AI for Excel and PowerPoint

This course is about dependable office work with AI, not a tour of buttons that may change with the next product release.

A useful spreadsheet is a small information system. A useful presentation is a decision interface. AI can accelerate both, but it cannot decide what your data means, repair missing facts with guesses, or take responsibility for a recommendation.

You will work with a fictional company called Aperture Workspaces. The practice workbook contains a realistic sales table plus four intentional defects: one duplicate row, inconsistent capitalization, trailing whitespace, and one missing owner. Your job is to identify and classify those issues, calculate defensible metrics, and present one bounded recommendation. You will also build controls for risks that are not present in the exercise data, because a clean result is not proof that an important failure mode was tested.

Download the course files

Terms the course builds on, in the order you will meet them

Agent Mode

Copilot’s edit mode in Excel: it writes formulas, sheets, charts and PivotTables into the workbook, so its proposals arrive as changes to review.

Metric definition

The formula, denominator, unit, rounding and zero-case behavior that fix what a number means.

Reconciliation

Showing that every summary adds back to the cleaned source total.

Excel Table

A named range with one header row and one record per row, which formulas and filters follow as it grows.

Grain

What one row represents: here, one month of one region’s activity, not one order.

Correction log

Each change’s location, original and corrected value, reason, source and authority.

Invariant

A rule every acceptable file obeys, such as returns never exceeding orders, written down and re-run on every file.

Denominator

What a rate divides by: for the return rate, total orders, not the number of rows.

REVIEW state

What a formula shows when its denominator is zero, so an undefined rate never passes as 0.

Lookup contract

What a lookup assumes: unique keys, a meaning for no match, and an owner who keeps the reference table current.

Weighted margin

(Total revenue − total cost) / total revenue: the margin of the business, not the average of row margins.

Observation and hypothesis

What the data measured, against an explanation that still needs evidence.

Formula Audit

A sheet of checks that report PASS or block release.

Claim register

The list tying each slide claim to its source cell, caveat, reviewer and date.

Claim title

A slide title that states the supported conclusion rather than the topic.

Evidence card

A slide’s source, period, population, metric, observation and unknowns, with its permitted and blocked claims, written before the title.

Most office-work errors happen before the formula or slide: the question, source, definition, or decision was never made explicit.

Imagine a manager asks, “Can you analyze sales and make a deck for tomorrow?” Before opening an AI assistant, translate that request into an operating contract. Which period? Which source is authoritative? Does “sales” mean booked revenue, invoiced revenue, or cash collected? Who will see the deck? What decision should they make?

The backward arrow matters. If a chart reveals an impossible spike, do not explain it with a story. Return to the data. If a formula produces a result that cannot be reconciled, do not round it until it looks right. Return to the definition and calculation. A professional workflow makes correction cheap before a decision makes the error expensive.

The fastest workflow is not the one that ignores privacy, retention, licensing, or authority.

Office files often contain more than their visible table. Hidden sheets, comments, formulas, named ranges, document properties, speaker notes, embedded files, and revision history can expose information you never intended to send. Treat upload as disclosure.

For this course, every company, person, transaction, and number is fictional. That is intentional. Synthetic data lets you practice the complete workflow without normalizing unsafe behavior. In real work, follow your employer's data classification, approved-tool, retention, and cross-border processing rules.

Useful prompts define the decision, evidence, constraints, output shape, and verification burden.

“Analyze this spreadsheet” gives the assistant permission to choose the question, metric, and interpretation. A stronger prompt narrows those choices. It provides only the minimum safe data, names the business rule, requires assumptions to be visible, and blocks unsupported completion.

Decision
Should we fund a 30-day East-region diagnostic before reallocating growth budget?

Verified inputs
- East gross margin: 30.6%
- East return rate: 5.0%
- West gross margin: 36.0%
- West return rate: 3.0%

Task
Separate observations from hypotheses. Propose one reversible next step.

Constraints
- Do not infer a cause from these aggregates.
- Do not create new numbers or forecasts.
- Mark missing evidence explicitly.

Output
Observation | possible explanation | evidence still needed | decision

This prompt does not ask AI to “be an expert.” It specifies an expert's discipline. The output remains a draft until you compare each claim against the workbook.

Five elements of a reliable prompt

1. Outcome: the decision or artifact the work must support.
2. Evidence: approved inputs, definitions, and reporting period.
3. Constraints: prohibited assumptions, data boundaries, and non-goals.
4. Format: the structure that makes review fast.
5. Verification: tests, reconciliation, caveats, and blocked states.

The downloadable prompt library gives you reusable versions for formula design, data quality, chart selection, slide critique, rehearsal, and final verification.

Fluency can hide arithmetic errors, invented causes, stale feature knowledge, and confident misuse of business definitions.

Use a four-part check. First, verify the source: are the rows complete, current, and authorized? Second, verify the calculation independently in Excel. Third, verify the interpretation: does the evidence support the wording? Fourth, verify the action: is the recommendation reversible and proportionate to uncertainty?

A good AI response often saves drafting time. Verification converts saved time into trustworthy work. Skipping verification merely transfers the time to a later incident, rework cycle, or embarrassing meeting.

Reliable analysis starts with rectangular data: one header row, one observation per row, one meaning per column.

Convert the practice data range into an Excel Table before building analysis. Tables expand with new rows, provide structured references, retain filters, and make formula intent more readable. Give the table a meaningful name such as SalesData.

AI works better when the structure is explicit, but the larger benefit is human: another analyst can understand and audit your workbook without reverse-engineering formatting conventions.

Read the grain before counting rows

The supplied table has 48 monthly regional records after cleaning: twelve months for each of four regions. A row summarizes activity for its displayed owner, channel, and product; it is not one customer order. Orders is a count and Returns is a returned-order count. For this synthetic exercise, Returns refers to the same order cohort as Orders. Thus summing Orders gives the denominator, while counting rows gives 48 reporting records. Real return data may arrive in a later month; choose an order-cohort rate or a period-flow ratio explicitly before comparing regions.

Copy only the header and data rectangle into an Excel Table: exclude the title, subtitle, blank rows, and totals. Check the table's row count after removing the known duplicate; a title accidentally treated as a record should fail that check.

Cleaning is the controlled resolution of defects, not permission to make incomplete data look complete.

Open the practice workbook and inspect the Raw Data sheet. Use filters, conditional formatting, Remove Duplicates on a copy, and validation rules to find defects. Ask AI to propose detection rules from the schema, but decide corrections from an authoritative source or a documented business rule.

The correction log

For every change, record where it is, what it was, what it became, why, on whose evidence and on whose authority. If the correct value is unknown, preserve the missing state and describe how downstream calculations handle it.

If the source column cannot be filled, the change is not a correction. That single test resolves most cleaning arguments, and it is what makes the next table decidable rather than a matter of confidence.

Rows four and five look similar and are opposites. An outlier is unusual; an invariant violation is impossible. Deleting the first loses a real sale and flatters the data; ignoring the second carries a contradiction into every calculation downstream. The distinction is whether a rule of the business has been broken, not whether the number looks surprising.

Observed defects are not the same as tested invariants

Write the invariants even when every current row passes. That is the whole distinction: fixing four rows cleans one file, and writing four rules cleans every file you receive afterwards — including the one where the defect appears for the first time and nobody is looking.

The practice workbook carries its own checker. Clean Raw Data, read what fails, and fix what it names until every rule passes.

Cleaning is easy to declare and hard to prove. The practice workbook’s Check My Cleaning sheet makes it provable: eleven rules read Raw Data directly and report PASS or CHECK, and the sheet reads READY only when every rule passes. Below them, a grid counts the rows for each month and region; every cell should be 1. Work in your working copy of Raw Data, in place, and keep the original file unchanged.

Untouched, the practice file reads NOT YET. This is what the sheet reports before any cleaning:

Check                                        Expected     Actual  Status
Records after cleaning                             48         49  CHECK
Month-region pairs present exactly once            48         46  CHECK
Rows whose region is not spelled exactly            0          1  CHECK
Rows whose product is not a catalog label           0          1  CHECK
Rows with a blank owner                             1          1  PASS
The blank owner is East, May 2026                   1          1  PASS
Revenue total                               8,837,000  9,057,500  CHECK
Cost total                                  5,763,552  5,904,672  CHECK
Orders total                                    8,952      9,161  CHECK
Returns total                                     310        316  CHECK
Rows where returns exceed orders                    0          0  PASS
Overall                                                           NOT YET

Read the failures before fixing anything. The duplicate West August row makes 49 records and inflates every total, revenue by ₹220,500, and the grid shows it as a 2. West’s February row spells its region west , with a trailing space, so it fails the spelling rule and leaves a 0 in the grid. July’s West row says Focus desk. And the blank owner passes: an unknown stays unknown, so the sheet allows exactly that one blank, at East, May 2026, and no other.

# The Check My Cleaning rules, rebuilt from the practice workbook. The five rows the defects
# touch are listed; the other 44 rows, which no rule asks you to change, are summed.
REGIONS, CATALOG = ["North", "South", "East", "West"], ["Focus Desk", "Flex Chair", "Team Pod"]
REFERENCE = {"revenue": 8_837_000, "cost": 5_763_552, "orders": 8_952, "returns": 310}
OTHER_44 = {"revenue": 8_097_700, "cost": 5_283_040, "orders": 8_245, "returns": 286}
RAW = [  # month, region, owner, product, orders, revenue, cost, returns
    ("2026-02", "west ", "Kabir", "Team Pod", 151, 159_200, 101_888, 5),
    ("2026-05", "East", "", "Flex Chair", 146, 147_200, 101_568, 7),
    ("2026-07", "West", "Kabir", "Focus desk", 201, 212_400, 135_936, 6),
    ("2026-08", "West", "Kabir", "Team Pod", 209, 220_500, 141_120, 6),
    ("2026-08", "West", "Kabir", "Team Pod", 209, 220_500, 141_120, 6),
]

def TRIM(text):        # Excel's TRIM: no spaces at either end, single spaces inside
    return " ".join(word for word in text.split(" ") if word)

def PROPER(text):      # Excel's PROPER: a capital after any non-letter, lower case elsewhere
    out, after_letter = "", False
    for ch in text:
        out += ch.lower() if after_letter else ch.upper()
        after_letter = ch.isalpha()
    return out

def lookup(text, allowed):   # the controlled list is the source; XLOOKUP also ignores case
    return next((value for value in allowed if value.lower() == TRIM(text).lower()), text)

def check(rows):             # every comparison is exact, as EXACT makes it
    pairs = [(row[0], row[1]) for row in rows]
    column = {"orders": 4, "revenue": 5, "cost": 6, "returns": 7}
    totals = {name: OTHER_44[name] + sum(row[column[name]] for row in rows) for name in REFERENCE}
    rules = [("records", 48, 44 + len(rows)),
             ("pairs once", 48, 44 + sum(pairs.count(p) == 1 and p[1] in REGIONS for p in pairs)),
             ("region spelling", 0, sum(row[1] not in REGIONS for row in rows)),
             ("product labels", 0, sum(row[3] not in CATALOG for row in rows)),
             ("blank owners", 1, sum(row[2] == "" for row in rows)),
             ("blank at East May", 1, sum(row[2] == "" and row[:2] == ("2026-05", "East") for row in rows)),
             *[(name, REFERENCE[name], totals[name]) for name in REFERENCE],
             ("returns over orders", 0, sum(row[7] > row[4] for row in rows))]
    return [f"{rule} {actual:,}" for rule, expected, actual in rules if actual != expected]

def clean(region=lambda text: lookup(text, REGIONS), duplicates="keep first", fill_owner=""):
    rows = []
    for month, reg, owner, product, *numbers in RAW:
        row = (month, region(reg), owner or fill_owner, lookup(product, CATALOG), *numbers)
        if not (duplicates == "keep first" and row in rows):
            rows.append(row)
    return [row for row in rows if duplicates != "delete both" or rows.count(row) == 1]

print(f"TRIM {TRIM('west ')!r}  PROPER {PROPER('west ')!r}  lookup {lookup('west ', REGIONS)!r}")
for name, rows in [("untouched file", RAW), ("owner filled with Leena", clean(fill_owner="Leena")),
                   ("both West August rows deleted", clean(duplicates="delete both")),
                   ("TRIM alone on the region", clean(region=TRIM)),
                   ("PROPER alone on the region", clean(region=PROPER)), ("the four decisions", clean())]:
    failed = check(rows)
    print(f"{name}: {'READY' if not failed else 'NOT YET'}, {len(failed)} of 11 read CHECK")
    for start in range(0, len(failed), 4):
        print("  " + ", ".join(failed[start:start + 4]))

TRIM 'west'  PROPER 'West '  lookup 'West'
untouched file: NOT YET, 8 of 11 read CHECK
  records 49, pairs once 46, region spelling 1, product labels 1
  revenue 9,057,500, cost 5,904,672, orders 9,161, returns 316
owner filled with Leena: NOT YET, 2 of 11 read CHECK
  blank owners 0, blank at East May 0
both West August rows deleted: NOT YET, 6 of 11 read CHECK
  records 47, pairs once 47, revenue 8,616,500, cost 5,622,432
  orders 8,743, returns 304
TRIM alone on the region: NOT YET, 2 of 11 read CHECK
  pairs once 47, region spelling 1
PROPER alone on the region: NOT YET, 2 of 11 read CHECK
  pairs once 47, region spelling 1
the four decisions: READY, 0 of 11 read CHECK

READY proves less than it seems to. It cannot see your correction log, a column sorted on its own (which keeps every total and scrambles the rows), or the formulas you build next. The reference workbook’s Formula Audit covers the analysis; a reviewer covers the log.

A formula can be syntactically valid and still implement the wrong business definition.

Before asking for a formula, write the rule in plain language. Gross margin is not “revenue minus cost” when expressed as a percentage. It is (Revenue - Cost) / Revenue, with defined behavior when revenue is zero. Return rate may mean returned units divided by units sold, not returned orders divided by all orders. The denominator is part of the product definition.

=IFERROR(([@Revenue]-[@Cost])/[@Revenue],0)

=SUMIFS(SalesData[Revenue],SalesData[Region],$A2)

=IF([@Revenue]=0,"REVIEW",([@Revenue]-[@Cost])/[@Revenue])

The third formula is often more honest than silently returning zero. Zero margin and undefined margin are different facts. Choose error behavior that helps the reviewer notice data requiring attention. The reference workbook uses the third form for every margin and rate, and its Formula Audit blocks release while any of them shows REVIEW.

Boundary cases AI should help enumerate

• blank values and blank-looking text;
• zero denominators and negative amounts;
• dates at month or financial-year boundaries;
• duplicate keys and missing lookup matches;
• numbers stored as text and inconsistent units;
• errors that should propagate versus errors that should be labeled.

Every lookup assumes something about key uniqueness, missing records, freshness, and ownership.

XLOOKUP is readable for many one-to-one joins. INDEX/MATCH remains useful for compatibility and flexible lookup direction. Power Query is better when data must be imported, shaped, merged, and refreshed repeatedly. The choice is not fashion; it follows from scale, refresh behavior, workbook compatibility, and auditability.

Ask AI to compare implementation options and generate test cases. Do not ask it to decide that duplicate keys are harmless. That decision belongs to the business rule and data owner.

Worked lookup: a match is not proof of a unique owner

In a separate small example, create a table named OwnerDirectory with columns Owner and Team: Meera → North, Kabir → West, Meera → Special Accounts. Put Meera in cell A2. The default XLOOKUP returns the first match, so it can hide this conflict. Guard the lookup with a count:

=IF(COUNTIF(OwnerDirectory[Owner],A2)<>1,"REVIEW",XLOOKUP(A2,OwnerDirectory[Owner],OwnerDirectory[Team],"REVIEW"))

Meera must produce REVIEW, Kabir must produce West, and an absent owner must produce REVIEW. Remove the conflicting Meera entry only when the directory owner resolves it. This teaching formula assumes exact ordinary names; production identifiers need normalization and an explicit policy for wildcard characters. In Excel editions without XLOOKUP, keep the uniqueness guard and use an exact INDEX/MATCH lookup. Microsoft documents XLOOKUP's matching behavior and availability.

A summary is only useful when each aggregation matches the metric's definition.

Revenue can usually be summed. Unit price may need a weighted average. Margin percentage should normally be calculated from total revenue and total cost, not averaged across rows. For this workbook, return rate uses total Returns divided by total Orders. Distinct customers require a distinct count, not a row count.

Design the PivotTable from the decision backward. If the question is where to investigate margin erosion, use Region as rows, total Revenue and Cost as values, a calculated margin measure, and Returns as supporting evidence. Add Product and Channel as filters for diagnosis. Validate the PivotTable grand total against an independent worksheet formula.

Calculate the difference before building the PivotTable

Consider two synthetic reporting rows: revenue 100 with cost 50, and revenue 900 with cost 810. Their margins are 50% and 10%. Averaging them gives 30%, which weights the small row as heavily as the large one. Total profit is 50 + 90 = 140 on revenue 1,000, so the combined margin is 14%. Recompute this from the PivotTable's summed Revenue and Cost outside the pivot, or use a measure that divides the sums. Do not use Average of a precomputed row-margin column.

rows = [(100, 50), (900, 810)]                          # (revenue, cost)
margins = [(revenue - cost) / revenue for revenue, cost in rows]
average = sum(margins) / len(margins)
weighted = (sum(r for r, _ in rows) - sum(c for _, c in rows)) / sum(r for r, _ in rows)
print("row margins:", ", ".join(f"{m:.0%}" for m in margins))
print(f"average of the row margins {average:.0%}; margin of the business {weighted:.0%}")

row margins: 50%, 10%
average of the row margins 30%; margin of the business 14%

An observation describes measured evidence. A hypothesis proposes an explanation that still needs evidence.

The reference workbook shows East with the lowest gross margin and highest return rate. That is an observation. “Discounting caused poor margin” is a hypothesis. So are product mix, damaged goods, customer mix, service quality, and unusual deal structure. Aggregate data cannot choose among them.

A strong analysis response should provide four separate outputs:

1. Observation: East margin is 30.6%, below the other regions, while its return rate is 5.0%, the highest regional rate.
2. Implication: East has lower gross profit per unit of revenue and a higher observed share of returned orders.
3. Uncertainty: the current table does not isolate price, product mix, discount, quality, or service causes.
4. Next decision: fund a bounded diagnostic before broad growth reallocation.

Worked case: why a percentage needs its denominator

In the reference workbook, East recorded 101 returns across 2,017 orders, so its return rate is 101 / 2,017 = 5.01%. West recorded 71 returns across 2,374 orders, so its rate is 71 / 2,374 = 2.99%. The dashboard rounds those values to 5.0% and 3.0%. The comparison is meaningful because both rates use returned-order counts divided by order counts for the same cohort and reporting period.

regions = {"North": (68, 2404), "South": (70, 2157), "East": (101, 2017), "West": (71, 2374)}   # (returns, orders)
for name, (returns, orders) in regions.items():
    print(f"{name:<6} {returns:>4} returns / {orders:>5,} orders = {returns / orders:.2%}")
returns = sum(r for r, _ in regions.values())
orders = sum(o for _, o in regions.values())
print(f"all    {returns:>4} returns / {orders:>5,} orders = {returns / orders:.2%}")
print(f"East has {101 - 71} more returns than West, and {2374 - 2017} fewer orders")

North    68 returns / 2,404 orders = 2.83%
South    70 returns / 2,157 orders = 3.25%
East    101 returns / 2,017 orders = 5.01%
West     71 returns / 2,374 orders = 2.99%
all     310 returns / 8,952 orders = 3.46%
East has 30 more returns than West, and 357 fewer orders

Now imagine an AI response says, “East has 30 more returns, therefore product quality is worse.” The arithmetic difference is real, but the conclusion is not. East and West processed different order volumes, the product mixes may differ, and the table does not include return reasons. The supported statement is narrower: East has the higher observed return rate and therefore deserves diagnosis.

Next, test whether the recommendation remains sensible under uncertainty. If most East returns come from one product introduced late in the year, investigate that product rather than the entire region. If return reasons are evenly distributed but deep discounting explains the margin gap, the intervention changes. If missing records are concentrated in East, fix data quality before making an operating decision.

This is the pattern to ask AI to follow: compute with a named denominator, compare like with like, list alternative explanations, identify the evidence that distinguishes them, and propose a reversible action while uncertainty remains. The numbers narrow the question; they do not automatically answer it.

Use AI as an adversarial reviewer: ask what else could explain the pattern, which denominator might be wrong, what selection bias exists, and what evidence would falsify the recommendation. This is more valuable than asking it for five impressive insights.

A dashboard should expose the state, comparison, exception, and next question without making the reader hunt.

The first row answers “Where are we?” with revenue, target variance, gross margin, and return rate. The trend chart answers “When did performance change?” The regional chart and table answer “Where is the exception?” The implication table prevents the visual from pretending that an observed pattern proves a cause.

Chart selection rules

• Use a line chart for change over an ordered time axis.
• Use bars for comparisons among discrete categories.
• Use scatterplots for relationships, while avoiding causal language.
• Use tables when exact values or multiple measures matter more than shape.
• Avoid pie charts when categories are numerous or differences are small.

Write a claim-based chart title only after verification. “Revenue crossed target in June and finished ₹231K above target” is useful if the chart and source table support it. “Strong momentum ahead” is vague, promotional, and unsupported.

months = ["Jan", "Feb", "Mar", "Apr", "May", "Jun", "Jul", "Aug", "Sep", "Oct", "Nov", "Dec"]
revenue = [522900, 551500, 580100, 608700, 637300, 719100, 750000, 780800, 871900, 905000, 938200, 971500]
target = [550000, 580400, 610800, 641200, 671600, 702000, 732400, 762800, 793200, 823600, 854000, 884400]
print("months above target:", ", ".join(m for m, r, t in zip(months, revenue, target) if r > t))
print(f"year: revenue {sum(revenue):,} against target {sum(target):,}, variance {sum(revenue) - sum(target):+,}")

months above target: Jun, Jul, Aug, Sep, Oct, Nov, Dec
year: revenue 8,837,000 against target 8,606,400, variance +230,600

east, west, baseline = 2_037_000, 2_504_900, 1_900_000
print(f"true ratio East / West: {east / west:.0%}")
print(f"bars from zero:  East {east / 1e6:.3f}, West {west / 1e6:.3f} (millions)")
print(f"bars from 1.9M:  East {(east - baseline) / 1e6:.3f}, West {(west - baseline) / 1e6:.3f}; "
      f"East looks {(east - baseline) / (west - baseline):.0%} of West")

true ratio East / West: 81%
bars from zero:  East 2.037, West 2.505 (millions)
bars from 1.9M:  East 0.137, West 0.605; East looks 23% of West

A workbook should make its own important assumptions and reconciliation visible.

The reference workbook contains a Formula Audit sheet. It checks whether monthly revenue equals total revenue, regional revenue equals total revenue, target variance equals revenue minus target, weighted margin matches its definition, returned-order counts do not exceed order counts, the only blank required field is the documented unknown owner, and no metric shows REVIEW. Each check reports PASS or a blocking status, and the publish gate reads READY only when all eleven pass.

Controls do not prove every value is correct, but they detect classes of failure before presentation. Add checks for the risks in your workbook rather than copying generic controls mechanically.

Build a verification packet, not just a dashboard

An experienced reviewer should be able to inspect the decision without reverse-engineering your workbook. Keep four small records with the deliverable:

This packet is deliberately lightweight. Its purpose is not ceremony; it is to make silent changes, stale data, ambiguous definitions, and unsupported slide edits visible before they reach a decision-maker.

The presentation should reduce the audience's decision effort without hiding uncertainty.

Do not copy the dashboard onto ten slides. Start with the decision required. Then select only the evidence necessary to understand the situation, compare alternatives, judge risk, and assign action.

A story is not drama added to numbers. It is the logic connecting evidence to a responsible action.

What the success gate commits the room to

“Approve a 30-day diagnostic with a success gate” is only a decision if the gate is written down before the diagnostic starts. A success gate is a measurable condition, agreed in advance, under which the intervention scales, with the action for each outcome fixed before anyone sees the result. Agreeing it first is what stops a disappointing result from being argued into a success afterwards. The reference deck states its gate on slides 5 and 6:

A gate also has to be hard to pass by luck, which brings back the denominator from the worked case above. Thirty days of East orders is about 200, close to its monthly volume late in 2026, and at that size the returns condition is noisy.

from math import comb

def chance_below_4_percent(orders, rate):
    """Chance that fewer than 4% of orders come back, if each order returns independently at rate."""
    return sum(comb(orders, k) * rate**k * (1 - rate)**(orders - k)
               for k in range(orders + 1) if k / orders < 0.04)

for orders in (200, 600):
    print(f"{orders} orders: passes with no real change {chance_below_4_percent(orders, 0.05):.0%} of the time; "
          f"passes after a real fall to 3.5% {chance_below_4_percent(orders, 0.035):.0%} of the time")

200 orders: passes with no real change 21% of the time; passes after a real fall to 3.5% 60% of the time
600 orders: passes with no real change 11% of the time; passes after a real fall to 3.5% 72% of the time

With one month of orders, about one diagnostic in five shows returns below 4% when nothing has changed, and a real fall to 3.5% is missed four times in ten. Read the count beside the rate at day 30, and agree in advance that a borderline result extends the diagnostic rather than passing it.

One slide should make one supported point, not store everything the analyst knows.

Three zones, three jobs

Use the title for the claim, the body for evidence, and the speaker notes for source, caveat, and delivery guidance.

The title row is the highest-return change available on any deck. A claim title lets a busy reader understand the argument before reading the chart, and it also disciplines the slide: if the evidence in the body does not support the title, that is now visible on the slide itself rather than only to the person who built it.

Layout rules, and the reason for each

The last row is the one that matters most in practice and gets the least attention, because the slide looks finished without it. Every deck that is any good gets forwarded, and at that moment the only caveats that exist are the written ones.

And the line to hold with an AI assistant: it can usefully draft alternate headlines, tighten dense text, or propose a reorganization — all genuine help with the framing. It must not change a number to make the layout cleaner. If a value does not fit, the honest moves are to simplify the message or split the slide; adjusting the figure to suit the design is the one edit that turns a presentation problem into a misrepresentation.

The distance between “the chart shows” and “therefore we should” is where unsupported narratives enter.

For each slide, write a small evidence card before writing the title:

Source: Analysis!regional_summary
Period: Jan-Dec 2026
Population: 48 cleaned fictional monthly regional records
Metric: returned orders / orders
Observation: East = 5.0%, highest region
Unknown: reason for return, product mix, discount, service level
Permitted claim: East requires diagnosis
Blocked claim: East quality problems caused weak margin

This card makes the boundary visible. It also helps AI produce better language because the permitted and blocked claims are explicit. Keep the evidence card in notes, an appendix, or a review worksheet even when it is not shown to the audience.

Feature availability depends on your Microsoft 365 plan, license, account, organization settings, app version, and file location. The course workflow works without Copilot.

In supported Microsoft 365 environments, Copilot can help create a presentation from a prompt or source, rewrite text, summarize a deck, and edit content. Use it after you have a verified brief. A good creation prompt names the audience, decision, approved facts, desired slide count, tone, and constraints against invented evidence.

Review the generated deck as a candidate, not a deliverable. Check every number, source, image right, title claim, chart scale, and speaker note. AI may choose a visually appealing structure that does not match your organization's template or may compress a caveat until it disappears.

If Copilot is unavailable, use the same brief manually: create six blank slides, add claim-based titles, insert verified charts from Excel, write implications, and record sources in notes. The professional method is independent of the tool.

Delivery quality includes timing, clarity, accessibility, objection handling, and clean handoff.

PowerPoint's Speaker Coach can provide rehearsal feedback such as pace and some delivery habits. Treat automated feedback as coaching input, not a judgment of expertise or communication quality. Rehearse the opening decision request, transitions, likely objections, and closing owner/date confirmation.

Ask these questions aloud

• Can I explain the recommendation in one sentence?
• Can I trace every important number to the workbook?
• What would change my recommendation?
• Which statement is an observation, and which is a hypothesis?
• Who owns the next action, and when is the next decision?

Before sharing, inspect hidden slides, speaker notes, comments, author metadata, linked workbooks, embedded objects, and accessibility. Open the files on a second machine or clean profile when possible. A presentation that works only on the creator's computer is not finished.

Six errors a reviewer should catch before the meeting

Each error below has appeared somewhere in this course. The third column names the check in the workbook or the verification packet that catches it, and the fourth names the story that lets it through.

Complete the exercise before opening the reference files. The objective is an evidence chain, not matching the sample's visual style.

1. Define the request. Write the audience, reporting period, decision, source, metric definitions, constraints, and required evidence.
2. Inspect and clean. Find every intentional defect in the practice workbook. Record corrections and preserve unknown values rather than guessing.
3. Calculate. Build monthly revenue and target, regional revenue, weighted gross margin, and return rate. Add boundary handling.
4. Reconcile. Create controls proving that monthly and regional summaries match the cleaned source and that operational invariants hold. Record them in a source manifest, metric contract, control gate, and claim register.
5. Analyze. Write supported observations, possible explanations, missing evidence, and one reversible next step.
6. Visualize. Build a compact dashboard with KPI state, trend, regional comparison, and a visible implication/caveat.
7. Present. Create a six-slide decision deck with claim-based titles, sources in notes, named owners, success measure, and date.
8. Review. Use the verification and privacy checklists, then compare your logic with the reference workbook and deck.

Release rubric

Product features and licensing can change. Use the current official documentation for availability and behavior.

• Microsoft Support: Get started with Copilot in Excel
• Microsoft Support: Analyze Data in Excel
• Microsoft Support: Edit with Copilot in PowerPoint
• Microsoft Support: Rehearse with Speaker Coach
• Microsoft Support: Detect formula errors in Excel
• Microsoft Learn: Microsoft 365 Copilot licensing
• Microsoft Support: Agent Mode in Excel — edit, plan and chat modes, and availability; checked 1 October 2026.
• Microsoft Support: COPILOT function — no longer available from 14 September 2026; checked 1 October 2026.
• Microsoft Support: Overview of Excel tables
• Microsoft Support: Top ten ways to clean your data
• Microsoft Support: TRIM function
• Microsoft Support: EXACT function
• Microsoft Support: XLOOKUP function
• Microsoft Support: Remove hidden data and personal information with the Document Inspector
• Microsoft Support: Make your PowerPoint presentations accessible
• W3C: Understanding WCAG 2.2 Success Criterion 1.4.3, Contrast (Minimum)

Course artifact provenance

The Aperture Workspaces dataset, workbooks, dashboard, prompts, checklists, and presentation were created for KnowledgeOS Academy. They are fictional educational materials and may be used to complete this course. No external company data or third-party stock imagery appears in the downloadable files.

The workbook metrics are formula-driven. On 1 October 2026 every formula in the reference workbook was recalculated in LibreOffice Calc without an error and all eleven audit checks passed; cleaning the practice data as this course teaches reproduces the reference data row for row; and the reference deck was rendered to confirm that no title or text box overflows. On 2 October 2026 the practice workbook’s Check My Cleaning sheet was recalculated in LibreOffice Calc: untouched it reads NOT YET with the eight failures shown above, a correct cleaning reads READY, and each of the four wrong fixes in the table fails the rules listed beside it. The files remain examples, not financial, legal, privacy, or employment advice.