1. Analytics, forecasting, and measurement have been the core of data investment for twenty years. You’re describing a layer that sits above all of that. What broke in the existing layer that created the need for something beyond it?
A decision layer is the part of the stack that weighs the options in front of a business, estimates what each one will cause, and records who owns the result. Everything underneath it was worth building. Twenty years of investment in data platforms, BI, and forecasting means large companies can now see what happened and what’s likely to come next.
The weak point is the handoff from that information to an actual choice. A forecast tells you what happens if nothing changes, but every budget meeting is about changing something. So people work out the real question, “what happens if we do this instead?”, by hand in the meeting itself, often from numbers that arrive after the plan is half-committed.
Gartner found in 2024 that only 48% of digital initiatives meet or exceed their business outcome targets. I don’t read that as an execution failure. In most of the programs I’ve seen, the insight was there, but acting on it was nobody’s job.
2. The organizations that are earliest to need a decision layer above their analytics are the ones whose decisions are most complex. Which business functions are feeling that need most right now and what does it look like when the existing analytics layer fails them?
The need shows up first wherever one decision moves several teams’ numbers. Think of the annual growth plan that ties marketing, sales, and finance together. Each of those functions has good analytics of its own, so the trouble starts at the boundaries, where each model sees only its part of the decision.
BCG’s 2026 planning research found that 78% of planning leaders rank forecast inaccuracy and misalignment among their three most pressing internal challenges. A typical failure would look like this: marketing commits to a heavier spring push, but the supply forecast, built on last year’s shipments, can’t see it.
Shelves run thin in the weeks the media lands, and finance books the quarter as a media miss. Each forecast was defensible on its own terms, but the decision failed because nobody owned all of it. Without an owner, choices like this drift to the most senior opinion in the room.
The same holds at the top. McKinsey found in 2025 that companies with a single customer- or growth-oriented role on the executive committee see up to 2.3 times more growth than those with several such roles.
3. Marketing measurement fragmentation is not a new problem. You’re describing that fragmentation is now producing conflicting answers to the same question, not just different metrics. What does it look like when an organization has to make a budget decision with three measurement systems pointing in different directions?
It stops being an analysis and becomes a negotiation. Take this question for instance: should we keep funding the summer connected TV flight? The marketing mix model shows a modest, steady lift, platform attribution gives it almost no credit because few viewers clicked, and a geo test says it worked in one region only.
Each of the three is doing its job. MMM has earned its place over decades and reads long-run effects across channels better than anything built on clicks, and the geo test is the cleanest evidence in the room. The trouble is that each one rests on a different assumption about what would have happened without the flight, and no one can settle which assumption holds from the reports alone.
So the CMO brings the reading that supports the plan, and the CFO discounts all three. The CMO Survey rates the CMO-CFO partnership at just 4.5 out of 7 for building the business case for marketing spend, and says that score has barely moved in four years. In practice, the line that gets cut is the one with the weakest advocate, not the weakest return.
4. A CMO sitting with three different measurement reports that give three different answers to “did this campaign work” has an immediate operational problem. What does causal decisioning change about that situation and what does it require the organization to have in place?
Causal decisioning replaces the three answers with one. The CMO asks what the campaign actually added and gets a single estimate, with price, promotions, distribution, seasonality, and every other channel that quarter accounted for in the same model. Because one model estimates all of those drivers together, a price cut that ran the same week gets credit for its own lift, and the media keeps only what it earned.
That changes the CMO’s role in the room. They arrive proposing a decision, with what each option should cause and a range around it, before any money moves. More reports can’t get you there, because the disagreement comes from how each report is built.
It does ask a lot of the organization. Finance has to help define the model, including what counts as revenue and margin, so the answer belongs to both functions. You also need data beyond media, like pricing, trade spend, distribution, and inventory, and the willingness to retire the old reports once the shared model has earned trust.
Getting there takes some hard conversations between marketing and finance, and I’d be wary of anyone who says it doesn’t.
5. Trust in an AI decision system is built through evidence accumulated over time. What does that trust-building process look like in practice and what breaks it when it goes wrong early in a deployment?
Trust comes from forecasts people can check: the system writes down what it expects, with a range, before the outcome, and the team checks it afterward. That’s why I’d start with frequent, easily reversed decisions in a recommend-only mode, where a visible track record builds quickly. Authority then widens one kind of decision at a time.
Early trust usually breaks on a confident call that contradicts what the team knows. Say the system advises cutting support for a product after a sales dip, when the category manager knows the dip came from a retailer delisting two pack sizes that month. The model never saw the delisting, and the team will retell that miss in every review for a year, because people forgive a colleague’s bad forecast faster than a machine’s.
Numbers that shift between runs without explanation do similar damage, and so do promises of autonomy before there’s any evidence. The fix is candor. A causal model is weakest where history is thin, such as a new channel or a market after a structural break, and it should tell people so.
6. Regulators in financial services, healthcare, and other sectors are moving toward requiring explainability and human accountability for automated decisions. How does that regulatory direction shape the architecture of a decision AI system?
Regulation is settling a design question companies should have settled on their own. Explanation and accountability have to be designed into the system from the start, not added afterward. The deadlines keep moving, but the direction is consistent: the person affected should get an explanation, and a named person should be able to override the output and answer for it.
Gartner’s warning this year, that “regulated industries and governments cannot rely on opaque ‘black box’ systems for consequential decisions,” applies just as well to any decision a company may one day have to defend.
That rules out the common shortcut of bolting an explainer onto a correlation-based model, which can only approximate what the model did after the fact. In a causal model, the explanation is the model itself: which drivers moved, by how much, and what the alternative would have produced.
So every recommendation should carry its reasoning in terms a business owner can challenge, with someone’s name on the approval before anything executes. Building that in at the start costs far less than piecing together a decision’s reasoning months later from logs nobody designed to explain it.
7. The decisions that are easiest to automate are not necessarily the decisions where automation creates the most value. And the decisions where automation creates the most value tend to be the ones with the consequence for getting it wrong. How do you think about that tradeoff?
Measuring automation in hours saved pulls companies toward the decisions that matter least. Easy decisions are frequent, well understood, and cheap to get wrong, which is exactly why automating them rarely moves the P&L. The valuable ones look different: how much promoted stock to send each region, or how to pace a quarter’s media budget as demand shifts.
Those decisions are interconnected and expensive to miss, so companies keep them manual, and that has a cost too. By the time someone approves a budget shift, the window it was meant for has often closed.
I’d rather lower the cost of being wrong. Let the system act within a range people set, start narrow, keep a person able to adjust what it does, and give it more room each quarter its calls hold up.
Pattern-based models struggle with these decisions, because they’re rare and the situations are often new. A causal model copes better, since it learns how each driver moves the result and can recombine those effects for a new situation, while saying where its evidence runs thin.
8. There is a significant accountability gap between AI that surfaces a recommendation and AI that executes a decision. When a human chooses to act on an AI recommendation, the human carries the accountability. When AI acts autonomously, accountability becomes ambiguous. How do enterprises need to think about that accountability structure before they hand over decision-making authority?
The gap looks ambiguous because companies still assign accountability one decision at a time. That works while a person approves every call. Once a system is making hundreds of small decisions a day, nobody can review each one, so accountability has to move to the rules the system follows and the person who sets them.
Supply chains already run this way. A head of supply chain never sees most purchase orders, but they own the reorder policy and answer for what it produces.
An AI system needs an owner like that before it gets any authority: someone who has written down what it’s trying to achieve and which tradeoffs are off the table. In my experience, if that person can’t explain the policy to the board in a couple of sentences, the system isn’t ready.
The last step is a record with every action, showing the policy it ran under and the option it passed over. When something goes wrong, that record leads back to the policy and its owner. Without it, blame lands on whoever happened to be closest to the system.
9. As CEO, you’ve made a specific bet on what enterprise AI looks like at the next level of maturity. What does the enterprise that has fully realized that vision look like from the inside?
The enterprises that pull ahead will run on one causal model of the business, one that finance, marketing, and supply chain trust enough to act on. That’s the bet behind DATA POEM. Day to day, those companies are calmer than you’d expect.
Planning meetings start from the same numbers and spend the time on choices. A supply planner sees the demand a campaign will create before it launches. When the CFO and CMO disagree, they disagree about something specific, like in-quarter margin against brand equity that pays back over years, with the cost of each path in plain view.
Analysts work on new questions, because last month’s reports already agree. People well beyond the analytics team use the model, so when it’s wrong, they can see why and fix it. MIT CISR’s 2025 research is consistent with that: where at least a third of employees use the company’s data, money made from it accounts for 15% of total revenue, against under 5% where fewer do.
What changes most is a habit. Leaders write down what they expect before they decide and let the record show when they were wrong. That’s how any decision layer gets better.
About Bharath Gaddam
Bharath Gaddam is the Founder and CEO of DATA POEM, the Enterprise Decision AI company he founded in 2019. Before founding DATA POEM, he worked in the Omnicom and IPG agency groups on brands including McDonald’s, Puma, Nike, Domino’s, Dove, and Volkswagen, and was involved in launch and media strategy for brands including ITC.
About DATA POEM
DATA POEM is an Enterprise Decision AI company, founded in 2019, whose platform, POEM365, gives enterprise leaders one unified causal model of what drives growth across the business. POEM365 is a Large Causal Model built on FOUNT’s Large Causal Architecture.
