India AI-Native Index — Methodology

Edition One · v2.3 · thecompounding.company · Published under CC BY-SA 4.0

1. What this is

The India AI-Native Index scores fifty well-known Indian companies on AI-nativeness using only publicly available evidence, applying the ANS-100 framework from The Compounding Company.

Scored companies never asked to be scored. The methodology therefore has to survive three attacks: this is just opinion, your data is wrong, and this is defamatory. Every design decision below exists to defeat one of those three.

The core discipline is simple. The Index distinguishes verifiable factual statements from evaluative scores. Scores and bands are opinions produced by applying the published rubric to the cited factual record, with a stated confidence level, after giving the company a chance to respond.

One sentence frames everything that follows: the Index measures what a company has publicly disclosed about its AI-nativeness. It does not measure company quality, business performance, or how much AI a company actually uses. Several companies in this edition delivered outstanding financial years and appear near the bottom, because they disclose almost nothing about AI. That is not a contradiction. It is the instrument working as designed.

2. The universe

Fifty companies, selected and frozen before any scoring began, on criteria published here so nobody can claim the list was built to engineer a story:

• NSE mainboard listed, or private with disclosed revenue above ~₹500 crore or 500+ employees
• Indian-headquartered or India-majority operations
• Seven sectors: IT services 10 · BFSI 9 · consumer internet 8 · manufacturing and mobility 8 · healthcare and pharma 6 · retail and consumer 6 · telecom, energy and infrastructure 3
• Operating-control exclusion: companies controlled by a foreign operating company that sets group strategy are excluded — Maruti Suzuki, Hindustan Unilever, Nestlé India, Oracle Financial Services. Financial-sponsor ownership does not exclude; a private-equity owner buys the strategy the company runs, so Hexaware and Mphasis are in.
• Private-company cap of four, of which two were used: Zoho and Zerodha.

One company was withdrawn after scoring, and the reason is published. L&T Technology Services was the universe's only ER&D-primary company. It was scored against IT-services anchors that its own evidence file records as ill-fitting for engineering-research disclosure norms, and — unlike every other sector — no ER&D calibration note was ever drafted or ratified. Scoring a single-company sub-sector under another sector's anchors is not defensible, so the entry was withdrawn rather than published. ER&D is a candidate sector for edition two.

Once published, the universe does not change mid-edition. Corporate actions after the freeze are footnoted, not acted upon.

3. The instrument

ANS-100 is thirty-five questions across seven weighted dimensions: Knowledge and context (15) · Agents and automation (15) · Model and infrastructure strategy (10) · Data, integration and AI security (15) · Organisation and operating model (10) · Economics and outcomes (17) · Business-model exposure (18).

The Full instrument is administered by interview. This Index uses ANS-100/P, the public-evidence form: each question is classified P (publicly observable), O (observable, able to reach the top score), or N (not publicly observable, unscored). P-questions cap at 3 of 4, because the top anchor requires internal evidence no outsider can verify.

This assessment is based exclusively on public information available during the stated evidence window. It does not assess undisclosed internal systems or practices. That line appears on every company file.

4. Scoring rules

• Evidence floor. Any score above 1 requires two independent items. A third-party assessment counts as one item of higher grade than self-declaration; two documents sharing a root source count as one.
• Recency. Evidence older than eighteen months is discounted unless continued in-window.
• Floor. Every scored question carries a minimum score of 1. A 1 is recorded in one of two circumstances, and the company file states which: the evidence pack met the completeness floor and no disclosure satisfying the level-2 anchor was located; or a single unsubstantiated claim or statement of intent was located. A 1 records what the public record contains. It is never a finding that the company does nothing. Questions that are not applicable or not scorable from public evidence are omitted from both numerator and denominator.
• The countable level-3 test. Level 3 on a P-question requires three distinct evidenced instances — not three documents about one instance — at least one in a revenue- or delivery-critical path, and at least one of third-party or primary-filing grade.
• The attribution test. A material economic improvement scores 3 only where a disclosed number links named AI systems to the reported outcome. Naming tools without quantifying, or quantifying without naming, is level 2.
• Pack completeness. Before any dimension is scored, the evidence pack must contain the full text of the latest annual report, four quarterly disclosures, earnings-call material, platform documentation, and a third-party assessment where one exists. "Likely underscored due to a pack gap" is not a permitted flag: fill the gap or leave the question unscored with the coverage consequence.

5. Sector calibration

Scoring the first non-software sector revealed that the public-evidence proxies were tuned to IT-services disclosure norms. Engineering firms publish patents and safety certifications, not AI revenue lines. Banks earn spreads, not billed hours. Steelmakers publish yield and downtime. Hospitals publish clinical outcomes.

So each sector received a published calibration note, ratified before its first company was scored — never after. The notes translate the anchors into that sector's evidence forms and, just as importantly, exclude categories that would award unearned credit:

• Automation is not AI — manufacturing and retail
• Digital channel share is not an automation rate — banking
• The algorithmic marketplace is not AI evidence — consumer internet
• GxP, NABH and JCI are not AI governance — healthcare
• Network automation is not necessarily AI — telecom

Anchors written after the scores they govern are worthless. Anchors written before are the whole point. All seven sector ruling sets, the anchor amendments and the v2.1 ratification and freeze record are published alongside this page.

6. Bands

Where a company's honest range crosses a band boundary, the placement carries a boundary marker and the range is stated. Companies are listed alphabetically within band; precise ordinal ranking is deferred to edition two.

7. Confidence grading

Every company receives a score and a data-coverage confidence grade, computed from how many of eight public-signal categories carried evidence: stated AI strategy and investment · engineering reality · talent shape · product evidence · operational adoption · data infrastructure · governance and risk posture · velocity.

The rule, stated so anyone can reproduce it: a category counts when the archived evidence pack holds at least one dated, attributable, AI-relevant item sourced from that category's channels. The grade is computed from the pack, not from the scores — otherwise the measure would be circular.

Grade A: six or more categories. Grade B: four or five. Grade C: three or fewer.

Grade C companies publish as a band, not a score, with the line: "Because public-evidence coverage was Grade C, no precise score or ordinal position is published." Nine companies publish this way.

The grade measures disclosure visibility, not company quality. In this edition, companies scoring in the forties grade C while others in the same range grade A. A Grade C is an invitation, not a verdict.

8. Ceilings — including the one that did not fire

Four ceilings cap the total regardless of everything else. Raw and capped scores are both reported, and every triggered ceiling is named.

Three findings are disclosed here rather than quietly retained.

The Exposure ceiling was predicted to do the editorial work in IT services. It never fired there. Discrimination came instead from the grain inside D7 and from dispersion in model-infrastructure and security disclosure. A pre-registered prediction that failed is published as such.

The economic ceilings are non-operative in edition one, and this is disclosed rather than quietly retained. The instrument's Outcome, Context and Exposure ceilings trigger when a dimension falls below 25%. Because edition one scores every question at a floor of 1, no dimension can fall below 25%, so none of the three ceilings can fire. They are designed for the full instrument, where absence can be scored at zero. An earlier draft of this page reported that the economic ceilings had fired at two companies after absence scoring was introduced late in the edition; that change was withdrawn before publication so that all fifty companies are scored on one rule, and the two affected companies were re-scored on the floor. Edition two will score absence at zero from the first company, with the rule published in advance.

The Security ceiling triggered ten times, and no score changed. Under the Index's public-evidence rubric, ten companies met the production-autonomy threshold while the reviewed public materials did not provide enough evidence to meet level 2 of the governance-disclosure measure. This is a finding about public disclosure, not a conclusion that the companies lack internal governance or controls. The ceiling did not alter any company's score. The triggers are concentrated in consumer internet.

9. Rater model — and what we are not claiming

Edition one is scored by a single rater against published anchors with a full audit trail. No inter-rater reliability claim is made. An inter-rater test was designed — blinded evidence packs, an eight-point tolerance per company — and deferred to edition two.

Instead, the material needed to check the work is published. The anchor set, the scoring rules and every sector calibration ruling are published under CC BY-SA, and the source archive register — every source relied on, with its archive reference — is available on request. Together these let any reader re-derive a scored company from the cited record and say where they would have scored differently.

What is not published, and why. Blinded evidence packs were designed for the deferred inter-rater test and do not exist for edition one: most sources were cited from the public record rather than assembled into per-company packs, so there was no pack to blind. Publishing a kit that does not exist would be worse than not offering one. Building blinded packs for a sample of companies is a candidate for edition two, alongside the inter-rater test itself.

Divergences are invited to the corrections contact and will be published.

10. Right of reply

Fourteen days before publication, every scored company received a one-page notice: its provisional score, the evidence list behind it, and an invitation to correct factual errors or supply public evidence that was missed.

Only public evidence can move a public score. Material shared confidentially is acknowledged and not scored.

Every notice sent and response received is logged. Silence means publication as assessed, noted neutrally as "company was invited to respond." Factual corrections are made and acknowledged by name in the corrections log.

11. Corrections — including our own

A standing corrections contact is published, every correction is logged, and the log is public.

Corrections made during edition one, all self-reported:

• A scoring inconsistency in the source framework was found and fixed during development.
• Three anchor ambiguities surfaced after the first three companies and were resolved by written amendment before scoring continued.
• Three entity rulings were corrected mid-sector after verifying corporate structure against annual reports: the correct listed entity following a demerger; an overlap where a parent's results include separately scored subsidiaries; and a company recorded as an associate that was in fact a consolidated subsidiary.
• Reading the source instrument after thirty-seven files had been scored revealed three errors: a wrong band set in working notes, a ceiling that had been recorded rather than applied, and a confidence system computed on the wrong basis. The last changed the published grades of twenty-two companies.
• A first-pass confidence recompute was itself wrong for eight companies and was corrected on verification against the archived packs.
• A retail evidence passage was misread from a multi-column extraction and corrected on re-reading, which sharpened a cross-sector finding rather than weakening it.
• Absence scoring (Batch-2 item B2-1) was withdrawn for edition one after review showed it had been applied to six of fifty companies and interacts incorrectly with the edition-one denominator rule. Six scores were revised upward; no company's confidence grade or publication form changed. The rule is pre-registered for edition two.
• Six company files carried arithmetic errors in a dimension percentage or total (range −0.7 to +1.8 points); all were corrected on reconciliation to a locked calculation sheet. No band or grade changed.

An index that publishes its own corrections is worth more than one that never finds any. These are listed because the audit trail is the product.

12. What edition one found

1. Disclosure breadth varies more by sector than by company size, performance or apparent capability. Every IT-services company received confidence Grade A; no manufacturing company did, and one of eight consumer-internet companies did.
2. Only seven of fifty companies connect a named AI system to a disclosed number in their own economics. Eighty-six percent cannot.
3. The floor is occupied by successful companies, not struggling ones. The lowest-placed companies all had strong commercial years. In several cases the placement rests on a complete reading of the annual report in which the specified AI terms did not occur in the machine-readable text; this establishes what the report discloses, not what the company does.
4. Across fourteen physical-first companies in retail and manufacturing, operational metrics are disclosed routinely and AI attribution for them is almost entirely absent. Companies publish the number or they name the AI, but rarely both — and where both appear, they are not connected.
5. The sector with the most consumer-facing autonomous action discloses the least about governing it. Seven of eight consumer-internet companies scored the minimum on AI governance; four trigger the Security ceiling.

13. Language, legal posture and licensing

Everything published describes evidence, never character: "in the public sources listed in the evidence annexure, reviewed through the stated cut-off, we did not locate disclosure sufficient to satisfy the Index's published criterion for AI governance; the company may maintain relevant practices that are not publicly disclosed" — never laggard, failing, or pretending. Every score-bearing claim carries its citation, and every floor score states its basis. Every source is archived at scoring date, because companies edit websites after publication. Archive references appear against each captured source on every company page; the complete source archive register is available on request via corrections@thecompounding.company.

Defamation in India carries civil and criminal exposure, so the design is conservative: only public, archived, cited evidence; opinion clearly framed as opinion applied through a published methodology; no pejorative labels; confidence grading that concedes incompleteness; a documented right of reply; a visible corrections policy. Claims about individuals, statements implying financial distress or dishonesty, use of non-public information, and publication without the notice period are prohibited. No leaked material is ever used — a leaked internal note about one scored company surfaced during research and was discarded for exactly that reason.

The original ANS-100 methodology and original editorial text — the rubric, weights, signal library, universe criteria and sector calibration notes — are published here under CC BY-SA 4.0. Company names, trademarks, quotations, source documents and other third-party material remain the property of their respective owners and are not licensed under that licence. The Index is not affiliated with, endorsed by, or sponsored by any assessed company. Nothing on this site is investment advice. The same open licensing that makes the framework adoptable makes the Index auditable, and auditable is what defensible looks like from the outside.