Why AI Winners in CPG Are Built Around Repeatable Problems, Not Vague Functions
By Simran Suri and Ryan Caldbeck
Most AI strategies in CPG still start with a broad function. AI for marketing. AI for sales. AI for operations. It sounds logical because functions are how companies are organized and how budgets are allocated, but this framing is also why so many AI initiatives stall after the demo.
Functions are not problems. They are bundles of problems, often pulling in different directions, owned by different stakeholders, and lacking a single definition of success. AI doesn’t fail because the models are insufficiently powerful. It fails because the problem is poorly defined. If humans can’t agree on the objective, AI won’t infer it.
The AI winners don’t start with functions. They start with specific, repeatable decisions surrounded by abundant unstructured information. The same decision, made hundreds or thousands of times, with similar inputs and clear stakes, is the sweet spot. Unstructured data only becomes valuable when the decision repeats, and patterns can be identified, otherwise known as data network effects. Otherwise, it remains noise.
The earliest and most durable AI wins are not trying to “transform” entire functions. They are attacking specific, repeatable workflows, unlocking previously disorganized and unstructured data, which ultimately unlocks broader platform opportunities.
In finance and accounting, AI has gained real traction in areas like invoice processing, three-way matching, expense auditing, accruals, and anomaly detection. Companies like Vic.ai focus narrowly on automating accounting judgments around invoices and GL coding. AppZen applies machine learning to expense reports and AP audits, flagging policy violations and leakage before payments go out. These tools don’t claim to reinvent finance. They take one decision that happens thousands of times a year and make it faster, more consistent, and more reliable by standardizing and analyzing disparate data sources
Legal followed a similar trajectory. AI did not succeed by “doing legal.” It succeeded by focusing on repeatable contract workflows, with product wedges ranging from clause extraction and deviation detection to approval routing. Platforms like Ironclad and Lexion gained adoption by living inside the contract lifecycle, where the same patterns appear across thousands of agreements, mostly embedded in unstructured documents.
Operations has seen the same dynamic. AI has been effective in ticket triage, exception handling, and compliance workflows—areas where decisions repeat, inputs are messy, and outcomes are measurable. These systems don’t rely on pristine data. They learn by observing the same type of work over and over again, letting their analysis intelligently drive expansion across other workflows and functions.
These categories share a critical set of properties. The decisions recur. The inputs are unstructured (PDFs, emails, attachments, free text) and the feedback loops are tight. Time saved, errors avoided, dollars recovered, risk reduced – value compounds through data network effects and process power, or process improvements that drive operational excellence over time.
Marketing, by contrast, is not discrete. While performance and attribution are objective, brand and creative are subjective. They require a human touch that AI cannot fully automate, or (in some cases) even optimize. Even strong teams disagree on the objective decisions, and those objectives change constantly by channel, by quarter, by leadership, and by market conditions. When so many decisions are bespoke, learning cannot compound. General-purpose AI can generate output, but it struggles to become authoritative because ambiguity and subjectivity are manageable for humans but corrosive for models. This is why invoice reconciliation has moved faster than campaign planning, why contract review has advanced further than brand strategy, and why operational exception handling is ahead of “AI for marketing.” These problems are narrow, frequent, and expensive. That combination matters more than how visible the function is.
This lens is especially important in CPG, because few industries generate more repeatable decisions inside messier workflows. CPG isn’t slow because teams lack sophistication, but because core work still runs through email, PDFs, and human judgment. In some cases, quotes arrive mid-thread with shifting assumptions, with specs changing quietly between versions. In others, COAs show up as attachments with inconsistent naming and certifications expire without warning. No matter how big the brand or experienced the team, a single missing document can stop a production line. These aren’t edge cases; they’re the operating rhythm of the industry. They require systems that can read messy reality as it happens and turn it into structured truth.
This is where workflow ownership becomes decisive—but only after focus. Owning the workflow means seeing the same decision again and again, capturing raw inputs by default, and learning from outcomes rather than intentions. You don’t learn by seeing everything. You learn by seeing the same thing repeatedly. Once decisions flow through a system consistently, compounding begins. More workflow creates more truth. More truth sharpens recommendations. Better recommendations drive better outcomes. Better outcomes pull more work into the system – this time, across functions. Repetition becomes the moat.
In CPG procurement and quality, companies like Waystation are applying this same playbook: focusing narrowly on repeatable supplier and compliance decisions, extracting structure from emails, specs, COAs, and certifications, and building a trusted record that procurement, QA, R&D, operations, and finance can rely on. Companies like DayZero, Jampack AI and Moselle are doing the same in CPG accounting, wholesale distribution and inventory management, respectively.
The takeaway is simple but often missed. If you are building or buying AI in CPG, stop starting with broad departments. Start with a single question: what specific decision do we want to make faster, safer, and more consistently every week, using information we already have but cannot reliably use or understand? Find the repeatable decision. Find the unstructured exhaust. Build the loop. The initial single decision the AI focuses on becomes the wedge that unlocks clean, structured data and its network effects, which can be used to drive greater stickiness, platform effects and retention.
That is how AI moves from interesting to inevitable. The future belongs to systems that master one decision before claiming many.
Simran Suri is a consumer investor at Maveron. Ryan Caldbeck is the founder and CEO at Waystation.