By industry
AI & Automation for Healthcare
Automate the administration, not the clinical judgement
We design around
- DPDP Act 2023
- ABDM and health data standards
- HIPAA (US-facing work)
- NABH documentation requirements
- ISO 27001
- 60%
- Pre-authorisation cycle time reduction
- 3x
- Coder throughput with assistance
- 0
- Clinical decisions made autonomously
A deliberate boundary, not a limitation
What we hear
The problems that bring people to us
If several of these describe your week, there is almost certainly something worth automating.
- Claims and pre-authorisation cycles measured in weeks
- Medical coding backlogs delaying revenue recognition
- Discharge summaries and records transcribed by hand
- Patient communication and follow-up handled inconsistently
- Patient data spread across systems that do not talk
What we build
Where automation pays off in healthcare
Claims and pre-authorisation
Extract from submitted documents, validate against policy terms, assemble the case file and flag what needs clinical review. Compresses cycle time without removing a clinician from any medical determination.
Coding assistance
Suggest ICD and procedure codes from the clinical record with the supporting text cited, for a coder to confirm. Assistive by design — the coder stays accountable and works considerably faster.
Administrative documentation
Draft discharge summaries and referral letters from the structured record, for clinician review and sign-off. Returns clinical time without touching clinical content.
Patient communication
Appointment reminders, preparation instructions, follow-up scheduling and results-ready notifications, automated with escalation to a human for anything clinical.
Records integration
Bring HIS, LIS, PACS and billing systems into one modelled layer so reporting and automation stop depending on manual exports.
The line we do not cross
Clinical decisions stay with clinicians. We do not build diagnostic tools, treatment recommendations or triage systems that act without a clinician — those are regulated medical devices requiring clinical validation and regulatory approval, and that is a different business.
What we automate is the administrative machinery around care: claims, coding, documentation, scheduling, communication and reporting. That is where most of the recoverable cost sits, and it does not require anyone to accept an algorithm's medical opinion.
Why claims and pre-authorisation come first
Both are document-heavy, rule-bound, and currently slow in a way that costs money at both ends — delayed revenue for providers, poor experience for patients. Both have a verifiable correct answer against policy terms, so accuracy can be measured.
And critically, neither requires a clinical judgement to be automated. The clinical review step stays; what disappears is the fortnight of document handling around it.
DPDP and health data
Health data carries heightened obligations, and the erasure right is the one with architectural consequences. If patient information sits inside embeddings in a retrieval index with no lineage back to source records, you cannot honour a deletion request. Chunk-level lineage is the fix and it has to go in at design time — see AI Governance for how we document this.
Sequencing a first year
The order below is not arbitrary. Each step makes the next one cheaper, and the first one is chosen because it can be measured against records you already hold.
Weeks 1–2 — audit. Volumes, current cycle times, where the coders and the claims team actually lose hours, and which of it is tractable. Output is a sequenced backlog with modelled savings, not a strategy document.
Weeks 3–6 — one process, in production. Usually pre-authorisation for a single payer, or coding assistance for one specialty. Narrow enough to finish, real enough to measure. This is document processing work at its core.
Months 3–6 — widen, then integrate. Additional payers and specialties reuse the extraction and evaluation scaffolding, so each costs a fraction of the first. In parallel, the records layer that makes reporting stop depending on manual exports.
Months 6–12 — the administrative surround. Patient communication, follow-up scheduling and the internal reconciliations that sit in back-office automation.
The failure mode to watch for
Not inaccuracy — over-trust. A coding suggestion accepted without reading it is worse than no suggestion, because it launders a machine output into a human-signed record.
We design against this by showing the source text alongside every suggestion, tracking reviewer agreement rate as a monitored metric, and treating a rising acceptance rate with a falling review time as a warning sign rather than an efficiency gain. If your coders start rubber-stamping, the system is failing even while the dashboard improves.
FAQ
Questions we get asked
Will you build diagnostic AI?
No. Diagnostic and treatment tools are regulated medical devices with clinical validation and approval pathways that are a different business from ours. We automate the administrative burden around care, which is where most of the recoverable waste actually sits.
How do you handle patient data?
Minimise what moves, de-identify wherever the task allows, keep the authoritative record in-region, and log every access. Under the DPDP Act health data attracts heightened obligations, including erasure — which means retrieval indexes need lineage back to source records from the start.
Our systems are on-premise and old. Is this feasible?
Usually, though it changes the integration approach — database-level reads, HL7 or FHIR interfaces where they exist, file-based exchange where they do not. We are explicit about which parts become brittle and what fixing the integration surface would cost.
Can this reduce clinician documentation time?
Yes, and it is one of the highest-value applications. Drafting administrative documentation from the structured record, with the clinician reviewing and signing, returns real time. The clinician remains the author and the accountable party.
What does a first engagement cost and how long does it take?
A readiness audit runs two weeks and produces a sequenced backlog with modelled savings against your actual volumes. A first automation in production — typically pre-authorisation or coding assistance for one specialty — is a four-week sprint. We deliberately scope the first one narrow, because a working narrow system teaches you more about your own data than a six-month programme plan does.
Our HIS vendor says they are adding AI features. Should we wait?
Ask them three questions before deciding. What is the measured accuracy on your case mix rather than a demo, what happens to your data, and when is the actual ship date rather than the roadmap slide. If the answers are good, wait — genuinely. Vendor-native features that work are cheaper than anything we would build. In practice the honest answer is often that the feature is eighteen months away and generic across every customer, which is a poor fit for the specialty mix that makes your coding hard.
How do you measure accuracy on something like coding assistance?
Against a labelled set of your own historic records, coded by your own coders, held out from anything the system was tuned on. We report agreement rate per code family rather than one headline number, because accuracy on common codes tells you nothing about the long tail where the revenue leakage actually is. That set gets versioned with the code and rerun on every change.
Who is accountable if an automated administrative output is wrong?
The person who signed it, which is why every output routes through a named reviewer before it has effect. That is not a liability dodge — it is the reason the design works. A system that produces unreviewed administrative records creates an accountability gap that no hospital governance committee will accept, and rightly.
Talk to someone who has worked in healthcare
Bring a process that annoys you. In 30 minutes we will tell you whether AI helps, what it would cost, and where it would fail — even if the answer is don't bother.
Or email [email protected] · we reply within 1 business day