AI · 7 min read
Where AI genuinely saves work in an Indian hospital — and where it does not
AAKZEN TECHNOLOGIES ·
Most conversations about AI in healthcare are about diagnosis. Almost none of the practical value in a working Indian hospital is there. It is in the paperwork surrounding care, which consumes an enormous amount of clinical and administrative time and almost never gets attention because it is not interesting.
Here is a straight assessment of which is which.
Where it genuinely helps
**Turning dictation and notes into structured summaries.** A doctor describes a discharge in two minutes of speech or half a page of shorthand. Producing the formatted discharge summary from that is a mechanical transformation, and it is one of the clearest wins available. The doctor still reviews and signs. The typing disappears.
**Reading documents that arrive as images.** Insurance pre-authorisation forms, referral letters, outside lab reports, old case papers. These arrive as scans and photographs and get re-typed by hand. Extracting structured fields from them is exactly what this technology is good at, and the volume in a mid-sized hospital is significant.
**Making old records findable.** A room of physical files, or fifteen years of scanned PDFs, is functionally unsearchable. Digitising and indexing them turns "somebody go and look" into a query. This is unglamorous and it is often the single highest-value project available.
**Triaging queues.** Incoming enquiries, appointment requests, insurance correspondence — sorting by category and urgency so a person handles the ones that need judgement. It does not replace the person. It stops them spending the morning on sorting.
**Coding support for billing.** Suggesting procedure and diagnosis codes from the clinical note, for a human to confirm. Claim rejections from coding errors are a real and recurring cost.
The pattern in all of these: **high volume, low judgement, and a human confirming the output.** That is where the technology is reliable and where the saving is measurable.
Where it is the wrong tool
**Anything where being wrong is unacceptable and there is no reviewer.** If the output goes straight into a clinical decision without a clinician between, the case has to be extraordinarily strong. In hospital operations it rarely is.
**Problems a rule solves.** A great deal of what gets proposed as AI is a decision tree. "If the patient is over 60 and the test is X, flag it" does not need a model. It needs an `if` statement, which is cheaper, faster, auditable, and cannot surprise you.
**Problems a better form solves.** If data is arriving badly structured because the form that collects it is bad, fix the form. Using a model to clean up input you could have captured correctly is paying twice for one problem.
**Anything trained on data you do not have.** Predictive models need substantial, clean, historical data. Most hospitals do not have it in usable form, and the honest first project is usually the digitisation that would make such a thing possible in three years — not the prediction itself.
The test we apply: if a competent person with a checklist could do this task reliably given enough time, and the only problem is volume, it is a good candidate. If the task needs judgement that varies case by case, it is not.
The practical constraints nobody mentions
**Patient data cannot leave casually.** Sending clinical records to a third-party service has consequences you must decide deliberately, not discover. Some of this work can run on your own infrastructure; some cannot. That decision belongs at the start of the project, not the end.
**Language is not only English.** Notes and patient interactions in Hindi, or in a mix, are normal. Accuracy varies by language and it needs testing on your actual material, not a demonstration on clean English.
**Handwriting is still hard.** Printed and typed documents are largely a solved problem. Doctors' handwriting genuinely is not, and any vendor promising otherwise should be asked to demonstrate on your real files before anything is signed.
**Somebody has to check the output.** Every one of the useful cases above assumes a reviewer. If the plan removes the reviewer to save more money, the risk changes completely and so should the decision.
How to start
Not with a strategy. With one process, chosen because it is repetitive and somebody can name how many hours it takes.
Measure the hours before. Automate that one thing. Measure again. If it did not save what was expected, you have lost a small project rather than a budget, and you have learned something concrete about your own data.
The organisations that get value from this technology are almost always the ones that started with something boring and specific. The ones that start with a platform tend to end up with a platform.
