If you are about to sign an AI-powered marketing retainer, the questions in your head are probably: which models do you use, how much time does it save, what's your approach? A prepared seller answers all three identically whether or not a single thing is automated. Those questions do not discriminate.
The ones that do share a single property: they require an artefact. Something on a screen. That distinction is the whole audit, and it is the test we would want applied to our own work before anyone signs anything.
There is also a legal point most Indian buyers do not know, and it changes what you should do in the meeting. We will get to it.
What have regulators already established about fake AI claims?
You do not have to invent a standard. Regulators built one, and it converts directly into questions.
In March 2024 the US Securities and Exchange Commission settled charges against two investment advisers for what it called AI washing. Delphia paid a $225,000 civil penalty; Global Predictions paid $175,000. Both settled without admitting or denying the findings.
What they were found to have claimed matters more than the money. Delphia said its investment advice was powered by predictive algorithms trained on collective client data. The SEC found it collected the data but never integrated AI or machine learning into the algorithms at all. Global Predictions marketed itself as the "first regulated AI financial advisor" and claimed AI-driven outperformance it could not substantiate.
Six months later the US Federal Trade Commission ran Operation AI Comply. The most transferable case is DoNotPay, which marketed an "AI lawyer." The FTC's core finding was not that the tool was bad. It was that the company had not conducted testing to establish that its output matched human professional standards before claiming it did. In the Evolv matter, the FTC alleged the company marketed its detection system's best-case performance while the settings required to achieve reliable detection produced enormous false-positive rates.
Three principles come out of this, and all three are askable:
- Intent is not capability. Planning to build an AI system does not entitle you to market it as working today.
- Using a third-party API is not proprietary technology. Having a subscription is not having a system.
- Claiming your output equals professional human work requires testing you did before you said it.
Nothing equivalent has happened in India. No Indian regulator or tribunal has publicly adjudicated an AI-washing claim against a marketing agency. That absence is itself a finding, and it leads straight to the part that matters for you.
Can you sue an Indian agency for false AI claims?
Not as a consumer. This surprises most buyers.
Section 2(7) of the Consumer Protection Act, 2019 defines a consumer to exclude a person who obtains goods or services for a commercial purpose. A business buying a marketing retainer is buying for a commercial purpose. So consumer forums are closed to you. No tribunal, no consumer complaint, none of the machinery people assume is there.
Your remedy is contractual, under the Indian Contract Act, 1872. Section 18 defines misrepresentation to include a positive assertion made in a manner not warranted by the information of the person making it — even where they believed it true. Section 19 makes a contract voidable at the option of the party whose consent was obtained that way.
Which produces the single most practical instruction in this article.
The Paper Trail Rule: if the AI claim is not written into the contract, you cannot prove it was made. A misrepresentation claim needs the assertion in evidence. A slide in a pitch deck that was never attached to the agreement, and a confident sentence in a meeting, are close to worthless when the relationship sours.
So the ask is simple, and an honest agency will not blink at it: put the specific capability claims into the service agreement as warranties. Not "AI-powered marketing." The actual sentences. Which workflows are automated, what is reviewed by a human, what the agency says it has built. If a seller resists writing down what they just told you, you have learned everything you came to learn, and you did not need to understand a single thing about machine learning.
The Three Claims, and what each one actually requires
Sales language deliberately merges three very different things. Separating them is most of the audit.
A tool subscription. The agency pays for access to a commercial platform and staff type prompts into it. Cost is a few thousand rupees per seat per month. It requires no engineering and no data expertise. Verifiable artefact: essentially none beyond a billing receipt and a browser window. This is a perfectly legitimate way to work — most good agencies work this way — but it is not proprietary technology and should not be priced as if it were.
An automated workflow. The agency connects models to its own data and logic through APIs, so a brief goes in and output comes out without a human copying and pasting between screens. The underlying model is unchanged; the value sits in the pipeline, the retrieval, and the routing. Verifiable artefacts: execution logs, API usage records, the routing code, an architecture diagram, and a live run on a brief the agency has not seen before.
A trained system. The agency has actually altered a model's weights for a specific task. This no longer requires millions — modern efficient fine-tuning techniques let a small team adapt an open model on modest hardware with a curated dataset, so a small honest agency genuinely can have one. Verifiable artefacts: the training dataset, the training loss curves showing error falling across runs, and version-controlled model files.
The discriminating question follows from the table above rather than from any technical knowledge: which of these three is it, and show me the artefact for that one. An agency on a subscription calling it a proprietary model cannot produce logs, a dataset or a loss curve, because none exist.
The Artefact Test: five questions that actually discriminate
Ask these in the meeting. The right answers are specific and come with something on screen.
1. "Run it on this brief, now." Hand over something they have not seen. A real pipeline executes in front of you. A subscription becomes a person typing into a chat window — which is fine, if that is what you are buying and paying for.
2. "Who reviews the output before it ships, at what stage, and what happens when they reject it?" This is the question AI theatre fails. Generative systems produce confidently wrong output at a rate that does not go away by choosing a bigger model, which is why every serious deployment routes a measurable share of output to human review and keeps a record of it. An agency with a real pipeline describes a review step, a queue, and a rejection path. An agency without one describes "our team checks everything," which is not a process.
3. "Which of the three is it — subscription, workflow, or trained system?" Then ask for that one's artefact. Note the answer, because it goes in the contract.
4. "What do I own in the deliverables?" Covered below, and almost nobody asks it.
5. "You say AI makes you faster. Where did that efficiency go?" If production cost fell materially, it went somewhere: more volume, lower cost per asset, or the agency's margin. Any of those can be a legitimate answer. No answer at all means the efficiency claim is decoration. We wrote about what agencies don't tell you on pricing separately.
And the questions to stop asking, because they discriminate between nothing: which model do you use, what's your AI approach, how much time does AI save you, do you use AI responsibly. Every one of those has a fluent answer available to a seller with no pipeline whatsoever.
Who is liable when AI creative goes wrong — you or the agency?
You are. This is the finding most buyers have not priced in.
ASCI published its Guidelines for Responsible Labelling of Synthetically Generated Content in Advertising on 29 September 2026, taking effect three months after publication — so from late December 2026. On the question of who carries the obligation, ASCI's Secretary General and CEO Manisha Kapoor was direct: "The responsibility remains with advertisers." The guidelines do not place the duty on the agency that produced the work.
So your brand carries the regulatory and reputational exposure for creative your agency generated. Which makes indemnification a contract term worth having rather than a technicality.
The full guidelines are worth reading before your next creative brief. What actually requires a label: synthetic influencers or ambassadors; a real person's likeness or voice used for personalised messaging even with consent; fabricated events or settings that change how a consumer understands the product; demonstrations of products that do not yet exist; synthetic sound effects central to a product's core features; and paid AI-generated product suggestions, which must carry "Sponsored by [Brand]".
What does not require a label, and this is the useful half: routine colour correction, lighting adjustment, noise reduction and minor blemish removal; decorative or ambient elements; obviously fantastical effects; accessibility uses such as subtitles and translations; and — importantly — administrative and text-based uses such as drafting ad copy. An agency using AI to write captions is not triggering a disclosure obligation. If someone is selling you a compliance package on that basis, they have not read the guidelines.
A label also does not rescue prohibited content. Fabricated endorsements, exaggerating results beyond what is substantiated, deepfakes, and unauthorised use of someone's likeness or copyrighted work remain banned whether disclosed or not.
What do you actually own in AI-generated work?
Possibly nothing, and your contract probably assumes otherwise.
Section 2(d)(vi) of the Copyright Act, 1957 says that for a computer-generated work, the author is the person who causes the work to be created. India's clearest signal on how that applies to generative AI is the "Suryast" matter, in which the Copyright Office rejected an application naming an AI system as author, then granted registration when a natural person was named as co-author alongside it, then issued a withdrawal notice — concluding that an AI system cannot qualify as an author under the Act. The position remains unsettled and no Indian court has decided it.
The practical consequence for a buyer: where a deliverable is substantially prompt-to-output with little human authorship, the copyright you believe you are being assigned may not exist to assign. Ask the agency to describe the extent of human authorship in the final work, and make the assignment clause reflect reality rather than aspiration.
Does AI content hurt your search rankings?
No, and the honest answer here is narrower than both the scare version and the sales version.
Google's published guidance is that it rewards high-quality content however it is produced, while automation used primarily to manipulate rankings remains a spam violation. Volume is therefore not an advantage — which removes the main thing agencies claim AI wins you in search. We set out what does work in our SEO guide for 2026.
One narrower requirement is real and often misstated as applying to all of Search: in Google Merchant Center, AI-generated product images are expected to carry the appropriate IPTC metadata identifying them as such, and that metadata should not be stripped. If you sell online this matters operationally — see our work with retail and ecommerce businesses — but it is a Merchant Center rule, not a sitewide search penalty for AI content.
The honest counter-argument: is this audit theatre?
There is a strong case that it is, and we should put it properly.
You are not a machine learning engineer. If you ask for training loss curves, you cannot tell whether they came from the agency's production runs or a public repository. If you ask for a live demonstration, a determined seller can stage one — a front end that looks automated while someone executes manually behind a loading screen. Regulators who establish AI washing do it with subpoena power over internal emails, code commits and server logs. You have none of that. The processing really is a black box from outside.
On that view, every question here mainly reassures you rather than protecting you, and the only test that truly discriminates is a short paid pilot judged on delivered outcomes — because if the cost per asset falls and the campaigns perform, it genuinely does not matter commercially whether a model or a fast team produced the work.
We think that is substantially right, and it does not make the audit worthless. Two things survive it. The Paper Trail Rule survives entirely: a written warranty is enforceable regardless of your technical ability, and a seller unwilling to sign one has told you something real. And the audit is cheap. Asking five questions costs one meeting; a pilot costs a month and a fee.
So the sequence is: run the Artefact Test to screen, write the claims into the contract to protect, and then buy a small pilot before a long retainer and judge it on outcomes. Judging on outcomes is its own discipline, which we cover in measuring marketing honestly.
What we would tell you to walk away from
Four signals, in order of seriousness.
They will not write the claim into the contract. Nothing else on this list matters as much.
They cannot describe a review step. Not "we check it" — who, at what point, and what happens on rejection.
They call a subscription a proprietary model. This is the specific conduct regulators have penalised elsewhere, and the artefact request settles it in thirty seconds.
They claim large AI efficiency and unchanged pricing with no explanation of where the gain went. It may be a fair answer; there should be one.
Where we would decline. If what you need is low-cost content volume, you should use tools directly and we would say so. If you want a system someone is accountable for, that is a different engagement — and we would expect you to run every question above on us, including the contract one.
Want this run on a proposal you have in hand? Send us the deck or scope you have been given, the monthly value, and what the deliverables are. We will tell you which claims are verifiable, which are structurally unverifiable, what we would ask in the next meeting, and which sentences should go into the contract as warranties. If the proposal is sound, we will say that too. Talk to Midgrow, or message +91 9632398896 with your sector and budget. Our own engagement structures are on the services page, our AI approach is set out in this guide, content production sits with our social work, and lead handling in why lead generation fails.
Frequently asked questions
How do I verify an agency's AI claims?
Ask for an artefact rather than an explanation. Have them run the pipeline live on a brief they have not seen, ask which of three things it is — a tool subscription, an automated workflow, or a genuinely trained model — and ask for that category's evidence: execution logs and routing code for a workflow, a training dataset and loss curves for a trained model. A subscription has no artefact beyond a browser window.
What questions should I ask an AI marketing agency?
The discriminating ones require something on screen: run it live on a fresh brief; who reviews output before it ships and what happens when they reject it; which of the three architectures is it and show the evidence; what do I own in the deliverables; and where did the claimed efficiency go. Questions about which model they use or how much time AI saves can be answered identically by an agency with no automation at all.
Can I take legal action if an agency's AI claims are false in India?
Not through consumer forums. Section 2(7) of the Consumer Protection Act, 2019 excludes services obtained for a commercial purpose, so a business buyer is not a consumer. The remedy is contractual: Section 18 of the Indian Contract Act, 1872 covers misrepresentation and Section 19 makes the contract voidable — but only if the claim was documented, which is why capability claims belong in the agreement as written warranties.
What is AI washing?
Marketing artificial-intelligence capability a company does not actually have. US regulators have penalised it: the SEC settled charges in March 2024 against two advisers for $225,000 and $175,000 respectively, in one case finding the firm collected client data but never integrated AI into its algorithms at all. The established principles are that intent is not capability and a third-party subscription is not proprietary technology.
Who is responsible for labelling AI-generated ads in India?
The advertiser. ASCI's guidelines on labelling synthetically generated content, published 29 September 2026 and effective three months later, place responsibility with advertisers rather than the agencies producing the work. Brands therefore carry the exposure for agency-generated creative, which makes indemnification worth negotiating into the contract.
Does AI-generated ad copy need a disclosure label in India?
Generally no. ASCI's guidelines exempt administrative and text-based uses such as drafting ad copy, along with routine colour correction, lighting adjustment, blemish removal, decorative elements, obviously fantastical effects, and accessibility uses like subtitles and translations. Labelling applies where synthetic content materially shapes a consumer's understanding.
Do I own the copyright in AI-generated marketing work?
Possibly not. Section 2(d)(vi) of the Copyright Act, 1957 makes the author the person who causes a computer-generated work to be created, and the Copyright Office has signalled that an AI system cannot be an author — registering and then issuing a withdrawal notice in the Suryast matter. Where a deliverable is largely prompt-to-output with minimal human authorship, there may be no copyright to assign.
Is a paid pilot better than auditing an agency's AI claims?
They do different jobs, and you should do both. An audit costs one meeting and screens out sellers who will not document their claims. A pilot tests what actually matters — delivered outcomes — because if cost per asset falls and campaigns perform, the underlying mechanism is commercially irrelevant. Screen with questions, protect with contract warranties, then buy small before committing to a long retainer.


