
Superintelligence Lens for Paint Stores: Formulas, Batches, Restocking
How paint stores record mixing formulas, batch numbers and reorder requests so an AI agent checks real data, not invented colors or quantities.
Want to apply this in your company?
Book a call with Juan →An AI agent in a paint store only helps when it checks recorded mixing formulas, batch data and stock. It should not try to remember colors from memory. To make that work, the store needs three sources in order: the mixing formula table, the batch register and the consumption history that feeds reordering. If any of these is incomplete, the agent should say it does not know and pass the question to the team.
Key takeaways
- A mixing formula is master data. Each color needs a code, base, colorants, quantities and version, and only formulas marked active in the table can be offered to a customer.
- A batch is traceable. The batch number, base, manufacturing date and expiry date stay linked to the product sold, because the shade can vary between batches.
- Stock and reorder figures come from the system, not from the conversation. The agent reads the current quantity, suggests a draft order, and a person still approves it.
- Without a confirmed record, there is no color or quantity answer. The case goes to a human with the context already summarized.
- Before the agent serves customers, test it with simulated formulas, batches and orders.
Why the agent invents colors and quantities
A language model produces plausible answers. When a customer asks for "a color like my sofa," the easiest response is a common color name or a code that sounds right. In a store, that becomes an expensive mistake: the wrong colorant in the machine, a returned can, rework, and a customer who stops trusting the store.
The problem is rarely the model itself. It is the lack of a source the agent can query. Agent design guidance generally treats tool use as a core part of the design: the agent should fetch data from a well-defined tool with clear inputs and outputs, not rely on what it absorbed during training. In a paint store, that tool is the lookup into the formula table and the stock record.
The practical rule is simple: every answer about color, stock or lead time must point to a record. If there is no record, the correct answer is "I don't have that information confirmed."
How to record mixing formulas
The formula table is the first asset to organize, and it is usually the messiest. Many stores keep notebooks, old spreadsheets and the counter staff's memory. Before connecting any agent, the work is to turn that into one master register.
Each formula should contain, at a minimum:
- A unique color code, with the commercial name and the internal name stored separately, so customers and staff refer to the same thing.
- The paint base and product line, such as matte, satin or gloss, because the same color can look different on different bases.
- Colorants with their codes and the quantity per container, exactly as the tinting machine uses them.
- The formula version and review date, so you know which version applied to each sale.
- The formula owner and its status: active, in testing or discontinued.
- A limitation note, for example colors that depend on a specific base batch or on surface preparation.
A discontinued or test formula should never appear to the agent as a sellable option. It can stay in the archive for history, but it should be outside the customer-facing lookup. This separation sounds like a small detail, yet it is exactly what prevents the store from offering a color it no longer mixes.
How to link paint batches to customer service
The batch number is what makes a sale traceable. Without it, when a customer returns months later asking for more of the same color, nobody knows which base was used or whether that shade still exists.
The batch record should link the batch number to the package sold, the formula applied, the base and the date. When the store does not have that link, the agent should not guess. It should say what it knows and bring in a person.
Illustrative scenario: a customer messages on WhatsApp asking for more paint for a wall painted some time ago. The agent checks the history, finds two sales of the same color with different base batches, and does not promise an identical shade. It explains the difference, asks the customer to bring a sample or a photo of the label to the store, and hands the case to the counter staff, who decide which formula to apply.
In this flow, the agent is useful without making a technical decision that belongs to the team. The record of who decided and why should also be saved, so the store can review the case later.
Stock and reorder requests: who decides what
Reordering is where invented quantities happen most easily. An agent that estimates consumption from memory can suggest an order of cans that nobody will check. The alternative is to separate clearly what the agent looks up, what it calculates and what only a person can authorize.
One possible design, to discuss with the team:
- Stock lookup: the agent replies with the recorded quantity and the time of the last update, and warns when the data may be outdated.
- Reorder suggestion: the agent calculates a suggestion from recorded consumption and the reorder point the store has defined, without using its own estimates.
- Draft order: the agent prepares the supplier, items, quantities and justification, but sends nothing.
- Supplier submission: only after approval by an authorized person.
- Price, promised delivery date or commercial terms: always outside the agent's scope.
There is also a duplication risk. If a customer asks twice, or if the system resends a message, the same order can be created twice. Payment APIs commonly use idempotency keys for this reason, so that repeating the same request does not produce a duplicate effect. The same idea applies to reorder requests: give each request a unique identifier.
How to test before serving customers
Testing should happen before any customer talks to the agent. Build a simulation environment with fictional formulas, batches and stock levels, and check whether the agent behaves the way the store expects in the difficult cases. The mock and sandbox approach is covered in more detail in another article on this blog.
Minimum cases for the test list:
- A color that exists under two different commercial names.
- A discontinued color, which must be declined with guidance.
- A color with two possible bases, which requires a clarifying question.
- A batch not found in the sale history.
- Zero stock, with and without a reorder already placed.
- A repeated order in the same conversation.
- A question outside scope, such as a renovation quote or a discount request.
The approach we use on projects like this has four stages: map the current counter, limit the agent's data and actions, pilot a real flow with logging and a metric agreed in advance, then transfer the operation and decide whether to expand, adjust or stop. This order avoids putting the agent into production before the store knows what it does. If you want to see how this work is organized, read about how the consultant works with companies.
The handoff to people also needs testing. Every transfer should carry a summary: what the customer asked, which formula or batch was checked, what the agent already said, and what still needs a decision. Without that summary, the counter staff starts from zero and the customer repeats everything.
Customer data and LGPD
Formulas and stock records are generally not personal data. But the conversations that query them often include names, phone numbers, the address of a job site and sometimes photos of rooms. That data falls under Brazil's General Data Protection Law (LGPD), and the store should know who is the controller, how long conversations are kept and who can access them.
The most useful principle is data minimization: the agent should record only what the order requires, and should not keep photos or personal data beyond a justified period. The National Data Protection Authority (ANPD) publishes guidance and is the official reference for questions on this subject in Brazil. For outbound messages, such as a notice that paint is ready or a reorder reminder, customer permission needs separate care. The topic is covered in WhatsApp opt-in and consent for outbound messages.
It also helps to organize the master data before anything else. Standardizing color names, batch codes and units of measure reduces lookup errors and makes the agent easier to maintain. The subject is treated in more detail in an article on data prepared for AI use.
Frequently asked questions
Can an AI agent give the right color from the name the customer says?
Not reliably. Names like "navy blue" or "brick gray" vary between manufacturers and between stores, so the agent should look up the registered color code and, when in doubt, ask for a sample or a photo of the surface. It should only give an answer when there is a clear match in the table.
What should happen when a customer's old formula no longer exists in the system?
The agent should say the record was not found and should not suggest an approximation on its own. The case goes to the team, which can check older files, compare samples and decide whether an equivalent formula exists. After that, the decision should become a new record in the table.
Do you need a sophisticated inventory system to start?
Not necessarily. What matters is that stock, batch and formula data exist somewhere the agent can query, with a date of last update. A well-structured spreadsheet can support a pilot, as long as it has an owner, an update rule, and the agent reads only confirmed data. More sophistication can come later, once use shows what is missing.
Can the agent place the supplier order by itself?
We recommend against it. The agent can prepare a draft with items, quantities and justification, but sending should depend on approval from an authorized person. That separation protects the store from wrong orders, duplicates and commitments on delivery or price that an automated system should not make.