AI Adoption Stalls When Measurement, Bidding, and Sales Don’t Agree on the Goal

When an AI-driven optimisation initiative underperforms, the post-mortem almost always looks in the wrong place. Teams audit the models, question the training data, and debate algorithm choices. Rarely does anyone ask the question that, in my experience, explains most of the failures: did the teams feeding, tuning, and selling this system ever agree on what it was supposed to optimise? 

I have spent over twelve years building large-scale optimisation and measurement systems for performance advertising, and I have led the effort to unify measurement, bidding, and go-to-market strategy across one of the largest advertising segments in the industry. The pattern I saw there repeats everywhere AI meets a complex organisation. The technology was never the constraint. The constraint was three groups of smart people optimising three different definitions of success. 

The Failure Mode Nobody Puts on a Slide 

Here is how it happens, and why it happens to sophisticated organisations rather than careless ones. 

The measurement team owns the definition of outcomes, and it optimises for rigour: more complete attribution, better data quality, cleaner methodology. The team that runs the optimisation systems owns the algorithms, and it optimises for model performance against whatever signal is available today. The go-to-market organisation owns customer adoption, and it optimises for what is easiest to explain and fastest to show results. Each team is competent, each is hitting its own targets, and each is pulling the system in a different direction. 

The result is quiet fragmentation. The models are trained on signals the measurement team considers outdated. The sales organisation drives customers toward configurations the optimisation team knows are second-best, because those are the ones customers understand. Measurement improvements ship but go unused, because nobody’s incentives depend on adopting them. From above, leadership sees an AI investment that keeps growing and adoption curves that keep flattening, and concludes the technology is not ready. 

The technology was ready. The organisation was optimising three different goals and calling it one strategy. 

Agreement Is Infrastructure 

The reframe that changed my own approach is this: a shared definition of the outcome is not a communication nicety, it is a component of the system. An AI optimisation loop connects a measured outcome, an algorithm that pursues it, and a channel that persuades users to adopt it. If those three are pointed at different targets, the loop is broken in a way no model improvement can fix. 

This is why the fix has to start before the automation layer, not after it. The most valuable work I have led in this space was not building a new model. It was conducting the unglamorous analysis across product data, revenue performance, and frontline feedback to understand where the definitions diverged, and then constructing a single north-star target that every function could recognise its own work inside. 

A good north-star target for AI optimisation has a specific character. It sits close to the real business outcome rather than a convenient proxy, so that improving it cannot be gamed by moving activity that does not matter. It is measurable with the signals you can actually collect, respecting privacy constraints and data realities rather than wishing them away. And it is adoptable, meaning the go-to-market organisation can explain to a customer, in one sentence, why moving toward it serves them. A target that fails any of these three tests will be silently abandoned by whichever team it fails. 

Sequencing: Why Measurement Comes First 

With a shared target agreed, the order of investment matters enormously, and most organisations get it backwards. The instinct is to buy the automation first, because automation is where the excitement lives. The discipline that actually compounds is the opposite: measurement first, automation second, diversification third. 

Measurement standardisation comes first because every downstream system inherits its definitions. Until the organisation records outcomes one way, with one taxonomy and one source of truth, every team’s automation will be trained on private data and produce private conclusions. Standardising measurement is slow, political, and utterly decisive; it is the step that makes every later step cumulative. 

Workflow automation comes second, and only where the standardised measurement already reaches. Automating on top of an agreed outcome definition means every efficiency gain reinforces the same goal, and every model retrained is retrained on data the whole organisation trusts. Automating ahead of measurement means encoding the fragmentation permanently, at machine speed. 

Diversification comes last: extending the proven loop into new channels, new formats, and new customer segments. This is where growth multiplies, but only because the earlier sequence made the loop portable. A measurement-and-optimisation loop that works end to end in one channel can be carried into the next one largely intact, and each new channel strengthens the shared dataset rather than fragmenting it further. 

The sequence is not merely tidy. It is what makes AI adoption compound instead of fragment, because every function’s progress lands in the same ledger. 

What Alignment Looked Like in Practice 

When I applied this framework across a business where measurement, bidding, and go-to-market had long operated independently, the change was structural rather than cosmetic. Objectives across product, sales, and operations were rewritten around the single north-star target, so a frontline conversation with a customer and a model-tuning decision made by an engineer were, for the first time, pushing the same direction. Adoption of the deeper measurement capabilities stopped being a side quest and became the path of least resistance, because it was the path everyone’s goals ran through. 

What convinced me most was not any single result but the shift in how decisions were made. Debates that had previously been jurisdictional, my team’s metric versus yours, became empirical: which choice moves the shared target. That is the cultural dividend of alignment, and no amount of model sophistication buys it. 

The lesson travels far beyond advertising. Every industry now has organisations investing heavily in AI-driven optimisation, and every one of them has a measurement function, an automation function, and a sales or operations function with historically separate goals. The companies frustrated with their AI results should audit their alignment before they audit their algorithms. 

The models will keep getting better on their own; the research world guarantees it. Whether your organisation can absorb that improvement is decided somewhere else entirely, in whether three teams can agree on one number worth optimising. 

AI doesn’t fail in the algorithm; it fails in the org chart. Alignment is the first model you have to build. 

Author bio:

Shrey Hatle is Director of Product, Performance Advertising at PubMatic, where he leads AI/ML-powered optimization, measurement, and performance advertising products. Previously at Google, he led large-scale advertising measurement and bidding initiatives serving millions of advertisers globally. His work spans AI/ML, programmatic advertising, attribution, automation, and product strategy.