Where AI actually fits in CPG operations.
What we believe about AI in CPG operations — and what it means for how we work.
Most mid-market CPG companies don't have an AI strategy, and most of the firms selling them one are selling the wrong thing. CEOs are anxious about falling behind, vendors are pitching strategy decks, and consultants are walking out of the building with retainers and no execution. Meanwhile the operations team is still hunting for data in the morning and reconciling spreadsheets in the afternoon.
We've been in those operations seats. We've built the systems. We've watched the AI projects that worked and the ones that didn't.
Our frame is the operations stack — tools, agents, and people. The four claims below are how we think about where AI earns its place across it.
You can't keep up with AI. That's not the goal.
AI is changing fast enough that nobody can stay current. New tools every few weeks, most of which won't matter in six months. The CEO trying to have an AI strategy is chasing a target that moves faster than they can read. But keeping up with AI isn't the work — specifying the work is. Most operations problems are a set of inputs, a process, and an output that nobody has ever described in writing. No AI tool, current or future, can automate a process that hasn't been specified. The companies that get value from AI start there. The ones chasing tools first burn quarters.
Stop trying to keep up. Start specifying the work. The tools will follow.
Operations expertise leads. AI follows.
The companies getting real results from AI didn't start with an AI strategy. They started with a clear operational problem and used AI as one tool among several, competing against off-the-shelf software, process redesign, and training before earning the build. The firms selling AI transformation to mid-market CPG pitch the inverse: strategy first, problem later, six-figure engagement, six months later no deployed AI. We start the other direction. The Operations X-Ray identifies where money is leaking, then asks where AI fits, where it doesn't, and where a different solution is cheaper. Some engagements end up AI-heavy. Some don't. Both are correct.
If a firm is selling you AI before they've understood your operations, you're being sold something that won't deploy.
Most AI failures aren't AI failures.
They're process failures and data failures that AI can't fix. A demand forecasting model trained on incomplete sales data isn't going to outperform the spreadsheet — it's going to be wrong faster. An AI agent that classifies invoice exceptions can't function if PO data and receipt data live in unreconciled systems. The model isn't the bottleneck. The data and the process are. Before we recommend AI for any operational function, we score how ripe the underlying process is for automation. Some score high enough to build immediately. Some get the process and data work first; the AI comes later, when it has a chance of working.
The work is almost always upstream of the model. That's where the engagement starts.
Build, buy, or train — not always build.
Some operations problems get solved by training the team to use existing tools well. Some by buying off-the-shelf software. Some by building a custom system. Most consulting firms push toward build — building bills more than recommending. Most software vendors push toward buy. Most training firms push toward train. Each is right sometimes and wrong the rest of the time. The honest version requires being able to recommend any of the three. We get paid for the diagnostic and the engagement that follows. We don't get paid more for recommending more building. That's deliberate.
Recommend the cheapest option that works. That's how you earn the next engagement.
If this matches how you'd want an operations partner to think, the Operations X-Ray is where the work starts.
30 minutes. No pitch deck. Just a conversation about your operations.