Artificial intelligence could offer an answer to the persistent question of whether marketing adds value for an organization i.e. proving and predicting incremental revenue. But not just any use of AI, says Hitesh Wadhwani.
Business leaders have leaned on traditional measurement stack for years combining correlational analysis and limited experimentation to guide big decisions regarding all aspects of the enterprise, and marketing in particular. If a campaign runs and sales tick up, the two get combined into a slide before the next budget period.
But correlation is not causation! The core issue is these legacy measurement tools trap marketers in retrospective reporting, says Wadhwani, General Manager, Platform at Alembic. That gap between the two is exactly where business decision-making keeps going wrong, he says.
Enterprise environments are far too noisy for basic correlative analytics,” Wadhwani explains. “When you have simultaneous media spend, pricing shifts, brand campaigns, and macroeconomic factors occurring at once, seeing two metrics move together doesn’t tell you what actually drove revenue. Unless you understand the true underlying causal chain, you can’t reliably predict what will happen when you reallocate capital tomorrow.
While correlation can indicate what effects followed what actions, causality can dive deeper and explain what actually caused the outcome based on counterfactual simulations. Causal AI analysis can separate the results of complex enterprise activities and trace their ultimate effects on sales and revenue to inform business decisions.
“We enable enterprises to gather all their granular marketing and operational data and derive insights out of it in order to enable a real-time Causal AI decision engine,” Wadhwani says.
Bridging the cause-and-effect gap
Alembic offers organizations a platform to enable causal analysis for all their operations, but marketing is a particular beneficiary of it. For CMOs under growing pressure to justify their budgets to their CFO, understanding what truly drives outcomes has become business-critical.
Even as marketers have become more comfortable using artificial intelligence to inform decisions, most insight-gathering processes are hobbled by legacy measurement approaches. “One of the biggest gaps in legacy frameworks is that they surface insights that may not reflect the true drivers of growth, falling short of providing that missing link of true causality,” Wadhwani says.
Three major challenges make it difficult for enterprises to deploy real causal analytics, he explains.
First is “Resource-Intensive testing” as it is costly to run true causal analysis at scale combining traditional tactics like incrementality testing and correlationals models. It only works if management can really narrow the focus only to a limited set of high priority channels with specific hypotheses to test.
“It’s not possible to run and design an always-on experiment for every part of your business. It will cost an enterprise a huge amount of resources to design these experiments on an ongoing basis,” says Wadhwani.
Even then, there is a “High-dimensionality bottleneck” as legacy measurement tools collapse with higher complexity and the underlying math of these models breaks down after incorporating more than 25-30 factors. Modern enterprises operate across hundreds of interconnected channels, campaigns, pricing shifts, and market signals and for a business to really make good decisions, they need to get insights at this granularity,” Wadhwani explains. “If a system can only evaluate only 25 aggregate factors at a time, it is forced to oversimplify reality, missing critical drivers of growth.”
Lastly, Latency plays a critical role as the current dynamic externalities require businesses to make real-time decisions to stay effective. As Wadhwani explains, it can be three months to sometimes one year for actionable insights to emerge through traditional methods.
“You’re essentially looking into a rearview mirror and not planning for the future, because by the time you’re getting insights, your campaign might be completely over,” he says.
The alternative requires a model that can look at all the factors that can affect an outcome across time. To solve this, Alembic models the enterprise using spatio-temporal causal graphs and patented Spiking Neural Networks (SNNs) approach – a system capable of detecting emerging market signals and calculating complex cause-and-effect relationships continuously, without waiting months for batch re-training. This model presents every event in an enterprise dataset as a node and generates a holistic causal graph after running millions of simulations, Wadhwani explains.
“It’s about getting a digital twin of your enterprise on which you can also run these millions of counterfactual simulations in real-time to see the exact downstream impact and get to the potential future answer before you commit capital” he says.
LLMs fall short
Creating these digital twins wasn’t feasible before, due to the amount of compute power necessary, but the processors powering artificial intelligence models have made it more accessible. Continuous causal discovery across massive enterprise datasets was historically computationally intractable. Alembic has partnered with Nvidia to use its infrastructure, including its NVL72 supercomputers and Vera Rubin platform, unlocking the ability to evaluate billions of halo interactions continuously across the full causal graph rather than in slow batch processing cycles.
The ability to deploy this level of compute power to granular data sets “is the foundation which was previously missing,” says Wadhwani. “Running these counterfactual simulations is now only possible because of this combination.”
While broad AI adoption has helped manage large data sets, traditional Large Language Models (LLMs) fall short of delivering mathematical causality, Wadhwani emphasizes.
LLMs serve a valuable role to surface insights as natural language interface layers allowing executives to ask natural questions and get plain-language summaries,” says Wadhwani.
However, as their name implies, LLMs are text-prediction models, not deterministic causal calculation engines. They are trained on text, not numbers.
Asking an LLM to evaluate complex numerical datasets for true causality carries a massive risk of hallucination. You cannot rely on a text generator to allocate a multi-million dollar marketing budget.
Instead of relying on LLMs for math, Alembic’s solution is to create a spatio-temporal graphical model trained specifically for that enterprise from the ground up, not an AI agent, says Wadhwani. “Only then can the model provide you with the reliability to make a C-level decision,” he says.
But Wadhwani adds these new innovations aren’t meant to replace the enterprises’ data science teams. Many Fortune 500 companies Alembic works with have extraordinary internal data science teams conducting rigorous experimentation and trying to solve some of the same cause-and-effect gaps through traditional media methods; Alembic just helps them solve them much faster as the platform acts as a force multiplier for them, he says.
“It’s a very close collaboration as we don’t only treat key signals coming from in-house teams as a critical context layer, but our causal graph uses those empirical data points as structural calibration anchors ,” says Wadhwani.
The two teams partner together, feed the ground-truth data directly into the graph to continuously calibrate and validate the insights from the causal analysis, especially if the in-house teams can provide additional context to the raw data, such as historical incrementality tests, which provide useful reference points for the causal links. Wadhwani adds that this approach is built to empower, not replace, internal enterprise data science teams.
Ultimately, all the work to solve causality at enterprise scale comes down to a question of uncovering true cause and effect, says Wadhwani, quoting Alembic’s motto: “The science behind why.”
