Every account manager running a merchant portfolio has the same artifact sitting in a folder: an export. It is complete, it is current, and it is not a review. Between that file and a conversation with a merchant’s executives lies a chain of small, unglamorous decisions — which columns matter, what counts as a bad number, which of the bad numbers deserves a slide, and how to say it in a sentence somebody will act on.
“AI generates your QBR” compresses that chain into four words. It is worth decompressing, because the interesting question is not whether software can produce a deck. It is which of those decisions the software is making on your behalf, and whether you can see them.
The gap an export leaves open
An export answers “what were the numbers?” A review has to answer three harder questions: which numbers changed enough to mention, what that implies about the merchant’s business, and what should happen before the next review. Those three are where preparation time actually goes — a pattern I wrote about in why merchant account managers lose days to QBR preparation.
Automating that gap responsibly means being explicit about a tradeoff. A system that reads any file and infers meaning is flexible and unpredictable. A system that demands a known shape and applies published rules is rigid and repeatable. For a document that gets presented to a merchant’s leadership team, repeatable wins. Everything below is a consequence of choosing that side.
Step 1: a schema strict enough to trust
The scorecard requires nine columns. Not nine suggested columns — nine required headers, checked before a single row is parsed. If one is missing, the run stops and names it, rather than generating a partial scorecard with a silently absent metric.
| Required column | What it carries | Formats accepted |
|---|---|---|
| Merchant_Name | The account the scorecard is about | Text, non-empty |
| Top_Category | Merchant category, used in the summary line | Text, non-empty |
| Auth_Rate | Authorization approval rate | 0.96 or 96% |
| Top_Decline_Reason | Leading decline reason, drives a recommended action | Text, non-empty |
| Chargeback_Rate | Disputes as a share of transactions | 0.001 or 0.10% |
| Loyalty_Spend_Lift | Spend uplift among loyalty participants | 1.4 or 1.4x |
| Interchange_Downgrade_bps | Acceptance cost leakage in basis points | 12 or 12 bps |
| AEO_Visibility_Score | Answer-engine visibility, scored 0 to 1 | 0.82 or 82% |
| Monthly_Volume | Monthly processed volume | 850000, $850,000, 850k, 1.2M |
The strictness is deliberately placed at the header level and nowhere else. Column order does not matter. Extra columns are ignored, so you can hand over a wider export without pruning it. Comma and tab delimited files are both detected automatically by comparing delimiter counts in the header line, which matters more than it sounds — a spreadsheet saved as “CSV” in a European locale is frequently neither.
Value formatting is forgiving on purpose, because the mismatch between
what a BI tool exports and what a parser wants is the single most common
reason an automation attempt dies in the first two minutes. Currency
symbols, thousands separators, trailing %, bps
and x suffixes, k/M/B magnitude suffixes, and accountancy
negatives in parentheses are all normalized before validation. A
percentage sign is not merely stripped: on the three rate fields it
divides by one hundred, so 96% and 0.96 mean
the same thing rather than differing by two orders of magnitude.
Step 2: validation that names the row
Once headers pass, each row is validated independently. Rates and the visibility score must fall between 0 and 1; volume must be non-negative; the three text fields must not be empty; the column count must match the header count. A row that fails is excluded from the output rather than repaired by guesswork.
Two design choices here matter more than the rules themselves. First, errors carry the row number and the field name, plus an example of the expected format — “Expected a numeric authorization rate such as 0.96 or 96%” is a fixable message in a way that “invalid input” is not. Second, one bad row does not fail the file. In a portfolio run, the accounts that parsed cleanly still produce scorecards, and the index written alongside them records which accounts need attention and why. For anyone who has watched a batch job abort on row 34 of 40, that distinction is the whole difference between a tool and an obstacle.
One quiet default is worth knowing: if you do not state a reporting period, the agent uses the most recently completed quarter rather than inventing one. Asked in September 2026, that is Q2 2026.
Step 3: fixed thresholds, not judgment
This is the step most likely to be assumed away. When a scorecard says authorization performance is “stable,” something decided that. It is not a language model forming an impression of the number. Each metric is placed in a band by a fixed comparison.
| Metric | Band and cutoff |
|---|---|
| Authorization rate | Strong at 95% and above · stable at 90% and above · otherwise at risk |
| Chargeback rate | Low below 0.3% · watch at 1% and below · otherwise elevated |
| AEO visibility | High at 0.8 and above · moderate at 0.5 and above · otherwise low |
| Interchange downgrade | Optimized below 10 bps · moderate leakage at 30 bps and below · otherwise high leakage |
Published cutoffs are arguable, which is the point. An account team can look at “at risk below 90%” and say that for their card-not- present portfolio the line belongs at 92%. That is a productive disagreement about a visible number. It is not available at all when the same judgment is buried inside a generated paragraph.
The practical consequence is determinism. The same row produces the same band, the same sentence, and the same recommended actions on Tuesday as it did on Monday. A QBR that quietly rewords itself between the dry run and the meeting is not a tool anyone should present from.
Step 4: the narrative and its receipt
From those bands the scorecard assembles a fixed set of fields: three executive summary lines, two visibility insights, and three recommended actions. Take a synthetic account — Riverbend Home Goods, Home & Garden, 91.2% authorization, 0.71% chargebacks, 1.4x loyalty spend lift, 22 bps of downgrade, 0.58 visibility, $640,000 monthly volume. The bands resolve to stable, watch, moderate leakage, and moderate visibility, and the summary comes out as:
Riverbend Home Goods posted 91.2% authorization with watch chargeback risk at 0.71%. Loyalty spend lift reached 1.4x while interchange downgrade exposure sits at 22 bps (moderate leakage). AEO visibility is moderate at 58.0% across $640,000 monthly volume in Home & Garden.
The recommended actions are generated the same way, each pinned to a measured value rather than a generic best practice — reduce “Do Not Honor” declines by tuning retry logic and payment validation; cut downgrade from 22 bps with richer transaction data and network qualification checks; raise visibility from 58.0% through category-page optimization and refreshed metadata.
The part I would argue is the most useful, and the part that is easiest to omit, is the receipt. Every generated line carries a structured explanation alongside it: which input fields it drew on, the rule applied in plain language, and the raw values used. The first summary line above reports its sources as the authorization and chargeback rates, its rule as the two banding scales, and its values as 0.912 and 0.0071.
That turns a challenge in a meeting from an argument into a lookup. When somebody asks why the deck called their disputes a watch item, the answer is a threshold and a number, available before the question is asked. An explanation produced on demand after the fact is a different and much weaker thing than one emitted at the moment the sentence was written.
Step 5: the deck and the folder it lives in
The generated fields are substituted into a real PowerPoint template, so
the output is an ordinary .pptx that opens, edits, and
re-brands like any other deck — not an image or a PDF of a slide.
The generator reports how many placeholders it filled and, importantly,
lists any it could not resolve, so a template drifting out of sync with
the field set surfaces as a warning instead of an empty box discovered
on stage.
Where the file goes matters as much as what is in it. Saved output lands
in one convention inside the signed-in user’s own OneDrive:
Merchant QBR AutoPilot › QBR Workspaces › account
› review period, with periods normalized so “Q1
2026” becomes the sortable folder 2026-Q1. Alongside
the deck sits the source workbook and a Markdown summary with empty
headings for meeting notes, decisions and commitments, and the next
review date.
Those empty headings are the quiet answer to the failure mode where last quarter’s commitments cannot be found when this quarter’s review starts. The deck alone is a document. The deck, its source data, and the notes in one predictable folder are a review that the next one can build on.
Nothing is written without an explicit yes. The agent asks before it saves, and if a save partially succeeds it reports exactly which artifacts landed and which failed rather than retrying silently against stale content. Browsing an existing workspace is strictly read-only.
What stays with the account manager
Being specific about limits is part of describing the work honestly.
The scorecard reports; it does not explain. It can tell you authorization sits at 91.2% and that the leading decline reason is Do Not Honor. It cannot know about the pricing change in month two, the processor migration, the outage, or the merchant’s new checkout vendor. Causation is the account manager’s contribution, and it is the part the merchant is actually paying attention to.
Thresholds are a yardstick, not a benchmark. The cutoffs are consistent, which makes them useful for comparison across a portfolio and across quarters. They are not industry standards, and a team with better internal baselines should say so out loud in the review.
Scale is bounded today. A portfolio run generates scorecards for the first eight merchant rows in a file. That is a real limit worth planning around: it is a way to cover a top-accounts tier quickly, not to process a book of two hundred in one pass.
Some capabilities are demonstrations. The multi-page executive deck preview, matched-period change detection, and meeting follow-up capture currently run on bundled sample content so you can see the output format. They do not yet read your files. Scorecard generation, single and batch, is the part that works on your own data today — and a product that is clear about which half is which is easier to trust than one that is not.
Practical next steps
- Open your current merchant export and check it against the nine columns above. Most portfolio exports already contain seven or eight of them under different names; the gap is usually a rename, not a new data pipeline.
- Write down your own cutoffs for authorization, chargebacks, and acceptance cost before you evaluate any tool. If your team cannot agree on when a number becomes a talking point, no automation will settle it for you — and having the list makes any vendor’s thresholds immediately assessable.
- Pick one account and time the path from export to a deck you would actually present. That single measurement is the baseline every claim about time saved should be checked against, including ours.
- If the workflow fits, Merchant QBR AutoPilot runs it inside Microsoft 365 Copilot against a CSV or TSV file in your own OneDrive. It is free on Microsoft Marketplace, and the get-started guide covers installation, including the IT approval your tenant may require.
Frequently asked questions
What columns does a merchant QBR export need?
Nine: Merchant_Name, Top_Category, Auth_Rate, Top_Decline_Reason, Chargeback_Rate, Loyalty_Spend_Lift, Interchange_Downgrade_bps, AEO_Visibility_Score, and Monthly_Volume. Column order does not matter and extra columns are ignored, but a missing required header stops the run before any row is read.
Do I have to reformat percentages and currency before uploading?
No. The parser normalizes the formats spreadsheets and BI tools actually emit: 96% and 0.96, $850,000 and 850k and 1.2M, 12 bps and 12, 1.4x and 1.4, and parenthetical negatives such as (1,200). Comma and tab delimited files are both detected automatically.
How does an automated QBR scorecard decide what to say?
Through fixed numeric thresholds, not model judgment. Authorization at or above 95% is strong and at or above 90% is stable; chargebacks below 0.3% are low and at or below 1% are watch; AEO visibility at or above 0.8 is high and at or above 0.5 is moderate. The same input always produces the same wording.
Can an automated merchant QBR be presented without review?
No. Automation should be treated as a prepared first draft. It reports what the export contains and cannot know about a pricing change, an outage, a migration, or a relationship issue that explains the numbers. The account manager still owns the causal claim and the commitment made in the room.