10decodersDocument intelligence for hospitals in India: what it is actually worth
Six questions a hospital board should be able to answer before funding a document AI project. Written for the people who will own the outcome, not the people who will demo the tool.
The document problem, as the hospital actually experiences it
Most hospital leaders in India have already sat through a document AI demo. Someone drops a discharge summary onto a screen, fields appear in a table, everyone nods. The demo is never the problem. The problem is that nobody in the room can say what it will be worth, what it will cost to run in month nine, who in the building will operate it, or what happens when the Data Protection Board asks where the patient data went.
This article answers those questions in the order a board tends to ask them.
A mid-sized hospital's TPA desk works across a dozen or more insurer and administrator portals, each with its own login, forms, document formats and query process. Claims travel as scanned PDFs and courier packets. Status is discovered by phone call. When a pre-authorisation enhancement is pending, a bed stays occupied by a patient who is clinically ready to leave.
Meanwhile the medical records department holds the hospital's legal position. Indian courts have consistently treated the absence of a record, or the inability to produce it, as evidence against the hospital rather than as a neutral administrative gap. Retrieval speed is not a filing convenience. It is a defence.
Both of these are document problems before they are software problems. That is the case for looking at document intelligence, and it is also the reason a generic tool rarely survives contact with the building.
Where the return actually sits.
There are four places where document handling converts into money or risk in an Indian hospital. Every credible business case sits in at least two of them.
Discharge turnaround and bed-hours
Discharge delay caused by pending claim approval is the most visible cost because everyone in the hospital can see it. A bed held for four extra hours is a bed not turned. If your average occupancy is high, that is direct revenue. If it is low, it is patient experience and referral reputation. Either way, the constraint is usually the completeness and speed of the document packet leaving your building, which is the part you control.
Deductions and denials that trace to paperwork
Insurers handled about 3.26 crore health claims in FY24, settling roughly 82 percent of them. Claims worth around ₹15,100 crore were disallowed and a further ₹10,937 crore repudiated, which is a combined value of roughly ₹26,000 crore against a total claimed value of about ₹1.17 lakh crore. The average amount paid per claim was ₹31,086.
Not all of that is recoverable by better documentation. Coverage terms, waiting periods and exclusions account for a large share, and no extraction engine changes a policy exclusion. But the documentation-attributable portion is real, and it is the portion sitting inside your control. Published breakdowns of Indian denial reasons put form and documentation errors in the low single digits as a percentage of denials, with unanswered insurer queries substantially higher, and a large majority of rejection reasons knowable before the patient is even admitted.
The point is not that AI recovers denied claims. The point is that a query answered in six hours instead of six days, with the right attachment first time, is a different receivables profile.
Medical records retrieval and audit response
Consumer forum matters, medical council complaints, NABH audits and insurer fraud investigations all resolve into the same request: produce the record, fast, complete, with its chain intact. A searchable, indexed, structured record archive changes that from a two-week scramble into a query.
Readiness for structured claim submission
The National Health Claims Exchange went live in June 2024 as a standardised digital gateway routing claims between hospitals, insurers, administrators and government schemes under the Ayushman Bharat Digital Mission. An IRDAI panel has recommended making integration with it universal. As of the NHA dashboard reading in April 2026, 83 payers and 42,687 provider facilities were registered, with more than 23.4 million claims processed.
Here is the part most hospitals miss. NHCX standardises how the claim travels. It does not clean your data for you. If your billing is partly manual and your clinical documentation is unstructured, NHCX readiness is a data quality project wearing an integration costume. Document intelligence is one of the few practical ways to get from scanned reality to structured submission without hiring another twenty people.
Calculated worked example for 5,000 claims / month
Below is a worked example for a hospital submitting 5,000 claims a month. Every input is an assumption you should replace with your own numbers before it means anything. We publish it because vendors rarely show the arithmetic, not because these are your figures.
| Input | Assumption | Where it comes from |
|---|---|---|
| Claims per month | 5,000 | Your HMS |
| Pages per claim packet | 18 | Sample 50 packets from your TPA desk |
| Pages processed per month | 90,000 | Derived |
| Average claim value | ₹31,086 | IRDAI FY24 national average paid per claim |
| Share of denials attributable to documentation or query handling | Set your own | Your denial log, last two quarters |
| Bed-hours released per avoided discharge delay | Set your own | Your discharge TAT report |
We deliberately do not fill in the last two rows. Any vendor who fills them in for you before reading your denial log is selling a spreadsheet, not an outcome.
The honest framing for a board is this. Document intelligence does not create revenue. It shortens the distance between work you already did and money you were already owed, and it lowers the number of times that distance is walked by a person retyping a field.
Your hospital is a data fiduciary now
The Digital Personal Data Protection Act was passed in August 2023, but it sat without teeth until the DPDP Rules were notified on 13 November 2025. The rollout is phased. The Data Protection Board became operational immediately. Consent Manager registration opens around 13 November 2026. The substantive obligations, covering notice, consent, security safeguards, breach reporting and data principal rights, commence on 13 May 2027, with penalties running up to ₹250 crore.
A hospital is a data fiduciary under this framework. Large hospital chains and health-tech platforms can be designated Significant Data Fiduciaries, which adds a data protection officer based in India, an independent data auditor, and data protection impact assessments for high-risk processing.
Feeding patient records into a document AI system is high-risk processing. That is not a reason to avoid it. It is a reason to specify it properly. These are the questions to put to any vendor, including us:
Where does the document physically go?
In your VPC, in an Indian region, or out to a shared multi-tenant endpoint. Get the region and the tenancy in writing, not the marketing phrase.
Is anything retained after processing?
Ask for a zero-retention configuration and ask what the default is if you do not ask. The two answers are often different.
Is patient data used for model training?
This must be a contractual no, not a policy page that can change. Purpose limitation under DPDP means treatment data cannot quietly become training data.
What is redacted before the model sees it?
Name, ABHA number, phone, address and identifiers can often be stripped before extraction and rejoined after. If the vendor has never been asked this, that is informative.
What does the audit log capture?
Who accessed which record, for what purpose, when, and how long the log is kept. You will need this for both DPDP and NABH.
How is consent withdrawal honoured downstream?
If a patient withdraws consent, what happens to the extracted structured data, the cached document and the index. A vague answer here is a future notice.
One useful piece of design context: NHCX is a router, not a repository. Clinical and financial payloads move encrypted end to end and the exchange reads routing headers, not patient data. Your document intelligence layer should follow the same principle. It should be the thing that reads and structures, not the thing that quietly accumulates a second copy of your medical records estate.
What it costs to run, line by line
Per-page inference pricing is public and low. It is also the smallest number in the project. Here are current published rates for the commercial document AI services most Indian hospitals will encounter through their cloud provider.
| Capability | Published rate | When you use it |
|---|---|---|
| Plain text extraction (OCR) | ~$1.50 per 1,000 pages | You only need the words, and a language model does the interpretation |
| Layout and structure | ~$10 per 1,000 pages | Tables, checkboxes, reading order matter (most claim packets) |
| Prebuilt models | ~$10 per 1,000 pages | Invoices, IDs, health insurance cards |
| Document classifier | ~$3 per 1,000 pages | Routing a mixed packet into its parts |
| Custom extraction | ~$30 per 1,000 pages | Your specific TPA form, your specific discharge template |
| Language model reasoning | ~$1 to $2 per million input tokens at the low end | Validation, cross-checking, query drafting |
Cloud document AI rates are US East list prices as re-verified in July 2026 and will differ by region and commitment tier. Language model rates are entry-tier published list prices. Both change. Re-check before signing.
On the 90,000 pages a month from the worked example above, running classification plus layout plus selective custom extraction lands the inference bill in the low hundreds of dollars a month. That is not the number that kills projects. These are:
- Integration engineering. Reading from your HMS, writing back to it, and handling the fifteen edge cases your HMS vendor did not document. This is almost always the largest first-year line.
- The human review queue. During calibration, a person checks a meaningful share of extractions. Budget this explicitly. Projects fail when this cost is hidden, because the review then does not happen and errors ship.
- Reprocessing. Every schema change, every new TPA form, every prompt revision means re-running documents. Cheap per page, easy to forget in the model.
- Storage and retention. DPDP-aligned logging carries its own retention floor, and logs at claim volume are not trivial.
- The exception path. What happens to the 4 percent the system will not touch. If the answer is "the TPA desk, as before," you have not removed work, you have moved it.
A defensible budget line for a hospital at this volume is dominated by engineering and change management in year one, and by review-queue labour and licence in year two onwards. If a proposal shows inference cost as the headline number, it has not been costed.
Why use document intelligence when I already have Claude
This question comes up in almost every scoping call now, and it deserves a straight answer rather than a defensive one.
If you paste a discharge summary into a chat window and ask for the fields, you will very often get them, and get them well. Modern language models read Indian clinical documents, handwriting included, better than most people expect. So the instinct is reasonable.
The gap is not intelligence. The gap is everything around the intelligence. A chat window gives you a reasoning engine. A production document pipeline gives you seven things a chat window structurally cannot:
| What you need | What the chat window gives you |
|---|---|
| Ingestion at volume | A human uploading files one at a time. At 5,000 claims a month, that person is the system. |
| A schema contract | Prose that is usually right. Your HMS needs the same field names, same types, every time, or the write-back fails. |
| Confidence and routing | Nothing. The model does not tell you which of today's 300 extractions to double-check. |
| An evaluation set | Nothing. You cannot answer "did accuracy drop this week" without a labelled holdout you run against continuously. |
| An audit trail | A chat history in someone's account. Not a DPDP-grade access log tied to purpose of use. |
| Cost and latency control | Whatever the model does. No routing of easy pages to a cheap path and hard pages to an expensive one. |
| A consent and access boundary | Whatever the person pasting decides. Which is the actual compliance exposure. |
There is a failure mode worth naming, because it is the one that hurts hospitals specifically. Extraction systems, both classical and model-based, fail silently. A wrong policy number or a total borrowed from the wrong page comes back marked successful. In a chat workflow, a person catches that because they are reading it. In an unsupervised pipeline, it goes to the insurer. Confidence scoring and an exception queue exist precisely to convert silent failures into flagged ones. That machinery is what you are buying.
It is not document intelligence or a language model. A well-built document intelligence system today has a language model inside it. What you are choosing is whether that model sits inside a controlled pipeline with schemas, confidence thresholds, audit logs and a review queue, or whether it sits in a browser tab with a staff member and no record of what happened.
A useful test: if the person who currently does the pasting resigns tomorrow, does the process still run on Monday. If not, you have a habit, not a system.
How to train your team, and who to train
The most common reason these projects stall in Indian hospitals is not model accuracy. It is that nobody owned the exception queue in week three.
Treat this as a job redesign for four groups, not a software rollout for one.
The TPA and insurance desk
Their job shifts from assembling packets to reviewing assembled packets and working exceptions. That is a harder job, not an easier one, and it should be recognised as such in how the role is described and paid. The skill to build is judgement about what a low-confidence flag means, not tool clicks.
The medical records team
They become custodians of an index rather than a shelf. The training that matters is retention rules, access logging, and what to do when a record cannot be matched to a patient identity with confidence.
Coders and billing
They gain structured data at the point of claim assembly instead of retrospectively. The training is about correcting at source, because a correction made in the structured record is worth ten made in the claim.
IT and the HMS owner
They own the write-back, the log retention, and the vendor relationship. If your HMS is vendor-hosted, get them into the room in week one, not week nine.
A six-week arc that works
Weeks 1 to 2: Shadow mode
The system processes real documents; humans process them too; nothing goes to the insurer from the system. You are building the disagreement log.
Weeks 3 to 4: Calibration
Every disagreement is triaged into model error, document quality issue, or human error. That third bucket is usually larger than anyone expects and is itself a finding.
Week 5: Threshold setting
You decide, with names attached, what confidence level auto-proceeds and what routes to review. This is a clinical-governance style decision, not an IT one.
Week 6: Go Live
Live on one payer or one department. Not all of them.
Four things to measure from day one, weekly, on one page: straight-through rate, exception queue age, disagreement rate against the holdout set, and query turnaround time to insurers. If you cannot produce that page, you are not running the system, you are hoping.
One caution. Do not promise headcount reduction in the first year. It is rarely true, it makes the team you need for calibration your adversary, and it is the fastest way to get a quietly sabotaged pilot.
Where document intelligence is not the right answer
We would rather say this now than in month four.
Where to start if this is worth exploring
Pull fifty claim packets from the last quarter and count the pages. Pull your denial log for the same period and mark which rejections were about documentation and which were about coverage terms. Pull your discharge turnaround report and find the delays that were waiting on a pre-authorisation query. Those three numbers, from your own building, decide whether any of this is worth funding, and no vendor can produce them for you.
If you have those numbers and want a second read on what they mean, book a discovery call. We will go through them with you and tell you what we think, including if we think the answer is to do nothing yet.
Book a discovery call
Thirty minutes, with your own numbers on the table. Bring your denial log, a handful of claim packets and your discharge turnaround report, and we will work through what they say about whether document intelligence is worth funding at your volume.
