Writing the Innovation Narrative for DPIIT

A Kochi founder pasted the Series A memo into the Startup India innovation box. The memo said they would 'build the sovereign AI layer for Bharat' and 'learn from every government deployment'. Six months later a state RFE asked whether the bidder trains on customer data. The evaluator opened the public-facing recognition extract, then the bid. The sentences did not match. The clarification was ugly.

DPIIT recognition under the Startup India programme is a real procurement asset when a bid writes the relaxations. The recognition form asks you to describe the innovation. That description is not a slogan contest. It is a representation to a government department. Write it like a representation.

This is a template for Indian product firms, especially those selling on-prem or air-gapped agents. Prcept AI is DPIIT recognised, does not train on customer data, and would fail its own test if the recognition narrative said otherwise. Use the structure. Change the nouns to yours.

This is not official DPIIT guidance. The live form on startupindia.gov.in wins. Portal fields change. Keep a copy of what you submitted.

What the innovation box is for

It is for a reader who must decide whether the entity is working on a new product, process or service, or on a significantly improved one, rather than on a generic trading or services shop dressed as a startup. That is the policy idea. Your job is to make the improvement legible in one screen.

It is not for TAM slides. It is not for a list of advisors. It is not for a claim that you are the only Indian firm in a category — that claim will be false and will follow you. It is not for customer names you do not have permission to use.

It will be read later by people you did not imagine: a tender evaluator, a journalist, a competitor, a diligence associate. Write as if they will. Because they will.

A structure that fits a government form

Block	Job	Bad fill	Better fill
Problem	Name the Indian institutional job	'AI is transforming governance'	'District officers draft grievance replies from circulars they cannot search'
What exists	Name the boring baseline	'Legacy is broken'	'Keyword search on a shared drive and a Word template'
What you built	Nouns: retrieval, tools, isolation, eval	'A proprietary AGI stack'	'An on-prem agent runtime that retrieves from a sealed corpus and calls only listed tools'
What is new	The delta, not the adjective	'First ever sovereign'	'Media-fed weights, no outbound licence daemon, customer data excluded from training by product default'
What you do not do	Kill future contradictions	Silence, or 'we learn from deployments'	'We do not train foundation models on customer prompts or documents'

A template you can steal and then ruin with honesty

Paragraph 1, problem: '[Institution type] in India must [job] using [system of record], under [constraint: residency / air-gap / language / audit]. Today's baseline is [manual step]. Errors show up as [citizen or officer harm].'

Paragraph 2, product: '[Company] builds [product class] that runs [where: on-prem / SDC / air-gapped VLAN]. It retrieves from [corpus class], proposes [artefact], and calls only [tool class] under an officer's identity. Logs are kept [where] for [purpose].'

Paragraph 3, innovation: 'The improvement over a generic chatbot or a hosted API is [two or three concrete mechanisms]. Example mechanisms: offline entitlement, hashed media updates, tool allow-lists, evaluation packs for hallucination on circulars, Hindi / scheduled-language retrieval that cites the circular number.'

Paragraph 4, boundary: 'We do not [train on customer data / send prompts to a foreign host / claim to replace the signing officer]. Intellectual property in customer corpora remains with the customer. Recognition is sought for the product and process above, not for a trading activity.'

Stay inside the character limit the live form shows. If you must cut, cut adjectives first, then market size, then history. Never cut the boundary paragraph.

1. Draft in a doc, not in the portal.
2. Remove every superlative. Search for first, only, best, revolutionary, sovereign if you have not defined it.
3. Cross-check against your website, GeM store, and last three bids. One story.
4. Have counsel or a careful COO mark any sentence that could be read as a security certification or a customer endorsement.
5. Paste, submit, export or screenshot the final text. Store it in 01-identity of the bid folder.

Special problems for agentic AI

If you use third-party models, say that you integrate and harden them, unless you truly train your own. 'We built a model' when you wrap an API is a representation problem. 'We built a runtime, retrieval and tool layer around licensed weights, deployed on the customer's metal' is a product.

If you are open-weight, do not claim you invented the weights. Claim the eval, the isolation, the corpus tooling, the Hindi administrative fine-tuning if you actually did it on your data, not theirs.

If your earlier website said 'improves with every conversation', change the website before you file. DPIIT will not come to your office. A tender evaluator will come to your URL.

After the certificate arrives

Put the recognition number on bids only where the clause exists. Do not stamp it on every page like a quality mark. It is an eligibility paper for startup relaxations, not a technical score.

When you materially change the product — hosted to on-prem, or you start training on customer data (do not) — review whether the narrative is still true. Update through the portal's process if one exists. Keep the old text. Diligence likes history.

Do not let sales rewrite the innovation sentence on LinkedIn in a way the form contradicts. The form is the senior document.

Objections you will hear — and what to do with them

These are the lines that stall the file. Answer them in the room, then put the answer in the note. A spoken answer without paper will be forgotten by the next officer.

If we are modest we will be rejected as not innovative.

Modest and specific is how technical readers recognise innovation. Vague grandeur is how they recognise a services shop. DPIIT has seen both.

We should claim we train our own foundation model.

Only if you do. A false claim is worse than a wrapper. Wrappers that isolate and evaluate can be innovative products.

The form is a formality; consultants will write it.

Consultants will write adjectives. You will sign. Write it yourself. Let a consultant check completeness, not poetry.

We can have a bold website and a careful form.

Evaluators have browsers. One story.

A one-week narrative sprint

Do this before you touch the portal, and again whenever the product's boundary changes.

1. Day 1: dump every public sentence about the product into one doc — site, GeM, decks, old bids.
2. Day 2: write the four paragraphs from the template. No superlatives.
3. Day 3: red-team with a colleague playing a hostile evaluator. Fix contradictions.
4. Day 4: legal or COO pass on certifications you do not hold and customers you cannot name.
5. Day 5: submit or update. Archive the text. Align the website the same week.

How this shows up in the file

Store: submitted text, date, application number, recognition number, and a one-line owner. Add a sticky note: 'Any bid that contradicts this text needs a founder signature.'

If a bid needs more technical depth, that depth belongs in the technical annexure, not in a rewritten origin myth.

This article is informational field guidance for Indian public institutions and their vendors, not legal, tax, procurement or engineering advice. Confirm the live circular, RFE, GCC, GeM term, state G.O. and your counsel before you file anything. Incentives, ceilings and portal screens change.

How to run the route without confusing the letterhead

“Writing the Innovation Narrative for DPIIT” is a route problem. A P5 Startup/Vendor should know which legal person they are talking to — NIC, NICSI, a state IT corporation, iDEX, or a GeM buyer — and which paper that person can actually issue. Searching “DPIIT innovation description” is not the same as being on a panel that can receive a work order.

The DPIIT box is not a YC application and not an RFP. It is a short, true account of what is new in the product. Write it so a later tender cannot be used to impeach you. Empanelment letters are not purchase orders. DPIIT recognition is not a technical score. Reserved startup seats, if a notice writes them, are local to that notice. IndiaAI compute empanelment is not NICSI application-software empanelment.

Keep a warm evidence pack: CIN, GST, DPIIT, Udyam, financials, work-completion letters, architecture one-pager, DPA draft. Renewals are lost by people who treat the panel as a trophy.

• Screenshot the live RFE paragraph you are relying on, dated.
• Match the bidding entity name across every certificate.
• Do not mix iDEX, TDF, NICSI and GeM clocks on one tracker cell.
• Record the validity end date 90 days before it dies.

Close this loop before the next CAB

Put “Writing the Innovation Narrative for DPIIT” on the next change-advisory or bid-opening agenda as a single line item with an owner. If it cannot earn a line item, it will not earn a control. The owner should be a P5 Startup/Vendor, not “the vendor.”

Revisit the item when the model, the GeM term, the region, or the SI changes. “DPIIT innovation description” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.