Two statistics drive almost every conversation about AI implementation failure. One says more than 80% of AI projects fail, with 85% of those failures caused by poor data quality. The other says 95% of generative AI pilots produce zero return.

We checked both against their source documents. Both are misquoted, and both are misquoted in the direction that sells remediation work — which is work we sell. So it is worth being precise about what the research actually found, because the real findings still support building a data layer first. They just support it for different reasons, with less certainty, and with a specific warning the popular version leaves out.

What do the failure statistics actually say?

The 80% figure is not RAND's finding. RAND's report on the root causes of AI project failure is real and worth reading. But it is based on 65 semi-structured interviews — 50 industry practitioners, 15 academics — conducted between August and December 2023. Where it says "by some estimates, more than 80 percent of AI projects fail," it is citing a 2022 magazine article, not reporting its own measurement. The precise-sounding "80.3%" that circulates with RAND's name attached appears nowhere in it, and neither does any analysis of 2,400 enterprise initiatives.

The 85% data-quality figure is not in the report at all. What RAND actually reports is that 30 of its 50 industry interviewees discussed persistent data quality problems. That is 60% of a small interview sample describing a problem they encounter — a meaningfully weaker claim than 85% of failures being caused by it, and a weaker claim is what the evidence supports.

RAND also flags a limitation nobody repeats. Its five industry root causes put leadership first: misunderstanding or miscommunicating the problem the AI is meant to solve. 84% of interviewees cited a leadership-driven cause as primary. But RAND notes in the same report that most of its interviewees were non-managerial, so the finding may skew toward blaming leadership. That is an unusually honest caveat, and it has been stripped out of every secondhand version we found.

The 95% figure counts organisations, not pilots. The MIT Media Lab's Project NANDA work reviewed 300-plus public AI initiatives, interviewed people at 52 organisations and collected 153 survey responses from senior leaders at four industry conferences, between January and June 2025. Its funnel: 60% of organisations evaluated task-specific AI tools, 20% reached a pilot, 5% reached production. Those are shares of organisations at each stage, not a 95% pilot failure rate. The report describes its own findings as preliminary and "directionally accurate based on individual interviews rather than official company reporting."

And Gartner's 60% is a forecast. Gartner's widely quoted position — that organisations will abandon 60% of AI projects unsupported by AI-ready data through 2026 — comes from a survey of data-management leaders and is framed as a prediction. It is a reasonable prediction from a credible firm. It is not a measurement of what happened.

None of this means the implementations aren't failing. It means the failure rate is not actually known to one decimal place, and anyone quoting it that way has not read the source.

What is the Demo Gap?

Here is the mechanic that explains the buying pattern, and it does not require a failure statistic.

A generation tool demonstrates beautifully in a sales meeting. Someone types a prompt, thirty ad variants appear, and the value is visible in ninety seconds. A data layer demonstrates nothing. It is a migration, a deduplication exercise, an identity-resolution problem and several uncomfortable conversations about who owns the customer record. There is no moment in which it looks impressive.

So purchasing follows demo quality rather than dependency order. That is the Demo Gap: the distance between how well a layer presents and how much the rest of the stack depends on it. The layers that depend on everything else present worst, and get bought last.

This is a judgement about organisational behaviour, not a research finding, and we should be straight about that. No peer-reviewed study has tested procurement sequence. Nobody has taken fifty companies, made half buy infrastructure first and half buy generation first, and measured the difference. The data-first principle is a strong practitioner heuristic supported by retrospective failure analysis. It is not proven, and the layer model it sits inside is itself an analyst artefact — the four-layer marketing stack taxonomy was popularised by Scott Brinker in industry writing between roughly 2009 and 2015, not derived from computer science.

We still think it is right. You are entitled to know it is a heuristic.

What are the four layers, honestly described?

The data layer. Everything you know about a customer, in one place, retrievable. For a mid-market Indian company this is usually a CRM used properly rather than a customer data platform — and "used properly" is the hard part. The specific obstacles here are local and real: phone number as de facto identity, with the same customer holding three numbers; walk-in and offline customers who never enter a system; WhatsApp conversations held on individual salespeople's personal devices and lost when they leave; and duplicate records created by every form on the website.

The generation layer. Content, creative, variants, first drafts. This is where the market is loudest and where the evidence is actually most specific — see below.

The automation layer. Routing, sequencing, handover, triggers. Most of what is sold as AI here is deterministic workflow automation, which is not a criticism — workflow automation is reliable and cheap, and calling it AI is a pricing decision rather than a technical one.

The measurement layer. This has become the hardest layer, not the easiest, because the platforms now require you to supply what they used to collect. Meta's Conversions API requires advertisers to hash customer identifiers with SHA-256 before transmission — Meta publishes no endpoint accepting raw personal data for attribution. Google has moved advertisers toward server-side tagging, consent signalling and modelled conversions. Both changes push the work down into your first-party data infrastructure, which means the measurement layer has a hard dependency on the data layer too. If anyone has told you measurement is a dashboard purchase, they are describing 2019. We go through what honest measurement now requires in our piece on marketing analytics.

What does the research actually say about where AI helps?

This is the best-evidenced part of the whole topic, and it is more specific than "AI makes you faster."

The Harvard Business School and BCG field experiment ran 758 professional consultants on real tasks. Inside what the researchers called the jagged technological frontier — the invisible boundary of what the model does well — participants produced 40% higher quality work and finished 25.1% faster. Outside it, they were 19 percentage points less likely to reach the correct answer. The frontier does not track human intuitions about difficulty, which is what makes it dangerous: a task that feels hard may be well inside it, and a task that feels trivial may be outside.

The Brynjolfsson, Li and Raymond study of 5,179 customer support agents found a 14% average increase in issues resolved per hour — and a 34% increase for novice and low-skilled workers, with minimal effect on experienced, highly skilled staff.

Put those together and you get a finding that should change how you budget. AI raises the floor far more than it raises the ceiling. It compresses the gap between your newest hire and your best one. It does comparatively little for the person who was already excellent. So if your marketing problem is that your senior people are stretched, a generation tool will not fix it. If your problem is that your junior output is inconsistent, it might. That distinction is the one we apply when scoping work, and it is the honest version of the delegation question we covered in our guide to AI-powered marketing.

What does an AI marketing stack actually cost an Indian company?

Three cost components, and the second one is the one nobody mentions.

Subscriptions. We are deliberately not publishing rupee figures here. Every India-specific price we could find for the major platforms traced to a reseller or partner blog rather than the vendor, and software list prices are region-gated and change constantly. What matters structurally is this: Indian pricing exists for some products and not others; the tiers that actually support automated workflows and cross-object reporting are rarely the advertised entry tiers; and at least one major platform charges mandatory one-time onboarding fees billed in US dollars, which convert without any adjustment for local purchasing power. Get the quote in writing, in rupees, with the onboarding line itemised.

Indirect tax, which is recoverable — and one levy that is now gone. When a GST-registered Indian business subscribes to foreign software, it accounts for IGST on the import of that service under the reverse charge mechanism and pays it directly to the government. Critically, a registered business using the software for business purposes is generally able to claim that back as input tax credit — so the real cost is the subscription, not the subscription plus 18%. Unregistered buyers absorb it as a sunk cost. The exact mechanism and your eligibility depend on your registration and use, so have your CA confirm it against the IGST Act and the relevant rate notifications rather than taking it from an article.

Two separate equalisation levies also mattered here, and the commentary conflates them. The 2% levy on non-resident e-commerce operators was withdrawn with effect from 1 August 2024. Separately — and this is the one that bears directly on marketing — the 6% levy on online advertising paid to non-resident platforms was abolished by the Finance Act, 2025, effective 1 April 2025. If your media cost model still carries a 6% line for payments to foreign ad platforms, it is overstating your cost. We cover the rest of the real cost picture in what agencies don't tell you about pricing.

The operator. Someone has to run this. We looked for defensible salary data for marketing operations roles in Indian tier-two cities and did not find any with disclosed methodology — the field is dominated by job-board aggregators. So we will not give you a number. We will say that in every stalled implementation we have looked at, the missing item was a person with time, not a tool with features.

Can you legally put Indian customer data in a foreign tool?

Yes, by default — and the widespread belief that you cannot is wrong in a way that costs companies money.

India's data protection framework does not impose general data localisation. Section 16 of the Digital Personal Data Protection Act, 2023 works on a negative-list basis: transfers abroad are permitted unless the Central Government notifies a restriction on a specific country or territory, and no such restricting notification has been issued. So a US-hosted CRM or a European automation platform is lawful.

Three qualifications, all of which matter.

Accountability does not travel with the data. You remain the accountable data fiduciary. Processing agreements with your overseas vendors are not paperwork, they are the mechanism by which you discharge an obligation you cannot delegate.

Sector rules override the default. Section 16 expressly preserves stricter existing localisation requirements — the Reserve Bank of India's payment-system data directive being the obvious one. If you are in a regulated sector, your regulator's rule wins.

The timeline is phased, and the near date is not your date. The DPDP Rules were notified on 13 November 2025. Rule 4, the consent-manager framework, commences 13 November 2026. The substantive obligations — the consent, notice, processing and data-principal-rights rules that actually govern a marketing database — commence 13 May 2027. That is the date to plan against. It is also far enough out that doing nothing until 2027 is a decision to rebuild your data layer twice, because consent-grade records cannot be reconstructed retroactively.

The One-Customer Test

Before buying anything, run this. It is free and it takes five minutes.

Pick one real customer by name. Without asking anyone else, and without opening more than one system, retrieve: every enquiry they ever made and through which channel, every quote sent, what they bought and when, every WhatsApp or phone conversation, which campaign brought them in, and whether they consented to marketing contact and when.

If you can do that, you have a data layer and you should go buy a generation tool. If it takes three people and a spreadsheet, no generation tool will help you, because there is nothing underneath it to generate from. And if you cannot answer the consent question at all, that is the item with a 2027 deadline attached.

When buying generation first is the right call

Here is the strongest case against our own argument, because it is sometimes correct.

If you are a small team producing a handful of assets a week, a generation tool scoped to a narrow workflow — ad variants, product descriptions, first-draft captions — is a sound first purchase. It costs little, it works immediately, and it needs no unified customer data because it is not personalising anything. An infrastructure project with a twelve-month payback is the wrong purchase for a company that needs output next Tuesday.

The failure mode is not buying generation first. It is buying generation first and then expecting it to do personalisation, segmentation, lifecycle messaging or attribution — all of which read from the data layer you skipped. Scope it narrowly, get value from it, and build the data layer when the limit of the narrow version is what is actually holding you back.

There is a related finding worth knowing. The MIT work found that organisations which bought specialised vendor tools succeeded more often than those that built internal ones, because internal builds tend to lose context between sessions and never integrate deeply into daily work. For a mid-market Indian company without an engineering team to spare, that is the practical answer to build-versus-buy: buy, and spend your scarce internal effort on the data layer, which is the part no vendor can do for you.

Four situations where we would tell you not to buy. When nobody has time to operate it — the tool will go unused and you will conclude AI does not work. When your customer data lives in three systems and the plan does not include consolidating them. When the pitch promises personalisation but the discovery call never asked where your customer data sits. And when the business problem has not been stated in a sentence that includes a number, which is RAND's leading root cause and the cheapest one to avoid.

Want your stack assessed before you sign anything? Send us what you currently pay for, where your customer data actually sits, who operates your marketing systems, and what the tool you are being pitched is supposed to fix. We will come back with which layer you are missing, whether the sequencing in the proposal makes sense, what the real cost of ownership looks like after tax recovery, and where your DPDP exposure sits ahead of 2027. If the honest answer is that you need one cheap narrow tool rather than a stack, we will tell you that. Talk to Midgrow, or message +91 9632398896 with your sector and current tooling. Engagement structures are on our services page, the lead-handling side sits with why lead generation fails, measurement with how to measure ROI properly, content production with our social work, and sector-specific context for manufacturers.

Frequently asked questions

What are the layers of an AI marketing stack?

Four: a data layer that unifies customer and campaign records, a generation layer for content and creative, an automation layer for routing and workflows, and a measurement layer. The generation and measurement layers both read from the data layer, which is why building them on fragmented data produces output that cannot be personalised or attributed.

Do 80% of AI projects really fail?

The figure is not as solid as its circulation suggests. RAND's report on AI project failure quotes "more than 80 percent" from a 2022 magazine article rather than measuring it, and RAND's own research was 65 interviews conducted in late 2023. The precise-sounding variants attributed to RAND, including 80.3%, do not appear in the report.

Did MIT find that 95% of AI pilots fail?

Not as usually stated. The MIT Media Lab's Project NANDA work found that 5% of organisations reached production with task-specific AI tools, against 20% reaching a pilot and 60% evaluating. Those are shares of organisations at each stage, not a pilot failure rate, and the report labels its own findings preliminary.

Why do AI marketing implementations fail?

The best-evidenced causes are organisational rather than technical: the business problem is not clearly defined before purchase, and customer data is fragmented or poor quality. RAND's interviewees named leadership-driven causes most often, though RAND notes most of its interviewees were non-managerial, which may skew that result.

Can Indian companies store customer data in foreign marketing software?

Yes, by default. Section 16 of the DPDP Act, 2023 permits cross-border transfer unless the Central Government notifies a restriction on a specific country, and no such restriction has been issued. Accountability stays with the Indian business, and stricter sector rules — such as RBI's payment-data localisation — override the default.

When does DPDP compliance become mandatory for marketing databases?

The Rules were notified on 13 November 2025 with phased commencement. The consent-manager framework in Rule 4 commences 13 November 2026, and the substantive consent, notice, processing and data-principal-rights obligations commence 13 May 2027. Consent records cannot be reconstructed retroactively, so the work starts before the deadline.

Is GST on foreign software subscriptions recoverable in India?

Generally yes for a GST-registered business. IGST on the imported service is accounted for under reverse charge and paid to the government, and is ordinarily available as input tax credit where the software is used for business. Unregistered buyers absorb it. Confirm your own position with your CA against the IGST Act.

Does AI help experienced marketers or junior ones more?

Junior, substantially. A study of 5,179 customer support agents found a 14% average productivity increase but 34% for novice and low-skilled workers, with minimal effect on experienced, highly skilled staff. AI raises the floor more than the ceiling, so it addresses inconsistent junior output rather than stretched senior capacity.