AI Do's and Don'ts for CPGs: Balancing Automation with Human Expertise

In my last blog post, we explored what AI can’t do when it comes to trade and deduction management. This post flips (and broadens) that conversation, focusing on what AI CAN and SHOULD do for CPG brands in this space - but not without guardrails. Yes, there's tremendous opportunity for brand teams to work faster and smarter than ever, but bringing it to fruition requires clarity on where to automate, where to stay cautious, and where human judgment must rule. Here, I summarize a list of AI Do's and Don'ts for CPG leaders considering AI for trade and deductions. Hopefully it helps in navigating these important scenarios and decisions for your business.

TL;DR

• DO use AI aggressively to accelerate repetitive, data-heavy work. Matching, analysis, forecasting, pattern recognition, summarization, and anomaly detection are all areas where AI can create real leverage.
• DON’T confuse speed with accuracy. An AI-generated answer still needs validation when the outcome affects trade spend, accruals, financial reporting, or customer relationships.
• DO use AI to make your people better. The biggest opportunity is giving experienced people better information faster—not removing expertise from the equation.
• DON’T expect technology to compensate for bad data or broken processes. AI can amplify a strong operating model. It can also amplify a bad one.
• DO know where accountability belongs. My rule is simple: let AI accelerate the work, but keep people accountable for consequential decisions.

In my last post, I made the case for why AI alone won't solve deduction management. You might say that the tone and title of that post sounded a bit negative against AI - but that couldn't be further from the truth. At Promomash, we've been incorporating AI and machine learning into our solutions for years. We're continuously investing in it for the benefit of our clients, using it internally across all our departments, and actively using it to build the next generation of our platform. Oh, and personally? I don't know where I'd be without the AI tools I count on each day to get work done and move Promomash forward.

All that is to say that we've seen AI work brilliantly in some areas and fail spectacularly in others. When it comes to CPG trade spend and deduction management, extra special care is required because financial statements are directly impacted. One mistake can compound into sizable financial damage. I truly believe at the end of the day that AI is going to positively impact CPG. However, I’m also convinced that brands that invest blindly are at risk of learning very expensive lessons.

McKinsey has studied more than 140 digital and AI use cases across the CPG value chain and estimates that generative AI could add another $160 billion to $270 billion in annual profit to CPG companies globally on top of the impact of traditional AI. The opportunities stretch from consumer insights and demand forecasting to product development, marketing, sales, supply chain, and back-office operations.

For CPG leaders, there's no question about whether to use AI. Not using it is not an option. But when it comes to deciding where to use it, how much authority to give it, and when a human decision is still needed...there are still many questions. These are the questions I hope to answer with my list of AI Do's and Don'ts below.
 

DO: Let AI do the first pass.

If there's an area where I'm most bullish on AI, it's this one. There are plenty of jobs in CPG that require someone to sift through an enormous amount of information before they can actually make a decision. And guess what? AI is fantastic at shortening that process.

Take deduction matching. A person may need to examine a deduction, search through promotional plans, compare dates and amounts, identify the likely program, and determine which promotion the deduction belongs against. AI can look across those inputs and say: “I think this deduction most likely belongs to this promotion.”

That's incredibly valuable. What might have taken someone hours or even days can happen in seconds. Elsewhere in CPG, AI can help analyze consumer feedback, synthesize research, interrogate large datasets, generate initial forecasts, summarize contracts or documents, and surface likely explanations for anomalies.

That's exactly what it should be doing, BUT... 

DON’T: Treat the first pass as the final answer.

Here's where people get into trouble. AI gives you an answer quickly, clearly, and confidently. That doesn't make the answer correct. Suppose you have three similar promotions with overlapping dates and similar spend. The AI selects Promotion A. It looks reasonable. But an experienced deduction professional recognizes something in the distributor's documentation that points to Promotion B. That small distinction matters.

If you automatically accept Promotion A, you haven't just made a matching mistake. You've potentially changed the remaining promotional balance, the accrual, the GL coding, and eventually the story your reporting tells about the effectiveness of that promotion. AI getting you 80% or 90% of the way there can be tremendously valuable. But the closer a decision gets to your financial statements, customer relationships, or significant business commitments, the more important that final 10% becomes.

DO: Use AI to find patterns people can't. 

Humans are great at context. Computers are great at scale. We should design workflows around that distinction.
Give AI thousands of deductions, invoices, promotional records, sales records, customer interactions, or consumer signals and ask it to find patterns. For example:

• Where are deductions increasing?

• Which customers are producing unusual activity?

• Are certain deduction codes suddenly appearing more frequently?

• Is promotional performance deviating from historical expectations?

• Are shipped volumes and sold-through volumes telling different stories?

• Is a recurring operational issue hiding inside hundreds of individually valid deductions?

Deloitte, for example, has highlighted the potential for AI in trade promotion management to continuously reconcile claims, identify unusual patterns, monitor fund allocation, compare shipped versus sold-through volume, and flag issues before they become larger financial problems. That's the kind of AI adoption I want to see more of.

DON’T: Ask AI to understand context it doesn't have.

Pattern recognition and business judgment are not the same thing. Imagine a system detects a spike in deductions from a distributor. That's useful. But why did it happen?

• Did the distributor introduce a new program?

• Did its deduction codes change?

• Was there a new contract?

• Was there a promotion that wasn't documented properly?

• Did a broker agree to something outside the normal process?

• Is the deduction valid but exposing an operational issue?

• Does the brand have a history with this distributor that changes how the issue should be handled?

The data can tell you something happened, but understanding what it means may require information that isn't in the dataset at all. That's where experienced people with the right skill sets, institutional knowledge, and context come in.

DO: Automate repetitive work.

No one should get an award for spending hours doing something a machine could safely do in minutes. If a task is repetitive, rules-based, high-volume, and relatively low-risk, we should be asking whether AI can take more of it. Deloitte makes a similar case specifically for TPM, describing routine and repetitive processes as strong candidates for AI while freeing sales and finance teams to spend more time on strategy, relationships, and growth.

That's not replacing people for the sake of replacing people. It's improving the job. I don't want an experienced deduction professional spending most of the day searching for documents or manually comparing hundreds of records. I want technology doing the searching, sorting, aggregating, and suggesting. Then I want the expert doing what they're actually good at: investigating, interpreting, deciding, communicating, and solving.

DON’T: Automate a consequential decision.

Quite often, the conversation around AI automation glosses over a very important distinction: automating repetitive processes vs automating decisions. There's a difference between automating predictable, repetitive, low-risk work and automating consequential decisions that humans should be accountable for.

For example: Should AI help identify which promotion a deduction probably belongs to? (Absolutely)

But should it make the actual decision of choosing which promotion to match the deduction to without the appropriate controls in place, simply because its confidence score is high enough? That's a different question.

Another example: Should AI flag a potentially invalid deduction? (Yes)

But should it automatically blast a distributor with disputes without understanding the validity of those deductions, and the relationship with that distributor? I'd be very careful.

The stakes should determine the level of oversight. The more consequential the outcome, the stronger the case for human review.

DO: Use AI to improve forecasting and decision-making.

Deduction management is only one piece of the opportunity. Some of the most exciting applications of AI in CPG happen before the transaction ever occurs. Forecasting is an obvious example. CPG companies are constantly trying to answer questions about future demand, inventory, promotions, pricing, assortment, and consumer behavior. AI can analyze more variables and more historical information than any individual planner could reasonably process. And we're already seeing evidence of what that can accomplish.

McKinsey describes one personal-care company that used digital and AI capabilities to incorporate internal and external data, improving forecast accuracy by 13%, reducing product shortages by 40%, and decreasing inventory by 35%.

That's real value when AI can provide humans a better starting point for decisions.

DON’T: Pretend a forecast eliminates uncertainty.

A better forecast is still a forecast. Consumer behavior changes. Retailers change programs. Competitors react. Weather happens. Supply chains break. Customers make unexpected decisions. And CPG data is notoriously imperfect.

The model can tell you what it thinks will happen based on the information available. But an experienced operator still needs to ask whether the output makes sense in the real world. The goal should be to provide the information needed to make the best judgement - not remove judgment from planning altogether.

DO: Make your experts dramatically more productive.

This might be the AI use case I'm most excited about. We spend so much time asking how many people can AI replace. But I think that's the wrong question. Instead, ask how much more valuable AI can make our best people!

• Give a great salesperson better customer intelligence.

• Give a deduction expert instant access to the documents and patterns relevant to an exception.

• Give a finance leader earlier visibility into an accrual problem.

• Give a marketer faster access to consumer insights.

• Give a supply-chain planner better demand signals.

• Give a product team the ability to synthesize thousands of pieces of consumer feedback.

Now that's leverage.

McKinsey has found potential AI value across virtually the entire CPG value chain, including consumer insights, demand shaping, innovation, marketing and sales, autonomous planning, and support functions. The common denominator isn't eliminating humans. It's increasing what humans can accomplish.

DON’T: Use AI to replace expertise you never had.

This one worries me. Let's say a growing CPG brand doesn't have anyone on their team who understands deductions deeply enough (believe me, I've talked to many). The brand chooses to invest in an AI deduction management platform promising to automate the entire process - instead of hiring or outsourcing a deduction expert.

Who's checking the AI?

If no one on the brand team knows what good deduction management looks like, how will anyone recognize when the system gets it wrong? Will they be able to rely on the vendor team for support to resolve the issue? If the errors continue unchecked, and soon the system is regarded as a source of truth because no one knows any better, what does that mean for the company's accounting and financials?

AI plus expertise creates leverage. But AI without expertise creates false confidence. Very different outcomes.

DO: Make AI more useful by fixing your data & processes.

If there's one thing nearly every AI conversation eventually comes back to, it's data. The better your inputs, the more useful AI becomes. For deduction management, that means having the promotional plan. The contract. The expected spend. The correct customer information. The historical deduction data. The appropriate coding. The documentation. For other areas of CPG, the specific inputs change, but the principle doesn't. You need reliable information and defined processes.

This isn't just my view. Deloitte identifies domain-specific knowledge, quality data, integrated workflows, clear business objectives, human oversight, and governance among the prerequisites for successfully applying agentic AI to trade promotion management. Technology doesn't eliminate the need for operational discipline; it amplifies the return you get from having it.

DON’T: Put AI on top of a broken process.

This is probably the simplest rule on the list. If you don't understand the process you're trying to automate, don't automate it yet. I've said before that if you can't manage your trade process reasonably well in Excel, buying a six-figure TPM platform isn't magically going to fix it. AI doesn't change that. A bad process running faster is still a bad process.

Bad data processed faster is still bad data. Unclear ownership automated at scale is still unclear ownership. Before asking what AI can automate, make sure you understand what the process is supposed to accomplish, who owns it, what good looks like, and where the exceptions occur. Then automate intelligently.

DO: Embrace AI and its possibilities.

Although I mention a lot of "Don'ts" here, I don't want my position to be misunderstood. CPG brands should embrace AI. As I mentioned earlier, we've embraced it at Promomash for years now. I'm using it personally. We're building with it. We're finding new ways to incorporate it into how work gets done. And the technology is going to become exponentially more capable.

CPG leaders clearly see that potential. Deloitte's 2026 survey of 200 retail and CPG executives found that 82% planned to increase their AI investment over the following 12 months. But the same research found enterprise-wide AI deployment remained in the single digits - evidence of a significant gap between enthusiasm and mature execution.

That's where I think the real conversation needs to happen. It's not about choosing between AI or humans. It's more about choosing WHAT the AI does, HOW it does it, WHO owns what, and how to make the combination of human experts + AI better together than either could be alone.

DON’T: Embrace AI for the sake of AI.

This is the hill I'll keep dying on. The goal isn't to use AI as a vanity metric, signal to the market, or simply because leadership asked you to. The goal is to build a better business, and AI offers many ways to assist in doing that.

If AI can reduce four hours of work to four minutes, use it. If AI can identify a pattern no person would realistically find, use it. If AI can improve your forecast, summarize thousands of records, surface an anomaly, suggest a promotional match, or give an expert better information faster, use it.

But if the decision requires context the model doesn't have, financial accountability it can't own, or a relationship it doesn't understand, put a qualified person in the loop. That's responsible adoption of AI.

The CPG brands that understand this distinction are going to get far more value from this technology than the ones racing to automate everything first.

Let AI accelerate the work. Let people own the outcome.

My view on AI ultimately comes down to one principle: AI should make experts more powerful, not make expertise optional. Use technology and machines for scale and speed. Use them to search, sort, summarize, detect, predict, and recommend. Then use people for context, judgment, relationships, accountability, and decisions where being wrong actually matters. That's the opportunity I see for AI in CPG. Smart people, equipped with better tools, spending more of their time on the things worth thinking about.