Is your hospital ready for AI? A readiness checklist for Indian hospitals

Every hospital is being offered AI now: scribes that write the consult note, tools that summarise a patient’s history, agents that watch the queue or the pharmacy shelf. Some hospitals will get real value from them. Others will buy a tool, run a pilot, and find that it does not work on their records.

The difference is rarely the AI. It is whether the hospital is ready for it. This article sets out what readiness means in practice, in six areas, with a checklist at the end. It is written for hospital owners, medical directors and IT heads in India.

The short version

• Records first. AI reads records. If yours are on paper or in scanned PDFs, start there.
• Trustworthy data. Duplicates, gaps and wrong codes become wrong answers.
• Standards. Coded diagnoses and tests, and a way to exchange data, make every later step cheaper.
• Security and consent. Patient data needs access control, logs and a lawful basis, before any AI touches it.
• Governance. Someone must check that a tool works on your patients, and keep checking.
• People. Pick one problem, measure it, and let the staff who live with it shape the fix.

1. Records that are digital, structured and in one place

AI works on data. A handwritten case sheet scanned into a PDF is an image: a person can read it, but a system cannot reliably pull the blood pressure, the diagnosis or the drug dose out of it. Readiness starts with records captured as data at the point of care: vitals as numbers, diagnoses as coded entries, prescriptions as lines with a drug, a dose and a frequency.

Three things to check:

• One patient, one record. Every patient should have one identifier inside the hospital (an MRN or UID), and visits, results and bills should hang off it. Two records for the same person split their history and halve what any tool can see.
• Everything in one system, or connected. Out-patient notes in one system, lab results in another and pharmacy on paper means no tool sees the whole patient.
• Linked to ABHA where the patient wants it. Ayushman Bharat Digital Mission links records to a patient’s ABHA with consent, which also makes records from elsewhere reachable.

2. Data you can trust

A model trained or run on bad data gives confident wrong answers. Researchers who study electronic health record data describe its quality in a handful of dimensions. One widely used review grouped them into five: whether the data is complete, correct, concordant with other sources, plausible, and current.

In a hospital this shows up as missing allergies, a weight entered in pounds in a kilogram field, a decimal slip in a lens power, the same patient registered twice under two spellings, or a diagnosis coded loosely because the right code took too long to find.

We cover how to measure and fix this in Clinical data quality: why AI starts with clean records. The short answer is to fix it at the point of entry, not later in a spreadsheet.

3. Coded and standard

Standards are what let data move and be compared:

• ICD-10 for diagnoses, so “type 2 diabetes” is one thing however it was typed.
• LOINC for lab tests, so a haemoglobin from your lab and one from an outside lab can sit on the same chart.
• FHIR R4 for exchanging clinical data, and the format ABDM uses.
• ABDM itself, for consent-based sharing with the patient’s other providers.

A hospital whose records are coded to these standards can adopt a new tool far faster, because the tool does not need a custom mapping for every field.

4. Infrastructure and security

Before any AI system reads patient data, the basics have to be in place:

• devices at the point of care, so data is captured during the visit rather than typed up later;
• reliable connectivity, and a plan for when it fails;
• access by role, so each person sees what their job needs;
• a log of who opened and changed what;
• a lawful basis and consent for how patient data is used, under the Digital Personal Data Protection Act, 2023.

Ask every AI vendor where the data goes, whether it leaves India, whether it is used to train their models, and how you would know if it was misused.

5. Governance: who decides, and who checks

The most common way clinical AI fails is quietly: it works where it was built and not where it is used. A widely cited example is a sepsis prediction model used in many US hospitals. When researchers at Michigan Medicine checked it against 38,455 hospital stays, it missed 67% of sepsis cases while alerting on 18% of all patients, with an AUC of 0.63.

The lesson for any hospital is to validate a tool on your own patients before relying on it, and to keep measuring after it goes live. In practice that means:

• a named person or small committee who approves AI tools and owns their review;
• a written purpose for each tool, and what it must never do;
• a check on local data before go-live, and regular checks after;
• a person who reviews AI output before it reaches a patient record, for anything clinical;
• a way for staff to report a wrong answer, and a record of what was done about it.

Several frameworks can guide this. ICMR’s 2023 Ethical Guidelines for Application of Artificial Intelligence in Biomedical Research and Healthcare are the Indian reference point. WHO’s 2021 guidance sets out six principles, from protecting human autonomy to ensuring inclusiveness and equity. NITI Aayog’s Responsible AI papers (2021) set out seven principles for India. For a formal management system, ISO/IEC 42001:2023 is the first international standard for managing AI in an organisation.

6. People and workflow

AI changes how work is done, so start with the work. Pick one problem the staff already feel, measure it today, try a tool on it, and measure again.

Documentation is the usual first choice, and the evidence is worth knowing. A randomised trial at UCLA published in NEJM AI in 2025 found that one ambient AI scribe cut time spent on notes by 9.5% compared with usual care, while another made no significant difference. A 2026 study in JAMA of 8,581 clinicians found that scribe users spent about 16 fewer minutes on documentation per eight scheduled hours. The authors called the effect modest. Useful, then, but not magic, and worth measuring in your own out-patient department, where volumes and consult lengths are very different from a US clinic.

Where the NABH Digital Health Standards fit

NABH’s Digital Health Standards are the closest thing India has to a digital maturity yardstick for hospitals. The first edition for hospitals was published in September 2023. It covers eight chapters, three clinical and five non-clinical:

1. Access, assessment and continuity of care
2. Care of patients
3. Management of medication
4. Digital infrastructure
5. Digital operations management
6. Finance and procurement management
7. Human resource management
8. Information management system

Any hospital with in-patient or day-care services can apply, whether or not it is NABH accredited. Accreditation is offered at maturity levels from Core to Excellence, and recognition lasts two years. A second edition was released in draft in September 2025. NABH also runs a separate standard for HIS and EMR software, which requires the software to hold ABDM compliance certification.

You do not need to apply to use the standards. They are a well-structured checklist for the work this article describes.

An AI readiness checklist

1. Are out-patient notes, prescriptions and lab results captured as structured data, not paper or scans?
2. Does every patient have one hospital identifier, and do you check for duplicates at registration?
3. Are diagnoses coded in ICD-10 and lab tests mapped to LOINC?
4. Can your systems exchange data in FHIR, and link records to ABHA with consent?
5. Do you know how complete your key fields are (allergies, vitals, diagnoses)?
6. Does each member of staff have their own login, with access by role, and is every access logged?
7. Do you have a lawful basis and consent process for patient data under the DPDP Act?
8. Is there a named person or committee who approves AI tools and owns their review?
9. Do you validate a tool on your own patients before go-live, and measure it after?
10. Does a person review AI output before it reaches a patient’s record?
11. Have you picked one problem, and measured how it performs today?
12. Do your staff know how to report an AI error, and does someone act on it?

Eight or more yeses and you are ready to pilot. Fewer, and the items you missed are the project.

How 92health works with hospitals on this

92health is an EMR and AI platform used by clinics, hospitals, pharmacies, labs and optical shops in India. Much of the readiness work above is built into how it records care: one patient record per hospital, ICD-10 coded diagnoses, lab results validated as FHIR R4, records linkable to ABHA, access by role with every change logged, and patient data that stays in India.

Its AI features follow the same rule throughout: they draft, read or answer, and a person checks the result before it counts.

For hospitals that want a structured path, we run AI readiness work in three steps:

1. Assess. We go through the six areas with your team and look at a sample of your records.
2. Plan. You get a written report: where you stand, what to fix first, and which AI use would pay off first.
3. Build. We help you put the foundations in place and run the first pilot, with a baseline measured before it starts.

Sources

• Weiskopf NG, Weng C. Methods and dimensions of electronic health record data quality assessment. JAMIA, 2013. doi:10.1136/amiajnl-2011-000681
• Wong A et al. External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Internal Medicine, 2021. doi:10.1001/jamainternmed.2021.2626
• Lukac PJ et al. Ambient AI scribes in clinical practice: a randomized trial. NEJM AI, 2025. doi:10.1056/AIoa2501000
• Rotenstein LS et al. JAMA, 2026;335(16):1408-1417.
• NABH. Digital Health Standards, and Digital Health Standards for HIS/EMR systems.
• WHO. Ethics and governance of artificial intelligence for health, 2021.
• NITI Aayog. Responsible AI #AIforAll, Part 1: Principles for Responsible AI, 2021.
• ISO/IEC 42001:2023, Artificial intelligence management system. IEC webstore.