Sovereign AI Glossary for Government Buyers

Committees stall on words. This glossary is the set we see on almost every Indian public-sector AI file. Statutory terms follow the Act. Engineering terms are defined the way an auditor will test them. Procurement terms are defined the way a file inspection will test them.

Attach it to the RFP as Annex A. Require bidders to use these meanings. Ambiguous vendor language then becomes a clarification item, not a scoring debate that lasts three meetings.

Statutory terms

• Personal data: data about an individual who is identifiable by or in relation to it. Format does not matter. A vector can qualify.
• Digital personal data: personal data in digital form, including data later digitised from paper.
• Data Principal: the individual the data is about — citizen, student, employee, vendor contact.
• Data Fiduciary: the person who decides the purpose and means of processing. Usually the department, not the vendor.
• Data Processor: a person who processes on behalf of a fiduciary. Platforms, SIs, model hosts, OCR and annotation vendors are usually processors.
• Consent: a free, specific, informed, unconditional and unambiguous indication of agreement. Withdrawal as easy as the yes.
• Legitimate use: a Section 7 purpose that can replace consent, including certain State functions and benefits. Not a general legitimate interests clause.
• Significant Data Fiduciary: a fiduciary or class designated by the Central Government under Section 10, with extra duties.
• Transfer: making personal data available outside India, including via an API call, a support session or a log shipper.
• Notice: the Section 5 information given to the Data Principal about who, what, why and how to complain.

Architecture terms

• Residency: the geographic location of stored data, including backups and disaster-recovery replicas.
• Localisation: a legal rule that restricts export or offshore storage of a data class. RBI payment data is the classic example.
• Sovereignty: operational and legal control, including who can compel access and who can turn the system off.
• Air-gap: no network path to systems outside the approved perimeter, including for updates and licence checks.
• On-prem: running on institution-controlled machines, which may still be networked. Not the same as air-gapped.
• India region: a cloud location. Not, by itself, sovereignty.
• Control plane: the brains that schedule jobs, users and billing. Its location is as important as the disk.
• Embedding: a numeric representation of content. Often still personal data.
• Fine-tune / adapter: extra weights specialised on your data. Treat as your artefact unless the contract says otherwise — and the contract should not.
• Telemetry: logs, traces, metrics and crash data that can include prompts and identifiers.
• Tool call: an action the agent takes in another system. Needs permission, a purpose tag and a log.
• Silent fallback: an undeclared second model or API used when the first one fails. Treat as a transfer until proven otherwise.

Procurement terms

• DPIIT recognised startup: eligibility and GFR relaxations. Not an architecture certificate.
• Make in India local content: a purchase-preference calculation. Separate from DPDP.
• DPA: the data processing agreement. Should list subprocessors, location, training bans and breach clocks.
• Material breach: the clause you want for silent egress or undeclared model fallbacks.
• Statement of work: where restore drills, key ceremonies and SIEM exports become countable deliverables.
• L1: the lowest-priced eligible bid. If your controls are not in eligibility, L1 will ignore them.

Phrases to reject unless defined

• Sovereign cloud — say which layers are where, and who holds keys.
• Data never leaves India — prove it for prompts, logs, support and backups.
• Anonymised — show there is no join key, no rare fact, no vector next to a user id.
• Enterprise grade — not a control.
• AI-enabled — say the workflow and the boundary.
• Compliant with all applicable laws — name the laws and the commencement dates.

How to annex this without a fight

Put the glossary in the RFP as Annex A and add one supremacy sentence: statutory terms take the meaning in the Digital Personal Data Protection Act, 2023 if they conflict with this annex. That sentence stops a vendor from redefining personal data in their MSA. It also stops your own team from arguing with the Act.

In the pre-bid meeting, walk through five words only: fiduciary, processor, transfer, residency, silent fallback. If bidders leave that meeting still using sovereign as a synonym for Indian, you have more work. If they leave using the five words correctly, scoring will be faster by weeks.

Translation and the file

If the evaluation committee works in a state language, translate the annex. Keep the statutory English in brackets so counsel can map back. A Hindi-only annex that loses Data Fiduciary will create two vocabularies. Two vocabularies create two arguments.

Words that belong in minutes

Minutes should record the defined term, not the slogan. Write the committee accepted residency in the SDC and rejected the bidder's sovereignty claim because the control plane is abroad. A later officer can act on that sentence. They cannot act on the committee was satisfied on sovereignty.

Get Annex A into one live file in ten days

Day 1: paste the glossary. Add the supremacy sentence. Translate if the committee needs it, with statutory English in brackets. Day 3: walk five words in the pre-bid or in an internal briefing: fiduciary, processor, transfer, residency, silent fallback.

Day 5: require bidders to recast any sentence that contains sovereign using Annex A terms. Day 7: reject a vendor glossary that redefines personal data or anonymised. Day 10: write minutes using the defined terms. The next officer should be able to score without asking what we meant.

Reuse the annex. Do not invent a new vocabulary per tender. Markets learn slowly. A stable annex is how they learn at all.

Objections you will hear — and what to do with them

Bidders will say a mandated glossary is unusual. Unusual is useful. Shared words are how you stop scoring poetry. If a bidder cannot work with defined terms, they cannot work with a DPA either.

Internal drafters will want to keep sovereign in eligibility because leadership likes it. Put it in the preamble. Eligibility needs words an evaluator can mark without a theological debate.

Translators will flatten Data Fiduciary into a generic owner. Do not let them. Keep the statutory English in brackets. A flattened term will collide with property-law owner in the same file.

Vendors will offer a joint glossary workshop. Workshops that produce a third vocabulary are delay. Offer a one-hour clarification on Annex A. Do not open the annex for redesign unless counsel finds a conflict with the Act.

Minutes writers will slip back into slogan language because it is shorter. Give them a cheat sheet of five terms. Review one set of minutes and send them back if the slogans returned. Minutes are the institutional memory. Memory that says sovereignty was discussed is not memory.

How this shows up in the file

Annex A only works if minutes use it. Review the next evaluation minutes and send them back if they say the committee was satisfied on sovereignty. Replace that sentence with the defined terms and the artefacts. That edit is the whole project.

Do not hold a workshop to redesign the annex. Hold one hour to walk five words. Workshops produce third vocabularies. Third vocabularies produce delay. Delay is how a chatbot with a foreign fallback wins on price.

Keep the annex stable across tenders. Markets learn slowly. A glossary that changes every file teaches them nothing except that you are negotiable on the meaning of personal data. You are not.

What the next noting must contain

“Sovereign AI Glossary for Government Buyers” belongs in a file, not only in a search result. A P2 Procurement should be able to point at one artefact that proves “sovereign AI glossary”: a packet capture, a processing schedule, a scored evaluation row, a dated notice, or a refusal rule. If the only evidence is a slide, you have a heading.

A shared vocabulary for the people who write notes, score bids and sign DPAs. Use these meanings in the next RFP instead of vendor English. DPDP 2023 does not define sovereign AI and does not write a blanket localisation rule for every model hop. CERT-In’s 28 April 2022 directions still set specified incident clocks and 180-day log retention in India for in-scope events. The November 2025 AI governance text is guidance, not a statute. A Proprietary Article Certificate, when it is lawful, lives in GFR Rule 166 — not Rule 161.

Write three dated sentences under C1 Sovereignty & Data Residency: what was decided, which designation owns it after the next posting order, and when it will be re-checked. Unsigned sentences are souvenirs. Dated sentences are controls.

• Name the designation that owns “sovereign AI glossary”, plus a deputy.
• Attach one artefact a stranger can open next year.
• Name the instrument you are actually using — Act, direction, GFR clause, GeM term, or guideline paragraph.
• Leave unsourced percentages, GMV slides and house forecasts out of the noting.
• Revisit when the model, the SI, the notice, the region or the posting changes.