Field notes on what's next.

Later than most teams do, and in fewer places. The decisions genuinely expensive to revisit are the data model and the tenancy model; nearly everything else can be changed once you know what the load looks like. Building for a scale you have not reached spends runway on optionality you may never use.

Usually the database and whatever humans were quietly doing by hand. The technical ceiling is visible and gets attention; the operational one does not, until a process that worked at fifty customers collapses at five hundred. We look at both when we assess a platform.

That you can explain any balance, prove who changed what, and lose a component without losing correctness. Encryption and access control are necessary and rarely the thing that fails. What fails is a system that cannot reconstruct how it arrived at a number, which is both a security problem and an audit one.

More than most teams plan for, and earlier. Data residency, retention and auditability are architectural constraints, not configuration, and retrofitting them is where fintech programmes lose quarters. We bring the compliance owner into design rather than presenting to them at the end.

The search step is moving away from the supplier's own site. Travellers increasingly arrive having already narrowed their choice somewhere else, which means the booking path has to be fast, trustworthy and capable of holding inventory accurately at the moment they commit. The competitive ground is shifting from discovery to execution.

Because availability is distributed across suppliers who each update on their own schedule, and the traveller experiences any disagreement as a failed booking. The platforms that handle this well treat supplier responses as untrusted input, normalise aggressively at the edge, and design what happens when a supplier is slow rather than assuming it will not be.

In documentation, triage support and the administrative traffic around care rather than in diagnosis. Ambient note-taking and discharge summarisation are the two with the clearest return, because they attack work clinicians never wanted to do and can verify in seconds. Anything that proposes a clinical decision carries a much heavier evidence and governance burden, and should be treated as a separate class of project.

A data estate you can trust and a named clinical owner. Most stalled projects we see failed on the first: the model was fine, but the records feeding it disagreed across departments. The second matters because an AI tool without a clinician who owns the workflow becomes a pilot nobody adopts.