose the metric that makes the presentation look best rather than the one that matters to the healthcare organisation.
The hospital does not buy the model alone
When a founder calculates unit economics, it is easy to focus on software subscription revenue and the marginal cost of running AI. Hospitals pay for an entire change programme, whether or not it appears on the invoice.
Staff must be trained. Identity management and privacy checks must be completed. Interfaces need testing. People need to monitor outputs and resolve exceptions. Clinical and operational leaders must agree who can rely on the system. In some settings, procurement and regulatory assessment add significant work before the first patient-facing use.
Those costs do not necessarily make the product a bad purchase. They make it a purchase that must be evaluated honestly. If a vendor claims it saves 30 minutes per case but the organisation needs a dedicated reviewer to verify every output, the net saving could be small or negative.
A good commercial case therefore states the full deployment requirements and the conditions under which the benefit was measured. It distinguishes an optimised pilot from routine operation. It also explains how performance will be monitored after the sales team has left.
That is one reason the financial discipline discussed on stage matters to healthcare. The cost of intelligence may fall, but implementation is not free.
Which outcomes deserve priority?
For operational software, a helpful approach is to follow the patient journey rather than a vendor's dashboard. If a tool claims to improve hospital capacity, can the buyer see changes in actual wait time, cancelled procedures or bed turnaround? If a product claims to improve aftercare, are fewer patients missing appropriate follow-up, or are more of them simply receiving automated messages?
For clinical products, accuracy in a benchmark may form part of the evidence, but it does not settle how the tool performs with real users. Workflow changes can alter clinician behaviour. An alert that fires frequently may be ignored. A score that looks reassuring may shift the threshold for escalation. The clinical evaluation should examine what happens when the information enters human decision-making.
The World Health Organization's guidance on AI in health stresses reliability, appropriate governance and recognition of the risk of inaccurate or incomplete outputs. Those requirements are not separate from business value. They define whether the claimed value can be delivered safely.
A practical buyer might ask for three layers of evidence. First, does the system perform the intended technical task? Second, does it improve the decision or workflow when used by real people? Third, does the improvement justify total cost and risk at the organisation's scale?
Missing one layer weakens the commercial case.
A better way to build an ROI story
Founders should start with a specific operating problem and agree the baseline before deploying the product. Consider an oncology referral-coordination tool. The commercial case might track the share of referrals complete on arrival, staff time spent chasing missing documentation, time to clinical triage and the proportion of urgent cases meeting a defined target.
Next, define the comparison. Is the tool being measured against the existing process, a simpler process change or another software option? If the hospital simultaneously hires more staff, changes triage rules and introduces automation, the evaluation should not credit all improvements to the AI system.
Then expose failure modes. Does the product mishandle uncommon cases? Does it perform differently across clinics or patient groups? Is performance stable when demand increases? Does integration break whenever another vendor updates its system? These questions may be uncomfortable during sales, but they become more expensive if they are discovered after deployment.
Finally, make the economics legible. Show the implementation cost, ongoing operating cost, avoided work and relevant patient outcomes. Where uncertainty is large, say so. A range supported by evidence is more credible than a precise number assembled from optimistic assumptions.
This kind of commercial narrative may be less dramatic than “our AI saves thousands of hours,” but it gives procurement teams something they can defend.
The new advantage is evidence discipline
Prince's comments were directed at the broader AI market. He questioned the sustainability of deals that relied on financial engineering and argued that companies with stronger execution and cost discipline would be better placed to endure. Osika's contribution was closely related: customers increasingly want to understand the measurable value they receive.
For HealthTech, the implication is not that every startup needs a randomised trial before writing its first line of code. Evidence should be proportionate to intended use and risk. A low-risk administrative assistant and a diagnostic medical device do not require identical evaluation designs.
But both require clarity. What does the product claim? Who benefits? What evidence would falsify the claim? What will it cost to maintain the improvement after the first enthusiastic team moves on?
Startups able to answer those questions may find procurement conversations easier, even when their competitors have more impressive demos.
What investors should ask
Healthcare investors should not confuse strong adoption signals with established clinical value. A company may have many paid accounts because its product is pleasant to use. That is commercially meaningful, but it does not prove patient benefit. Conversely, excellent clinical evidence without a credible distribution and reimbursement pathway may not create a sustainable business.
The investable story connects evidence, workflow, buyer and economics. Who has the problem? Who can approve the purchase? Which budget pays for it? How quickly can the organisation implement? What data supports renewal? How does the company keep the value proposition intact as it expands to new sites?
Those questions were important before the current AI boom. The difference is that faster software creation and cheaper model access may bring more competitors into the market. Claims will be easier to make. Strong evidence will be harder to imitate.
At Web Summit, the broader AI industry sounded as though it was rediscovering the importance of return on investment. Healthcare has an opportunity to insist on a more complete definition: economic value that survives routine use, with patient safety and clinical quality firmly in view.
The next winning pitch may be the least theatrical one in the room. It will show precisely what changed after the technology was introduced, and why the buyer should believe the improvement will last.
Key Signals
The AI buying conversation has shifted from capability to proof: customers and investors now ask what measurable value a product produces, not what it can do in a demo. In healthcare that question has a clinical edge, because a tool can look efficient while shifting work, risk or cost onto clinicians and patients. The companies that survive this phase will be the ones that arrive with outcome data, not slideware.
