Every agency deck you have seen in the last two years has said "AI-powered." You cannot verify a single one of them, and you suspect it means they are using the same tools you could open yourself, while charging you more for the privilege.
That suspicion is largely correct, and this article is written on the assumption that you hold it. We are an agency that describes itself as AI-powered, so treat everything below as a claim requiring evidence — including ours. What follows is the evidence, a test you can run on any agency including ours, and the part of this story that nobody in our category is telling you: producing marketing assets with generative tools now carries disclosure obligations in India, and the ownership position on what you receive is unsettled.
What does AI actually change in an agency, according to evidence?
The best available study is a pre-registered randomised field experiment, not a vendor survey. Dell'Acqua and colleagues at Harvard Business School, working with Boston Consulting Group, ran 758 BCG consultants through three conditions — no AI, GPT-4, and GPT-4 with prompt-engineering guidance — across two tasks, one chosen to sit inside the model's competence and one deliberately outside it. The paper is public and dated September 2023.
Inside the model's competence, the effect was large:
- 12.2% more tasks completed
- 22.5% to 27.6% faster
- Human-graded quality up more than 40%
Outside it, the effect reversed hard. The control group answered correctly 84.5% of the time. The AI-assisted groups were 19 percentage points less likely to be correct on average.
Read that number twice, because it is the whole argument. The same tool that made people meaningfully better at one task made them substantially worse at another, and the people doing the work could not reliably tell which was which. The authors call this the jagged frontier: capability that is uneven in ways you cannot see from the outside.
There is a detail in the results that almost nobody quotes and that we find the most instructive. The group given prompt-engineering guidance did worse outside the frontier than the group given none — 24 percentage points below control, against 13. Greater fluency with the tool produced greater misplaced confidence in it.
So the honest version of the category claim is narrow: AI changes throughput on structured production. It does not change judgement. Anyone telling you it improves strategy is describing the condition where correctness fell by nineteen points.
What is the Delegation Map?
If capability is jagged, then the only thing that matters about an agency's AI use is where it has drawn the line — what it hands to machines and what it refuses to.
That map is specific, and it is auditable. On our side of it, the work that genuinely industrialises is:
- Creative variant production. Many executions of a tested concept, at a volume that makes real testing possible rather than theatrical.
- Content drafting and research synthesis. First drafts and source gathering, reviewed and corrected by whoever will put their name on the output.
- Performance analysis and reporting. Pattern-finding across accounts that a human then interprets.
- Conversational enquiry qualification. Routing and first-pass qualification, with defined escalation.
And the work that does not, because the evidence says delegating it degrades the output:
- Positioning, and the decision about what the business is actually selling
- The offer and the price
- Brand decisions with multi-year consequences
- Which channel to abandon
- Anything where being wrong is expensive and the error is hard to spot
Here is the one question worth asking any agency, including us: what do you refuse to use AI for? An agency that cannot answer that in specifics has not drawn a map. It has bought tool subscriptions and written a new line in its deck. The answer tells you more than any case study, because it is the only part of the pitch that costs the agency something to say.
If production costs less, why isn't the invoice smaller?
This is the sharpest objection and it deserves a direct answer rather than a deflection.
The honest position is that the saving should go somewhere you can see. We think it should show up as volume of tested work rather than a smaller fee, and that is a defensible position only if you can count it. So count it:
- How many distinct creative variants were produced this month, and how many were actually run rather than filed?
- How many tests reached a decision, and how long did each take?
- How many pieces of content were published, and what share needed rewriting before they went out?
If those numbers are not materially higher than the agency's pre-AI baseline — or than a conventional competitor's — then the efficiency is being retained upstream and you are funding someone's margin. That is a legitimate thing to discover, and it is why we would rather you measured it. The general principle, which applies to every line of a retainer, is in how to actually measure marketing, and our pricing logic is set out in what agencies do not tell you.
There is a counter-argument we take seriously: more variants is not self-evidently better. Spread the same budget across forty creatives instead of eight and each one gets thinner data, significance takes longer, and you may learn more slowly while feeling more productive. Volume is only an advantage where it is gated by human selection before it reaches a platform. Unfiltered volume is not throughput; it is noise with a higher invoice.
Does Google penalise AI-generated content?
Not for being AI-generated. For being made at scale to game rankings.
Google's position, set out when it replaced its automatically-generated content policy in March 2024, is a scaled content abuse policy: producing many pages primarily to manipulate rankings rather than help users. The policy applies whether the content was made by automation, by humans, or by both. Method is irrelevant; purpose and value are what is assessed.
This matters commercially in a way most agencies gloss over. It means the throughput advantage does not transfer to content volume. An agency that uses AI to publish two hundred thin pages against programmatic keyword permutations is not exploiting an edge — it is walking into a documented policy. The throughput advantage transfers to creative variants in paid media, where volume genuinely helps, and to drafting speed on content that still has to clear a quality bar. Those are different things, and an agency that conflates them will cost you an organic footprint. That distinction runs through what SEO actually is in 2026.
What do you now have to disclose on AI-made creative in India?
This is the section no competitor's article contains, and it is the reason to read this one.
Google now expects advertisers to declare generative AI in creative. In July 2026 Google introduced a "How this ad was made" panel in My Ad Center across Search, YouTube and Discover, showing whether an ad was created or edited with AI. Google's own announcement states that advertisers must use the provided tools to label ads made with generative AI, with a control for declaring AI used outside Google's own tools. Ads built with Google's generative advertising tools are disclosed automatically. Depending on local requirements, a label may appear on the ad itself and not only in the transparency panel.
Note what that is not: it is not limited to election advertising. The 2023 election-ads rule still exists separately, but this is a general expectation across advertiser creative. If an agency is running AI-produced creative for you and has not mentioned the declaration, that is a gap in your account, not a technicality.
ASCI's labelling guidelines bind you, the advertiser. The Advertising Standards Council of India released its final Guidelines for Responsible Labelling of Synthetically Generated Content in Advertising on 29 September 2026, following a May 2026 draft, and they take effect 29 December 2026. Khaitan & Co's analysis sets out the structure; the guidelines themselves are published.
The test is deliberately about the consumer rather than the technology: a label is required where the synthetic content materially influences consumer decisions and leaving it undisclosed would mislead. Three practical buckets:
Label required. Replicating a real person's face or voice for a consented endorsement in personalised messaging. Demonstrating a product that does not yet exist. AI-generated sponsored product recommendations, which must identify the sponsor as "Sponsored by [Brand]". Suggested wording includes "Video created using AI" or "Video enhanced using AI".
No label required. Minor modifications. Routine edits that do not change core claims. Decorative elements unrelated to the product's capabilities. Accessibility features such as subtitles and captions. This bucket is wider than the panic suggests, and it is the part most commentary omits.
Prohibited regardless of labelling. Fabricated endorsements. Synthetic depictions that exaggerate results — the example given is unachievable weight-loss outcomes. Anything otherwise breaching the ASCI Code. A label does not cure a misleading advertisement. If an agency proposes generating a result your product does not produce and labelling it, the label is irrelevant and you are the one in breach.
The IT Rules bind platforms, not your agency — but they reach your asset. MeitY notified the Information Technology Amendment Rules, 2026 in February, in force from 20 February 2026. They define synthetically generated information as audio, visual or audio-visual content made or altered to appear real and indistinguishable from a real person or event. Khaitan's summary of the notified rules sets out the obligations, which sit on intermediaries and on significant social media intermediaries — user declarations, verification, labelling, and sharply shortened takedown timelines, including three hours for court orders or reasoned government intimations.
Two things to get right here, because they are widely misreported:
Text is outside the definition. AI-drafted blog posts and copy do not trigger the synthetic-media labelling duties. Routine and good-faith editing, document preparation, and tools that improve clarity, quality, translation or accessibility without manipulating a material part of the content are also outside it.
The "10% of screen area" rule you may have read about is not law. The October 2025 draft required visual labels to cover 10% of the display area. The notified rules replaced that with a qualitative standard — labels must be prominent, easily noticeable and adequately perceivable. Rödl & Partner's draft-versus-final comparison sets out the change, and the notified rules are on MeitY's site. We flag this specifically because the 10% figure is circulating widely as though it were in force, including in research we commissioned for this article. If a vendor is selling you compliance against a 10% watermark, they have read a draft.
Can you even own AI-generated creative?
Probably not as copyright, and this is the part with real commercial teeth.
Indian copyright law requires a human author. For a computer-generated work, the author is the person who causes the work to be created, and Indian law does not treat an algorithm as a person. The administrative signal came from the Copyright Office's handling of the artwork Suryast, where an AI system was initially recorded as a co-author alongside its human operator and that registration was subsequently withdrawn.
The practical consequence for you: platform terms may grant you commercial rights to use an output, but a commercial licence is not copyright. If your campaign creative is purely generated, a competitor who copies it may face no enforceable claim from you. Where a professional has exercised substantial skill, judgement and arrangement over the output, the position improves — which makes documenting the human contribution a legal step, not an administrative one.
Three things to put in your agency contract as a result:
- Which assets are generated, which are substantially human-authored, and who holds what rights in each
- Which tool tier was used, since vendor commercial rights and indemnities differ by tier and a consumer-tier account used for a large client's work may not carry the licence the work requires
- Who carries the ASCI exposure if a labelling decision turns out wrong
Ask for all three in writing. An agency that will not put them in writing is telling you where it thinks the risk should sit.
So where does an AI-native agency actually leave you better off?
Four places, stated narrowly enough to be checked:
Tested creative volume in paid media. More executions reaching a platform, so decisions rest on data rather than on the account manager's taste. This is the clearest gain and it shows up in a variant count.
Drafting and research speed on content that still clears a bar. Faster to first draft, with the same editorial standard. It shows up in publishing cadence, not in word count.
Analysis across more surface. Pattern-finding over more accounts, campaigns and periods than a human would read, with a human deciding what it means. The reasoning behind this is in why analytics is the real lever.
Production capacity for formats that used to be prohibitive. Video and social creative at a volume a mid-market budget could not previously reach — with the disclosure obligations above attached, which is why we raise them before the shoot rather than after.
And the honest list of what has not changed: your positioning, your offer, your pricing, your distribution, and the quality of the thing you sell. Those determine your outcome. No part of this article touches them. If your problem is that the offer is wrong, a thousand variants will find that out faster and more expensively. That is the real version of the problem behind most failed lead generation.
When you should not hire an AI-powered agency
Three situations, and we would rather say them now.
Your constraint is trust or compliance, not content volume. In regulated and relationship-led sectors, throughput is the wrong axis. A chartered accountancy practice or a legal firm competes on precision and permission, not on variant count.
You could hire one competent person instead. The tools are generally available and cheap. If your need is steady production within a settled strategy, an in-house hire with the same subscriptions may beat a retainer, and the advantage an agency holds here is genuinely narrow and genuinely temporary.
You want someone to own the strategy. Then buy judgement, and evaluate the agency on the quality of its thinking in the first conversation. The AI part is irrelevant to that purchase, and an agency leading with it for a strategy brief is answering a question you did not ask.
For the wider picture of how this applies to a business in this market rather than to agency selection, we wrote a fuller guide to AI-powered marketing.
Want to test us against this article? Ask us what we refuse to use AI for, ask for a variant count and test cadence from a live account, and ask us to put the ownership and labelling allocation in the contract. If the answers are thin, you will know, and that is the point of publishing the test. Talk to Midgrow, or message +91 9632398896. How engagements are structured is on our services page.
Frequently asked questions
Does an AI-powered marketing agency produce better results?
Not better judgement — better throughput on structured production. The strongest available evidence, a randomised field experiment with 758 BCG consultants, found large gains inside the tool's competence and a 19 percentage point fall in correctness outside it. So the gain is real and bounded. An agency claiming AI improves strategy is describing the condition where accuracy dropped.
What should I ask an agency that says it is AI-powered?
Ask what it refuses to use AI for. An agency that cannot answer in specifics has bought subscriptions rather than built a method. Then ask for a monthly variant count, how many variants actually ran, and how many tests reached a decision — the countable evidence that throughput reached you rather than stopping at their margin.
Does Google penalise AI-generated content?
No, not for being AI-generated. Google's scaled content abuse policy targets content produced at volume primarily to manipulate rankings, and it applies whether the content was made by automation, humans, or both. Purpose and value are assessed, not method. The practical effect is that AI throughput does not safely transfer to publishing content at scale.
Do I have to label AI-generated ads in India?
ASCI's final guidelines, published 29 September 2026 and effective 29 December 2026, require labelling where synthetically generated content materially influences consumer decisions and non-disclosure would mislead. Separately, Google introduced advertiser AI declarations across Search, YouTube and Discover in July 2026. Minor edits, decorative elements and accessibility features generally do not require a label.
Is the 10% screen-area AI watermark rule real?
No. That requirement appeared in the October 2025 draft of India's IT Rules amendment. The notified rules, in force from 20 February 2026, replaced it with a qualitative standard requiring labels to be prominent, easily noticeable and adequately perceivable. The figure is still circulating widely as though it were law.
Does AI-generated marketing content have to be labelled if it is text?
No. India's notified definition of synthetically generated information covers audio, visual and audio-visual content, so AI-drafted copy and blog posts fall outside those specific labelling duties. General advertising law and the ASCI Code still apply to what the text claims.
Who owns AI-generated creative that an agency delivers?
Indian copyright requires a human author, and the Copyright Office withdrew a registration that had recorded an AI system as co-author. A platform licence to use an output is not copyright, so purely generated creative may not be enforceable against a competitor who copies it. Substantial human skill and judgement over the output improves the position, which is why the human contribution should be documented and the rights allocation written into the contract.
Will using AI-generated ads get my brand in trouble with ASCI?
Only if the advertisement is misleading. ASCI is explicit that a label does not cure a prohibited advertisement — fabricated endorsements and synthetic depictions exaggerating results remain breaches however prominently they are labelled. The exposure sits with the advertiser, so the labelling and liability allocation belongs in your agency agreement.
