Podcast: AI price optimization - what it actually automates (and what it doesn't)

AI price optimization is on every vendor deck this year - AI-powered pricing, agentic pricing, AI everywhere. Ask a retailer whether they use AI in pricing and the answer is yes. Ask what it changed last Tuesday and the answer gets quieter. This episode is the practical version: what AI actually does for a pricing team today, what it still cannot do, and how to tell the two apart before you buy anything.

In episode 5 of the Retailgrid Podcast, Retailgrid founder Maxim separates the four words that get used as synonyms (AI, machine learning, LLM, data science), walks through where the machine is already better than a person, where the human still has to sign off, how to move from manual to propose-and-approve to full auto without losing control, and closes with a lightning round on five myths about AI in pricing.

AI, ML, LLM and data science are four different jobs

AI is the umbrella: anything that decides, predicts, and helps a human do cognitive work. Machine learning is one specific technology under it - it learns from data and improves, and it is the workhorse of forecasting and price optimization. A large language model is another technology: the thing you talk to in an AI assistant, a neural network trained on billions of data points that answers like a person. Data science is everything related to data - any mathematical method counts. When a vendor says "our AI", it is worth asking which of the four they mean.

Can AI replace the analyst?

At least for now, no - but it takes a lot off the desk. AI reads huge data sets quickly, every row, and summarizes and analyzes them. It is very good at matching, because an LLM has a large capacity to find similar products, which makes it useful for product matching and price matching. It handles unstructured analysis well: propose, rank, explain. And it never gets bored with routine work.

What it cannot do is define the strategy. Humans still decide the optimization goal, what the business wants to achieve, and the long-term direction - that context is given to the AI, not discovered by it. AI also does not know the details: the conversations in meetings, the negotiations with suppliers, the situations specific to your business. And in the end it comes down to responsibility. AI helps, but AI is not responsible. When something goes wrong you can say "the AI suggested it", but you will still be the one in the responsible position.

Three use cases where AI price optimization is already better than a person

• Collecting and matching data. AI is changing the whole price monitoring and market intelligence market. You no longer need a room of people manually matching products; the machine does it better at scale. You still control it and put guardrails around it, but it is a game changer.
• Classification. Product roles - which SKUs are KVIs, traffic drivers, profit drivers. This is analysis rather than constrained optimization, and AI does it well, as long as the context it is given stays limited.
• Elasticity. Yes and no. Machine learning is used a lot here; the LLM is not really used for the calculation itself, but it helps with grouping, clusters and demand groups. Classic math, regression models and machine learning work, and the LLM can improve them. The combination is what gives real power in elasticity and demand forecasting. For the calculation itself, see the episode on how to calculate price elasticity.

Guardrails, hallucinations and why you keep the sign-off

Trying to run pricing inside a chat assistant directly is where it goes wrong. As the context gets bigger, at some point the model starts making mistakes, sometimes stupid ones, and those mistakes can cost real money. So AI belongs in a safe environment: guardrails, and an infrastructure that prevents the mistakes from reaching a live price. Do not hand all your data to an LLM and expect a deterministic machine - it is not one.

The practical rule is human in the loop. At least at the beginning, approve everything, check and double check, and understand how it works. "This is the model, it thinks this way" is not an answer a business accepts. You need full control.

Manual, then propose and approve, then full auto

Mass-market retail is a red ocean - many competitors selling the same products, where price, assortment, promo and brand all matter and the retailer has to be proactive. That is exactly where AI helps most, and also where the autonomy question matters most. The path is always the same: start manual, move to propose-and-approve ("here is a price suggestion, looks good, approved"), and only when the process really runs like a machine, switch to full auto and hand daily pricing to the agent. The side effect: the more complex the system, the more sophisticated the tools need to be, and the less direct control remains. Balancing that is the ongoing job.

The black box trade-off

"The black box suggests this and we cannot explain why" is hard to defend inside an organization. An LLM is not explainable - it is a neural network with millions of weights that returns the most probable answer, with no explanation behind it. The balance that works is to start with good math: understand the guardrails, the elasticities, the seasonalities, the factors that drive the business. Then add small black boxes inside that controlled infrastructure where they earn their place - base demand calculation is the classic example, where gradient boosting gives the best forecast accuracy and is a complete black box. Build the trust, keep the control infrastructure, and you can still defend the proposal - because the business still owns the result.

The process: rules, optimize, review, publish, measure

Every pricing process has the same steps: rules, goals, optimization, review, publish, measure. Each step can be automated, and each should be transparent. The way in is to understand the current state first - the process, the bottlenecks, the data readiness - and then replace blocks one at a time. Rules unclear? Define them, analyze the current price architecture, compare it with competitors. Then optimization: with many rules it is hard to find the optimal price against all of them, so bring in algorithms and constraint optimization. Then, to optimize for gross profit rather than for rules, move to demand modeling - seasonality, base demand, elasticity. Step by step, learning as you go, rather than trying to get it all done on day one. That staged path is what agentic pricing in Retailgrid is built around: rules first, proposals inside them, and an approval queue before anything goes live.

Five myths about AI in pricing

1. AI will replace pricing and category management entirely. Not yet, despite the pitches. What has changed is that companies that could not afford pricing experts can now run complex price optimization projects with fewer resources - a 100 million retailer with a commercial director and no pricing specialist can start today, because most of the work can be automated. Most, not all.
2. You need a data science team first. If you are very large with lots of data, in-house is good. A mid-sized retailer needs a good CIO and good infrastructure, so the data sits cleanly in the ERP and databases - then vendors and consultants can do the rest.
3. More data means better prices. Statistically more is better, but a small, clean data set beats big data full of mistakes and inconsistencies that cannot be monetized.
4. The LLM does the math. The number one myth. An LLM is not a calculator. It reads and writes; asked to calculate, it either predicts an answer (not deterministic) or calls a mathematical function or writes code to do it. By nature it is not a mathematical tool.
5. AI pricing means a price war. That is the story of every technology: early adopters first, then everyone, then it is a commodity and a cost of doing business, like a computer. Adoption is still small; in a few years it will simply be table stakes. The only decision is whether to be early, in the majority, or last.

No magic without data

Many catalogues contain products whose price has not changed in two years. Mathematically an elasticity can still be produced; from a business perspective its confidence will be low. The honest approach is to shape the data and the algorithm - borrow elasticity from similar products, use other methods - and say so. Data should drive the idea, and AI cannot solve a data problem. If there is no data, AI cannot give you anything.

Episode timestamps

• 0:00 - AI is everywhere. What does it mean in practice?
• 0:57 - AI vs ML vs LLM vs data science
• 2:37 - Can AI replace the analyst?
• 3:35 - What it still cannot do
• 5:01 - Use cases: matching, classification, elasticity
• 7:23 - Guardrails and hallucinations
• 8:44 - Sign off: human in the loop
• 9:14 - Manual, propose and approve, full auto
• 10:44 - The black box trade-off
• 13:03 - The process: rules, optimize, review, publish, measure
• 14:53 - Five myths about AI in pricing
• 20:15 - Thin data, low confidence, no magic
• 21:21 - Be practical, ignore the hype

Where this leaves pricing teams

Maxim was building neural networks and computer vision for retail price monitoring before AI became a buzzword, and the advice at the end of the episode is the same as it was then: test it, try it, be practical about it, take the real benefits and ignore the hype. AI replaces the spreadsheet work, not the pricing manager - it moves the team from calculating to deciding.

Watch the full episode on YouTube, and subscribe to the Retailgrid channel for future episodes on pricing strategy and AI in retail.