Multi-store numbers and AI decisions: 4 steps to align the counts first
Fix how stores count and calculate first, and AI's explanations can support decisions like adding peak-time slots.
Key points
- Align counts: members, months, refunds
- Fix sums in sheets; AI gets checked data
- Let AI explain gaps and ask, not compute
- Judge the trial by the capacity decision
Contents
Align how every store counts before anything else
A&A perspective
Before you ask AI to explain your store numbers, make every store count members, the reporting month and refunds the same way. Aligning the counts here means making a number with the same name measure the same thing in every store. The reporting month means which month's figures a sale or refund is counted in.
From the sources
One example of pulling store numbers into one place appears in an announcement Anthropic published for small businesses on September 15, 2026. At Mothership Coffee Roasters, which runs six cafés plus wholesale and catering in Las Vegas, the numbers were split across the register, the accounting software, team chat and email.
From the sources
There was no single place to see them week by week. The founder connected those sources to Claude and made it the place where her team does forecasting and weekly planning. The announcement says she now runs all six locations from one consolidated view.
A&A perspective
The announcement does not say how the stores' counting rules were aligned. Automate & Augment (A&A), the company writing this article, thinks a step that aligns the counts is needed before the numbers are pulled together and compared, because stores can use the same name while counting different things.
Hypothetical example
Take, for example, a hypothetical company that runs three teaching studios. Store A counts members as people enrolled at month-end; store B counts anyone who came at least once that month. For refunds, stores A and B deduct them from the original month's sales, while store C deducts them from the month the money was returned.
A&A perspective
If you ask AI to explain the gaps between stores from this table as it is, it may describe differences in counting as differences in customer behavior. The more readable the explanation, the harder it is to notice the counting mismatch underneath.
A&A perspective
Choose what to align from the decision you want to make at the next meeting. If the question is whether to add class slots at popular times, you need three counts aligned: members, the reporting month and refunds. Try to align every number at once and you will miss the first meeting.
A&A perspective
Put the counting rules in one table. For each item, write each store's current rule, the aligned rule, the month it changed and who decided it. This does not change your accounting. It is a management rule for comparing the business.
Fix the sums in a spreadsheet, check them, then hand them to AI
A&A perspective
Once the counts are aligned, produce amounts and ratios with a spreadsheet or a small aggregation script, and give AI only the tables that pass the checks. If AI does the arithmetic in each conversation, the same source data will not always give the same numbers.
From the sources
Anthropic's engineering article "Building effective agents" recommends starting with the simplest solution and adding complexity only when needed. For well-defined tasks, it says, having the model and tools follow predefined code paths makes behavior predictable and results consistent.
From the sources
The same article says that when a task is split into steps that run in sequence, programmatic checks can be placed between the steps.
From the sources
Another Anthropic article on evaluating AI, "Demystifying evals for AI agents," describes code-based checks as fast, cheap and reproducible. It says checks performed by a model do not give the same result every time.
A&A perspective
A&A considers store reporting to be one of those well-defined tasks, because it repeats the same calculation in the same shape every week. Fix the calculation in a script written with Claude Code or in spreadsheet formulas, and run that same calculation every time.
A&A perspective
Before anything goes to AI, a machine checks three things. Does the total across stores match the source total? Is any store missing? Are any rows left that fit no category? Rows that cannot be classified are not folded into "other"; they are listed separately as waiting for review.
From the sources
The same announcement also describes a US firm that handles bookkeeping, payroll and HR for small businesses. It says the firm used Claude to narrow a $13,000 discrepancy, buried in two and a half years of tens of thousands of payment records, down to 16 transactions.
A&A perspective
A&A reads this as an example of AI being useful at the checking step too. Its role, though, stops at finding where a discrepancy is; people decide how to fix it. Corrected numbers are recalculated with the same calculation before they go back into the table. The table below shows who owns what across the four steps.
| Step | Owner or tool | What it covers |
|---|---|---|
| Align the counts | Owner | How members, reporting month and refunds are counted, and the month a rule changed |
| Calculate and check | Spreadsheet or script | Amounts and ratios, matching totals, missing stores, unclassified rows |
| Explain gaps and ask | AI | Explanations of gaps, questions to check next, causes labeled as hypotheses |
| Decide | Owner | Whether to add slots, or whether to collect the missing numbers |
Have AI explain the gaps and ask questions, not make numbers
A&A perspective
Once the checked table is handed over, what AI writes is an explanation of the gaps between stores and the questions to check next. There is no need to have AI recalculate amounts or ratios.
A&A perspective
Give AI three things: the checked table, the list of rows waiting for review, and the one decision you want to make at the next meeting. Including the decision keeps AI's explanation focused on the gaps that matter for it.
A&A perspective
Have AI show which column of the table each number in its explanation came from. When it names a cause that is not in the table, such as the weather or a nearby competitor, have it label that cause a "hypothesis." That way, when you read it, you can tell what the table supports from what is a guess.
Hypothetical example
Back in the three-studio example, suppose store A's sales fell from the previous month while attendance rose. Recalculated with the aligned rules, it turns out that refunds belonging to the previous month had landed in this month.
Hypothetical example
What you want from AI here is not a general comment like "Strengthen your marketing." You want the explanation that most of the sales gap comes from refund timing, and the question "Do we record how many people were turned away because classes were full?"
From the sources
Anthropic's announcement says that, by default, Claude does the work but waits for the user's approval before anything is sent, posted or paid.
A&A perspective
A&A likewise keeps AI's explanation as material for a decision: whether to add slots is something the owner decides.
Judge the trial by whether the capacity decision got made
A&A perspective
When you first try this flow, judge it by whether the planned decision got made, not by whether the table was produced automatically. If the tables line up but you still cannot decide because you do not know the staffing cost of extra slots, the trial is not finished.
From the sources
Anthropic's article on evaluations separates the record of an AI's interactions from the state at the end. Its example is a flight-booking agent: even if the agent says "Your flight has been booked," the outcome is whether a reservation actually exists.
A&A perspective
A&A applies this to store reporting. Rather than AI's report that it "finished the summary," record as the outcome whether the meeting could make the capacity decision.
A&A perspective
Keep the record in two levels. Level one is whether the table is done: do the totals match, is no store missing, and are unclassified rows listed separately? Level two is whether the planned decision was made. If level one passes but level two stalls, write down why it stalled and which numbers were missing.
A&A perspective
Limit the trial to one meeting and the most recent month. Run it next to the table you build by hand today, and do not add stores or numbers until the person in charge can explain the differences.
A&A perspective
If the trial stalled for lack of numbers, spend just the next week collecting the missing ones. In the studio example, that is the number of people turned away because classes were full. If it stalled because the counts still disagree, go back to the counting table.
A&A perspective
People and spreadsheets own the counts and the calculations, AI writes the explanation, and people decide. Where that line sits moves with what AI can do today.
Summary
A&A perspective
Align how members, the reporting month and refunds are counted across stores, in one table.
A&A perspective
Fix amounts with a spreadsheet or script, and check totals, missing stores and unclassified rows before handing them to AI.
A&A perspective
Have AI explain the gaps and ask what to check next, and label causes that are not in the table as hypotheses.
A&A perspective
Record the trial in two parts: whether the table was done, and whether the capacity decision was made.
When you use store numbers with AI to make business decisions, people and spreadsheets settle the numbers before AI sees them. Align how members, the reporting month and refunds are counted, then fix amounts with a spreadsheet or script that repeats the same calculation, and check them. Have AI explain the gaps and ask what to check next, and judge the trial by whether the capacity decision got made.
FAQ
Can't I let AI read the raw exports from the register or booking system and do the totals too?
For a one-off investigation, that is enough. Numbers you compare every week and use for investment decisions, such as adding slots, should come from a spreadsheet or script that repeats the same calculation. Put AI to work finding where discrepancies are, and you keep its speed.
Do I need a counting table if I only have two stores?
Yes, if each store uses a different register or booking system. If both use the same system with the same settings, one line about which month refunds go back to may be enough.
Can the counting table differ from how my accounting handles things?
Yes. The table holds management rules for comparing the business and does not change your accounting. For tax and year-end treatment, follow your accountant's judgment.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- Claude for Small Business launches new workflows, integrations, and training programs
Anthropic · 2026-09-15
Accessed 2026-10-22 - Building effective agents
Anthropic · 2024-12-19
Accessed 2026-10-22 - Demystifying evals for AI agents
Anthropic · 2026-01-09
Accessed 2026-10-22