Agentic Campaigns, One Year Later: Are Indian Startups’ AI-Run Ad Platforms Actually Delivering ROI

A year ago, the pitch decks all sounded the same. Feed the machine a budget and a brand objective, the story went, and an autonomous agent would plan, buy, optimise and report on a campaign with almost no human in the loop. A handful of Indian AdTech startups built entire product lines around this promise, positioning themselves as the local answer to the agentic-AI wave sweeping through Silicon Valley’s marketing stacks. Founders spoke of “self-driving media plans” and “always-on optimisation engines” in the same breathless register once reserved for programmatic itself. Investors, hungry for a category to bet on after the DSP consolidation of the mid-2020s, wrote cheques accordingly.

Twelve months on, the language has cooled considerably. Walk the halls of any Mumbai or Bengaluru marketing conference today and the word “agentic” still gets used, but it arrives with more caveats and fewer superlatives. Marketers who signed on early are now the ones best placed to answer the only question that ever really mattered: did the machine actually deliver better returns than the humans it was meant to replace, or at least assist more cheaply?

The honest answer, gathered from conversations across brand and agency teams that ran real budgets through these platforms, is a qualified yes with an asterisk the size of a billboard.

The efficiency case held up better than the intelligence case

Where agentic platforms have earned their keep is in the unglamorous middle of the funnel — bid management, budget pacing, creative rotation and cross-channel reallocation at a frequency no planner could sustain manually. A performance marketing lead at a Gurugram-based D2C personal care brand, who has run two full fiscal quarters of spend through one such platform, describes the gain in blunt terms: fewer wasted impressions during dayparts that never converted, and a pacing discipline that kept the brand from the familiar end-of-month scramble to spend down leftover budget at inflated CPMs.

That is a real, measurable win, and it is not a small one in a market where media inflation on platforms like Meta and Google has made every rupee of wastage more expensive than it was two years ago. Several agency operators independently arrived at the same figure: agentic pacing and bid optimisation alone accounted for single-digit to low-double-digit percentage improvements in cost-per-acquisition, mostly by eliminating the kind of human lag that comes from someone checking a dashboard once every few hours instead of every few minutes.

But efficiency is not the same thing as intelligence, and this is where the agentic story gets more complicated. The platforms that promised genuine strategic judgment — audience discovery beyond what the walled gardens already surface, creative concepting that goes past template remixing, or channel-mix decisions that account for brand-building rather than last-click attribution — have had a rougher year. Multiple marketers described a pattern that has become almost a punchline in private WhatsApp groups: the agent optimises beautifully toward whatever metric it is given, and just as beautifully away from everything it was not told to protect.

“It’s a brilliant intern with no institutional memory,” is how one Delhi-based media agency director put it. “It will hit your CPA target every single time. It has no idea it’s doing that by draining the brand’s search share of voice to zero.”

That comment captures the central tension of the past year better than any case study deck. Agentic systems are, at bottom, optimisation engines. They are extraordinarily good at converging on a stated objective and considerably less good at inferring the unstated ones — brand safety nuance, category context, the difference between a discount-seeking searcher and a genuinely high-intent one. In categories with thin margins for error, that gap has been expensive.

Where the ROI numbers actually come from

Ask three different startups for their headline ROI claim and you will get three different denominators, which is part of why the sector’s own marketing has started to feel less trustworthy than the products themselves. Some measure improvement against a brand’s own historical baseline, which flatters the platform if the previous year’s campaigns were badly run to begin with. Others benchmark against category averages pulled from third-party data that clients cannot independently verify. A smaller number have submitted to genuine incrementality testing — holdout groups, geo-split experiments, the unglamorous statistical rigour that actually tells you whether spend caused an outcome rather than merely preceding it.

Unsurprisingly, it is this last group whose ROI claims have survived scrutiny best, and also the group willing to admit their numbers are more modest than the sales deck implied a year ago. One Bengaluru-headquartered platform that agreed to run a controlled geo-holdout test with a mid-sized fintech client landed on a genuinely defensible 14 percent lift in incremental sign-ups — respectable, real, and about a third of the figure the same company had floated in its Series A pitch materials before it had a single client on full autopilot.

That compression between pitch-deck ROI and audited ROI is not unique to India, but it lands harder here because so much of the category’s initial credibility was built on comparisons to overseas case studies that Indian buyers had no way to replicate locally. A platform that cites a 40 percent efficiency gain from a US retail campaign is not making a claim about the Indian festive-season media market, where inventory dynamics, regional language fragmentation and the sheer scale of app-based commerce behave quite differently.

The human-in-the-loop compromise nobody predicted

Perhaps the most interesting shift of the past twelve months is structural rather than statistical. Almost none of the marketers interviewed for this piece are running fully autonomous campaigns anymore, even at brands that started out chasing exactly that. What has emerged instead, almost without anyone formally deciding it, is a tiered model: agents handle the high-frequency, low-stakes decisions — bid adjustments, budget shifts between near-identical ad sets, pacing corrections — while anything touching brand narrative, new-audience testing, or spend above a certain threshold routes back to a human for sign-off. This is, in effect, a quiet retreat from the original agentic promise, and most people building these platforms will now say so if asked directly rather than defensively. The framing has shifted from “autonomous media buying” to something closer to “supervised autonomy,” a phrase that would have sounded like marketing cowardice at last year’s conferences and now reads as basic operational maturity.

It mirrors, in a smaller and faster-moving way, what happened with programmatic itself a decade earlier — a technology sold on the promise of removing humans from the buying chain that ultimately succeeded by giving humans better instruments rather than replacing them outright. The startups that seem to be building durable businesses are the ones that internalised this early, designing dashboards and override controls as a core product feature rather than an embarrassing admission that the AI needed supervision.

The categories where it genuinely works, and the ones where it doesn’t

Performance is not evenly distributed across sectors, and the pattern that has emerged is fairly intuitive once you see it laid out. Agentic platforms deliver their clearest wins in categories with high transaction frequency, dense first-party conversion data, and relatively low creative complexity — quick commerce, subscription apps, lending and other financial products with clear funnel events. These are environments where the agent has enough signal to learn quickly and where the cost of an occasional bad decision is small and quickly correctable.

The picture is murkier for considered-purchase categories — automobiles, real estate, premium consumer durables — where conversion cycles stretch across weeks or months and the signal an agent needs to learn from arrives too slowly and too sparsely to compound into genuine intelligence. Several marketers in these categories described reverting to largely manual planning after disappointing pilots, not because the technology was broken but because their business simply did not generate enough conversion events per week for the system to learn anything useful before the next planning cycle overtook it.

Brand campaigns present their own distinct problem, one that is more philosophical than technical. An agent optimising toward measurable proxies — video completion rate, click-through, even survey-based brand lift where available — will always be optimising toward what can be measured rather than what actually builds a brand over time. Several CMOs made a version of the same point: the platforms are excellent at telling you whether last week’s spend was efficient, and structurally unable to tell you whether this year’s campaign will still be remembered in three years. That is not a criticism unique to Indian startups; it is close to an unsolved problem in marketing measurement generally. But it does mean the “agentic” label has been applied more broadly than the technology’s actual competence justifies.

What year two will need to prove

The startups that survive the next round of client renewals will likely be the ones that stopped selling autonomy and started selling accountability — transparent reporting on what the agent actually decided and why, real incrementality testing rather than baseline comparisons, and pricing models tied to audited outcomes rather than platform fees charged regardless of performance. A few are already moving this direction, offering performance-linked pricing tiers that put real skin in the game, a shift that tellingly did not feature in any of last year’s fundraising narratives.

The other clear signal from this year’s conversations is that clients have become considerably more sophisticated buyers of this technology than they were twelve months ago. Procurement teams now routinely ask for holdout test results before signing, a question almost nobody thought to ask in the initial rush. That alone has done more to separate genuine platforms from repackaged rules-based automation than any amount of vendor marketing could.

None of this amounts to a verdict that agentic advertising has failed to deliver on its promise in India. It has delivered, but on a narrower and more mundane promise than the one originally sold — meaningful efficiency gains in well-instrumented, high-frequency categories, real but modest incrementality where anyone has bothered to measure it properly, and a genuine reduction in the operational drudgery that used to consume junior media planners’ days. What it has not delivered, at least not yet, is the wholesale replacement of strategic human judgment that the category’s most excitable pitch decks promised a year ago. The machines got faster. The good ones got honest about what that speed is actually worth.