Begin with the decision

Start with the choice you need to make.

A company may have a CRM, accounting software, spreadsheets, call records, web analytics, inventory exports, and years of customer history. That does not mean it has an answer. Data becomes useful when a specific person needs to make a recurring decision and the information arrives in time to help.

Southern Utah has many types of firms. Each one needs different numbers. But all should name the choice before they build a chart.

A better starting question is painfully specific.

“Which estimate sources produced the most gross profit last quarter after cancellations?” is useful. “Can AI find insights in our CRM?” is a demo request.

DecisionWhat will someone choose, change, approve, stop, or investigate?
OwnerWho has the authority and responsibility to act on the result?
CadenceDaily, weekly, monthly, per job, per campaign, or after a threshold is crossed.
EvidenceWhich fields, systems, calculations, and time periods can support the decision?
ConsequenceWhat does a wrong conclusion cost, and what human review is required?
Make the numbers explainable

Define each number before you make a guess.

Small-business data often looks obvious until two employees define “customer,” “lead,” “booked,” “revenue,” or “complete” differently. AI cannot resolve a business definition the company has never agreed on.

01

Name the field

Use a plain business name and the exact source column. Record whether the field is entered by a person, calculated, or imported.

Example: “Qualified lead” is not the same as “form submission.”
02

Define the formula

Write the numerator, denominator, filters, date used, exclusions, and rounding rules.

Example: close rate may use created date, decision date, or completed-job date.
03

Record the grain

Determine whether each row represents a person, lead, call, invoice, line item, appointment, job, or day.

Mixing grains creates totals that look reasonable and are wrong.
04

Mark missing data

Distinguish zero, unknown, not applicable, not yet recorded, and intentionally withheld.

A blank is not always zero.
05

Track changes

Note when software, pricing, categories, staffing, campaigns, or data-entry habits changed.

A trend break may be a process change, not customer behavior.
Questions worth answering

Use AI for work you explain and do often.

LEADS

Which sources create qualified, profitable work?

Separate volume, qualification, close rate, average value, gross profit, and time to conversion.
SALES

Where do good opportunities stall?

Compare stage age, follow-up activity, objections, owner, and outcome without treating correlation as cause.
OPERATIONS

Which work creates rework or delay?

Group reasons, handoffs, job types, teams, and time periods. Review category quality before blaming a department.
CUSTOMERS

What questions and problems repeat?

Cluster calls, tickets, reviews, and emails after removing or protecting personal information.
FINANCE

What changed from plan to actual?

Explain the variance by volume, price, mix, timing, cost, and one-time events. Keep accounting review in place.
CAPACITY

Where is staff time being absorbed?

Estimate frequency, handling time, wait time, correction time, and the portion that is actually recoverable.

AI can write queries, explain formulas, group data, and sum up patterns. A person should check the work. Do not trust a guess when the sample is small or the data has changed.

A weekly operating habit

Keep facts, ideas, and choices apart.

A decision brief should make it difficult to confuse an observed number with a theory. Use four sections every time:

Observed

What changed, by how much, during which period, and according to which defined metric.

Possible explanation

What may have contributed. Label this as a hypothesis until additional evidence supports it.

Unknown

Missing fields, small samples, quality concerns, outside events, and questions the dataset cannot answer.

End with a decision record.

What will we do, who owns it, when will we know whether it worked, and what result would cause us to change course?

A ten-day pilot

Test one report before you link every tool.

Days 1 and 2Choose the decision

Write the exact question, current method, owner, timing, and cost of uncertainty.

Days 3 and 4Prepare the dataset

Export only the needed fields. Remove duplicates, identify blanks, normalize dates, and protect sensitive data.

Day 5Verify totals

Reconcile row counts, revenue or job totals, date ranges, and key calculations with the source system.

Days 6 and 7Ask and challenge

Run the analysis, test alternative explanations, inspect outliers, and verify calculations independently.

Day 8Build the brief

Show observations, hypotheses, unknowns, source notes, and proposed actions.

Days 9 and 10Use it in a real meeting

Record the decision, owner, due date, and questions that should improve the next dataset.

The scorecard

Check the choice and the time it took.

Preparation timeHours spent collecting, cleaning, reconciling, calculating, explaining, and formatting.
Reconciliation pass rateWhether key totals match the source system within an approved tolerance.
Definition coveragePercentage of important fields and metrics with an agreed business definition.
Correction rateMaterial errors found in calculations, labels, time periods, or interpretations.
Decision completionReviews that end with a documented owner, action, due date, and success measure.
Outcome reviewDecisions later checked against what actually happened.
Accuracy is a workflow

Trust checked math more than smooth words.

Protect customer and employee information, limit datasets to the task, and use an AI environment approved for the data involved. Important calculations should be reproducible outside the model. Important conclusions should show the source period, filters, definitions, and uncertainty.

Verify math

Recalculate key totals and formulas with deterministic tools. Language models can produce incorrect arithmetic and code.

Protect identity

Remove personal information when names are not needed. Apply role-based access and retention rules.

Challenge the story

Ask what evidence would disprove the explanation and whether a process or measurement change created the pattern.

Questions owners ask

Useful questions before you build.

Can AI analyze an Excel or CSV file?

Yes, many tools can summarize fields, write formulas, create charts, and test questions against a file. The result is only as reliable as the definitions, completeness, calculations, and permissions. Verify important totals and conclusions independently.

What data should a small business analyze first?

Start with data tied to one choice you make often. You might study leads, bids, jobs, stock, repeat buyers, or bills. Use one set of data and give it one owner.

Can AI predict future sales?

It can assist with scenarios and forecasting, but a forecast is not a promise. Small datasets, seasonality, pricing changes, one-time events, and changing market conditions can make confident projections misleading. Show assumptions and ranges.

Do we need a data warehouse or business-intelligence platform?

Not for every first project. A controlled spreadsheet or export may be enough to prove value. Invest in larger infrastructure when repeated manual preparation, access control, data volume, or freshness becomes the measured bottleneck.

START WITH ONE REAL TASK

Bring one job and one task you do often.

We teach owners and teams with real work. Your people keep the steps. They also learn what to check and how to get better on their own.

Sources and further reading

We use local talks as clues, not proof. Numbers are examples unless we link to a source.