Podcast episode

Forget The Shiny Objects And Focus On Fundamentals

by Ken Rona ·July 22, 2026 ·AdExchanger Talks

agency ai-in-adtech attribution measurement

TL;DR

Sharona Sankar-King, Havas Media Network's Chief Data and Product Officer, makes the case for measurement rigor and AI fundamentals over hype — specifically walking through how "large causal models" advance beyond traditional marketing mix modeling (MMM). The episode is most relevant to measurement practitioners and agency operators evaluating AI tooling maturity; it's light on hard market data or breaking news.

What was covered

"Black book, not black box" transparency philosophy at Havas. Sankar-King described Havas's proprietary end-to-end workflow platform (Converge.AI) as deliberately modular — built to swap data partners and tools rather than lock into acquisitions. The platform has over 500 partner integrations and uses a "build and partner" approach rather than large-scale M&A.
"Fit for purpose" data as a strategic principle. Sankar-King argued that data volume is not a proxy for data value, emphasizing that AI and predictive models perform better with curated, outcome-relevant data than with massive undifferentiated datasets — framing data quality as the primary fuel for AI performance.
MMM maturity curve and the case for large causal models. She laid out a measurement progression: basic dashboards → digital attribution → standard MMM → nested MMM → large causal models. Nested MMMs suffer from multicollinearity (when too many variables correlate with each other, making it hard to isolate what's actually driving outcomes); large causal models use AI to trace probabilistic chains of events across portfolio, channel, and creative layers simultaneously — a structural advance over stitching together multiple separate models.
Agentic AI definitions and practical deployment. Sankar-King drew a clear line: automation = rules-based; agentic AI = multi-step reasoning with API actions and governance agents; fully autonomous AI = no human oversight. She cited a Bain study putting fully autonomous AI adoption at roughly 7% of corporations.
Havas' internal AI training mandate. Every Havas employee — including administrators — is required to complete "Prof AI" training. Havas University provides deeper tracks for power users, and employees receive a stipend for external AI courses. Partnership with "Accio" (likely Accenture or a vendor by that name — transcript spelling unclear) supports multi-layer agentic architecture builds.
Performance Max / Advantage Plus tension. Host Allison Schiff pressed Sankar-King on how agencies navigate clients' reliance on Google's Performance Max and Meta's Advantage Plus — both opaque, AI-driven systems — against transparency commitments. Sankar-King's answer: treat performance channels as one slice of a full-funnel mix, not as a standalone strategy.

Notable claims & predictions

Sankar-King on causal AI adoption: "Nine out of ten clients I talk to are very bullish on this because they are hitting up against a wall with MMMs — they've squeezed all the juice they can out of nested models." Implies mainstream MMM is reaching its practical ceiling at sophisticated advertisers.
Sankar-King on autonomous AI: "Only 7% or so based on a Bain study of corporations have any kind of systems running fully autonomously with agents — we are far from being able to really realize the end-to-end autonomous agent world." Positions fully autonomous ad-tech as further off than vendor pitches suggest.
Sankar-King on AI startup red flags: "When a company says they can completely replace people without oversight, that is a red flag — everything we build today still needs oversight." A direct signal to how Havas evaluates vendor pitches.
Sankar-King on the Havas platform build philosophy: "We didn't build a one-size-fits-all system — we didn't want to be tied to a certain dataset for life... [competitors] built it for their workflow and how they operate, but not really around how their different clients work based on industry vertical." Implicit critique of holdco platforms built around acquisitions.
Sankar-King on AI and employee productivity: "Something we built saved over a month of time for one planner" — cited as a representative example of agentic tools reducing manual tasks in the planning workflow.

Fact check

Sankar-King's 7% autonomous AI adoption figure (attributed to a Bain study): Unverified. The specific Bain study, its date, methodology, and exact statistic are not named or sourced in the transcript. "7% of corporations" running fully autonomous AI is a precise claim with a specific origin that listeners cannot check. It may well be real — Bain does publish AI adoption research — but the figure should be treated as unverified until the specific study is identified. The number also does roughly comport with directionally similar industry surveys, so there is no strong basis to call it false.
Sankar-King on large causal models being "tried and true" with "biggest logos": Contested framing. She simultaneously acknowledges "a lot of causal AI players in the space are not that good" and that there is significant noise in the market — which is in tension with calling large causal models "tried and true." She is also, as Havas's Chief Data and Product Officer, directly talking her own book: Havas is positioning this methodology as a differentiator and selling it to clients. The claim that it is mature and widespread should be weighed against her commercial incentive to frame it that way.
Sankar-King's description of MMM and nested MMM multicollinearity: Accurate as a general technical description. Multicollinearity is a well-established limitation of regression-based econometric models used in MMM. No issues found.
Sankar-King's LLM analogy for large causal models ("similar to an LLM predicting tokens, but predicting causal events"): Plausible framing but simplified. The analogy is directionally useful for a lay audience but glosses over meaningful architectural differences between transformer-based language models and causal graph or Bayesian network approaches. Not false, but listeners shouldn't treat the comparison as technically precise.

Why this matters for ad-tech operators

MMM ceiling is becoming a real procurement trigger. Sankar-King's claim that sophisticated clients have "squeezed all the juice" from nested MMMs — and are now evaluating large causal model vendors — is a signal for measurement vendors and DSPs (software advertisers use to buy digital ads) that the next measurement RFP cycle may look different. Operators building or selling incrementality and causal inference products should expect increased scrutiny and a higher bar for methodology.
Holdco platform architecture is a competitive battleground. Havas's explicit positioning — modular, partner-integrated, no big data acquisitions — is a direct counternarrative to the acquisition-heavy platform strategies at Publicis (Epsilon), Omnicom (Flywheel), and others. Publishers and data partners pitching holdcos should understand which architecture model they're walking into and tailor integration stories accordingly.
"Everyone is an AI developer" is becoming a staffing and training norm, not a differentiator. Havas requiring universal AI training (including non-technical staff) signals this is table stakes for agencies. For ad-tech vendors, this raises the floor on what clients expect from UIs and explainability — products that require specialist technical users to operate will face faster displacement.
Impact for this episode is moderate and indirect. No breaking news, no M&A, no regulatory developments, no earnings signals. The primary value is as a directional read on where a mid-sized holdco's data leadership sits on measurement methodology maturity and AI deployment philosophy — useful calibration for vendors selling into agencies, but not an immediate operational trigger.

Full analysis

Havas Media Network's Chief Data and Product Officer, Sharona Sankar-King, sat down with AdExchanger's Allison Schiff to argue that agencies should stop chasing AI hype and get their measurement fundamentals right — and that the next real advance is "large causal models" that go beyond traditional marketing mix modeling (MMM). MMM is the statistical method advertisers use to figure out how much each channel — TV, search, social — actually drove sales.

The real question for operators: is the MMM ceiling Sankar-King describes a genuine procurement shift that measurement vendors need to prepare for, or is this a holdco talking its own book? This is a Type 2 (easily reversible) read — nobody's signing anything today. The forcing function is soft: the next measurement RFP cycle, not a deadline.

The Market Analyst — This episode is a positioning signal, not a news event. Havas is drawing a line in the sand against the acquisition-heavy platform playbooks at the bigger holdcos — Publicis with Epsilon, Omnicom with Flywheel — by selling "modular, we-partner-not-buy" as a virtue. In plain terms: Havas can't out-spend its larger rivals on data acquisitions, so it's reframing not-buying as principled. That's smart narrative for a #4-ish network. For measurement vendors, the useful read is that a holdco data chief is publicly saying nested MMM has hit its ceiling — that's a demand signal for incrementality and causal-inference products, regardless of whether Havas's own tool is as mature as she claims.

The Skeptic — The load-bearing assumption is that "large causal models" are a real, deployable category — and Sankar-King undercuts it herself. She calls them "tried and true" with the "biggest logos" in one breath and admits "a lot of causal AI players are not that good" in the next. Those don't reconcile. She's the Chief Data and Product Officer selling this methodology as Havas's differentiator; treat "nine out of ten clients are bullish" as a sales anecdote, not a market stat. The most honest thing she said cuts against her own pitch: only ~7% of corporations run fully autonomous AI (a Bain figure the transcript doesn't source). The hype gap is real — but so is the fog around what's actually shipping.

The Operator — Strip the strategy and ask what a planner does Tuesday. The concrete win she cited — "saved over a month of time for one planner" — is a workflow automation story, not a causal-modeling breakthrough. That's the tell: the value showing up today is agentic tooling clearing manual grunt work, while large causal models remain aspirational. In plain English: the robot is doing the boring spreadsheet work, not rewriting the science of measurement. The universal "Prof AI" training mandate is the more durable signal — when non-technical staff are required to be AI-literate, vendors whose products need a specialist to operate get displaced faster. Build for the planner, not the data scientist.

The Customer / End User — Two customers here. The advertiser wants transparency and a straight answer on what worked; Havas's "black book, not black box" pitch speaks directly to CMOs burned by opaque systems. But those same advertisers pour budget into Google's Performance Max and Meta's Advantage Plus — the most opaque AI black boxes going — and Sankar-King's answer is essentially "treat them as one slice of the funnel." That's a dodge dressed as strategy. The client's real problem is unchanged: they can't see inside the walled gardens, and no causal model fixes that. The measurement vendor customer, meanwhile, hears a higher methodology bar coming — a good thing if your incrementality math is real.

Where they part ways: The Market Analyst sees a savvy demand signal worth acting on; the Skeptic sees a vendor conflating a real workflow win with an unproven modeling category to sell a differentiator. The Operator splits the difference — the automation is real and shipping, the causal-model revolution isn't. And the Customer tension is sharpest of all: you can't sell "full transparency" while your clients' biggest spend sits inside black boxes you don't control.

What this actually hinges on: whether "large causal models" become a named line item in real measurement RFPs, or stay a boutique differentiator that a handful of holdcos market. The council leans skeptical on the category maturity but takes the underlying demand signal seriously: sophisticated advertisers genuinely are frustrated with nested MMM's limits, and that frustration will pull budget toward whoever can credibly claim causal, incremental measurement.

For vendors: don't rebuild your pitch around "causal AI" as a buzzword — the buyers can smell it. Do sharpen your incrementality methodology and your explainability, because the bar in the next RFP cycle is rising. For publishers and data partners pitching holdcos: know whether you're walking into an acquisition shop or a partner-integration shop, and tailor the integration story.

Impact of this specific episode is moderate and indirect — no M&A, no earnings, no regulation. It's calibration, not a trigger.

No high-conviction prediction this week.

The episode is a philosophy conversation, not an event with an observable outcome. "Large causal models enter mainstream RFPs" is directionally plausible but has no clean anchor date or falsifiable threshold I'd stake a call on — and the source is a vendor talking her own book, which is exactly the kind of signal the framework tells us not to launder into a confident prediction.

Listen to the episode →