All insights

PSU & CPSE

Why CPSE AI Spend Fell 22% in FY26

· 10 minute read

The title promised a 22% fall. We will not invent that number. If you saw 22% in a title bank or a slide, treat it as unsourced until you rebuild it from GeM, DoE and CPSE reports.

The working title for this piece arrived from a title bank: Why CPSE AI Spend Fell 22% in FY26. It is a clean headline. It is also, as of 17 August 2026, a figure we cannot defend. We looked for a single official series that defines AI spend across Central Public Sector Enterprises for FY 2025–26 and reports a twenty-two percent decline. We did not find one we would put in a board note.

That is the first lesson of a data study in this market. Headlines travel faster than taxonomies. GeM categories change. Board notes bury software under digital transformation or consultancy. A pilot paid from a plant maintenance head will never appear in an AI dashboard. Anyone who quotes a national percentage without showing the query is selling a mood.

This article therefore does two honest things. It corrects the reader who came for twenty-two percent. And it gives a method you can run on papers you actually control: GeM extracts, DoE and DPE publications, CPSE annual reports, and your own board minutes. If a fall is real inside your enterprise, you will see it there. If it is not, you will stop repeating a number that was born in a spreadsheet of titles.

Prcept AI has an interest. We sell on-prem agents. We would love a story about cautious, quality-seeking CPSEs. We still will not mint a national percentage to make that story look like Reuters.

Correct the title before you quote it

If you arrived from a social card that said CPSE AI spend fell 22% in FY26, pause. Ask the author for the series, the definition of AI, the set of CPSEs, the treatment of amendments and cancellations, and whether FY26 is complete. If those answers are missing, the card is marketing.

We are the author of this page. Our answer is: we do not have that series. The 22% was a title-bank figure. It is not a finding of this study. Do not put it in a cabinet note, a bid, or a journalist briefing as if Prcept measured it.

You may still be seeing a real slowdown in your own company: stalled GeM bids, board deferrals after DPDP notification, L1 contests that collapsed, or pilots that never converted. Those are local facts. Local facts do not licence a national percentage.

What AI spend even means in a CPSE ledger

There is no single budget object in most CPSEs called Artificial Intelligence. Money lands under IT, consultancy, R&M, training, plant modernisation, cybersecurity, or a board-approved scheme with a patriotic name. A language-model subscription and a historian upgrade can sit three heads apart.

GeM will show you product categories and service titles as they existed on the day of the bid. Those labels are not a research taxonomy. A chatbot in 2024 and an agentic platform in 2026 may be the same work or completely different isolation stories. Keyword searches will both overcount (RPA, BI, ordinary IT) and undercount (silent line items).

FY26, if you mean 1 April 2025 to 31 March 2026, may still be getting restated in annual reports you will read later in 2026. Quoting a full-year national fall in August 2026 requires a source that has already closed that year for the whole CPSE set. Be sure that source exists before you brief.

A method you can run without inventing a census

  1. Define the entity set. Maharatna, Navratna, Miniratna, uncategorised — say which, and use DPE's current list, not memory.
  2. Define the year. Financial year, not calendar year. Write whether you include spillover payments for FY25 contracts.
  3. Define AI for this study only. Example: software, services or consultancy whose statement of work names model inference, RAG, or agent workflows. Exclude ordinary ERP unless the SOW is explicit.
  4. Pull GeM extracts you are allowed to pull. Record category, bid status, value, amendments, and whether the bid was cancelled. Do not treat bid value as spend.
  5. Read the CPSE's own annual report and board notes for digital, IT and consultancy. Tag line items with the same definition. Note where the report is silent.
  6. Check DoE and DPE dashboards and press notes for anything they actually published. Quote them as they stand. Do not splice them into 22%.
  7. Separate committed, spent, and slipped. A deferred board item is not a decline in spend until the year closes without it.
  8. Write the confidence. If half your entities have no usable line item, you do not have a percentage. You have a gap.

Reasons a real fall might appear — hypotheses, not findings

If your own extract shows a decline, do not jump to India does not want AI. Test these hypotheses against the file. They are explanations to check, not national conclusions.

  • Reclassification. Work moved from an AI label to cybersecurity, data platform, or consultancy.
  • L1 theatre that failed. Bids cancelled after isolation and exit could not be scored.
  • Board caution after DPDP notification and after early hosted pilots leaked traces.
  • Vigilance questions on haste, brand-lock and gifts of free POCs on live data.
  • Capex bunching in FY25 that was never going to repeat.
  • GeM category churn that makes year-on-year keyword charts lie.
  • Genuine substitution: fewer chat tenants, more unglamorous data cleaning, which some studies will not count as AI.

How founders should brief this without becoming the source of the 22%

If an investor asks why public-sector pipelines slipped, tell the local truth: longer board notes, more DPO questions, more insistence on air-gap and human gates. That is a market shape. It is not a Reuters print.

If a journalist asks you to confirm the 22%, say you cannot. Offer the method. Offer your own booked-versus-slipped numbers if you are willing to stand behind them. Do not confirm a title.

If a CPSE director asks whether everyone is cutting 22%, tell them to open their GeM and their board calendar. Peer folklore is how bad caps get set.

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.

Everyone is using 22%, we will look pedantic

Pedantry is the job when the number might enter a sanction. Looking careful is cheaper than retracting a briefing.

Approximate numbers are how strategy works

Approximate is allowed when the error bar is written. 22% with no series is not approximate. It is decorative.

You are talking down a real slowdown

We are talking down a fake precision. If your plant's AI indent sat for two quarters, say that. It is a better story than a national ghost.

Change the title so people are not misled

We kept the title because the title bank is already in circulation. The first duty is to correct the reader who arrived with it, not to pretend the phrase was never printed.

Four weeks to a number you can stand behind

  1. Week 1: write the entity set, the year, and the definition of AI on one page. Get finance and IT to initial it.
  2. Week 2: extract GeM bids and purchase orders you are allowed to see. Tag spent versus bid versus cancelled. Save the query.
  3. Week 3: read the last two annual reports and the live board calendar. Tag line items with the same definition. List silences.
  4. Week 4: produce a company-level table with confidence notes. If you cannot produce a national percentage, write that sentence in bold and stop.

How this shows up in the file

Subject: Treatment of unsourced 22% FY26 CPSE AI spend claims.

This organisation will not cite a twenty-two percent national decline in CPSE AI expenditure unless a dated official series is attached, with definitions of entity set, financial year and AI. Title-bank figures and vendor headlines are not series.

Internal decisions will use this company's extracts from GeM, the ledger and board notes. This note is an internal aid. It is not legal, investment or statistical advice.

What we will not do with this headline

We will not quietly change 22% to about a fifth in later paragraphs as if that were sourcing. We will not attribute the figure to industry. We will not let a salesperson use this URL as if it confirmed the fall. The URL is the correction.

If your company did spend less on labelled AI this year, write the local reason: a cancelled L1, a DPO hold, a board that wanted the three spines. Those sentences are better strategy than a ghost percentage. They also survive a director who can read.

This article is informational field guidance for Indian public institutions, not legal, procurement, security-accreditation, engineering, investment or statistical advice. Confirm against the live Gazette, GFR, GeM term, CVC instruction, CERT-In direction, UIDAI regulation, DPE guideline, DPDP text, departmental charter, plant safety manual and your counsel before you file it.

Questions this usually raises

Did CPSE AI spend fall 22% in FY26?
We do not have a verified official series that says so. The figure appeared in a title bank. Do not repeat it as a finding. Measure your own entity with a written definition.
Where would an official series come from?
Possibly GeM analytics, DoE, DPE compilations, or a research note that publishes its query. As of 17 August 2026 we are not treating any of those as having given us this percentage. Check again when you file; do not invent the check.
Is bid value on GeM the same as spend?
No. Bids are cancelled, amended, split and sometimes never awarded. Spend is what the ledger and the completed order say.
Why might my CPSE's AI indent have slipped even if 22% is false?
Board caution, DPDP, failed L1 contests, vigilance questions, category changes, or a one-time FY25 bunching. Read your minutes. Do not outsource the reason to a headline.
Will Prcept update this if a series appears?
If a dated official series is published, cite that series and retire the correction. Until then the honest sentence is that we do not know the national percentage.

Sources