B Builderlog
Builderlog · ·Buying Decisions ·Builderlog Field Manual 188 ·Sep 9, 2026 ·6 min read

AI Table Totals Need a Human Review Before You Compare

#ai#human-review#table#totals#quote-checking
AI Table Totals Need a Human Review Before You Compare

An AI comparison table can list the right items while leaving its total unsupported. The useful response is to rebuild the calculation from the source quote, then check whether every charge appears with the correct quantity, unit, and billing period. A human review needs to explain the difference, not just replace the final number.

Recalculate: derive each line amount from the original quote, then add the eligible amounts independently.
Reconcile: mark missing charges, repeated charges, and incompatible units beside the affected rows.
Release: use the comparison only after the remaining differences are explained.

This playbook provides a worksheet for that review. It requires no specialist purchase: use an available calculator or spreadsheet alongside the source document.

The receipt starts with an honest boundary

Prepared: 2026-09-09.
Tested date: unavailable; no completed numerical test was supplied.
Conditions: the supplied evidence contains no source quote, AI comparison table, prices, quantities, or calculated results.

That boundary matters. There is no verified arithmetic discrepancy to report here, and no basis for claiming that a correction worked.

The artifact below is therefore an unfilled review worksheet, supported by an explicitly fictional quote structure. Its letters stand for amounts the reviewer must establish. They are not measured values, and they do not demonstrate a successful test.

A completed receipt would contain the source entries, the preserved AI output, the independent calculation, and an explanation of each difference. Until those exist, the appropriate status is unverified.

A total becomes reviewable when every included amount has a traceable source.

Freeze the table before correcting it

Keep the original quote and the AI table visible side by side. Preserve the AI output before editing anything. Otherwise, the evidence of what changed disappears into the corrected version.

Record the comparison basis in plain language:

  • Which currency applies?
  • Is the comparison for an initial purchase, a recurring period, or a defined contract?
  • Are taxes, delivery, discounts, and optional charges included?
  • Are displayed amounts unit prices or extended line amounts?

These are calculation inputs. A sum cannot resolve an undefined billing period.

For a fictional digital storefront project, the quote might contain design work, implementation, hosting, support, and handover. Those labels provide a structure for the worksheet; they are not evidence of an actual purchase or supplier offer.

Give each source entry a stable identifier. Preserve that identifier even when the wording in the comparison is shortened.

Build the review table around the source

Start with the source quote, because starting with the AI table can hide an omitted entry. Populate every source row, then map the AI entries back to it.

Source itemSource calculation to establishIndependent amountAI entry and amountOmission checkDuplicate checkUnit or period check
Design workQuoted quantity × price per deliverableA: pendingPendingPendingPendingDeliverable basis unresolved
ImplementationQuoted quantity × applicable unit priceB: pendingPendingPendingPendingConfirm fixed scope or usage
HostingPrice per billing period × included periodsC: pendingPendingPendingPendingMatch comparison period
SupportQuoted support quantity × applicable rateD: pendingPendingPendingPendingConfirm included allowance
HandoverQuoted fixed fee or stated quantity calculationE: pendingPendingPendingPendingConfirm separate charge

Artifact caption: An unfilled numerical review worksheet. Each source item receives an independent amount and explicit checks for omission, duplication, and unit differences.

“Pending” is deliberate. Replace it with an observation only after inspecting the relevant evidence.

If an AI entry has no source counterpart, record it separately as unsupported. Do not force it into the closest-looking row. Likewise, an empty AI cell does not establish that a charge is free.

The worksheet is complete only when the mapping works in both directions: source entries have destinations, and AI entries have sources.

Rebuild the amount before comparing the total

For each row, copy the original quantity, rate, and unit into your working calculation. Do not copy an AI-derived line amount into the independent amount column.

For a quantity-priced entry:

Line amount = quoted quantity × price per matching unit

For a fixed fee, use the quoted fee under its stated conditions. Multiplying a fixed project price by a quantity intended for another purpose would introduce a new error during review.

Once the amounts share the intended comparison basis, the fictional worksheet has this symbolic subtotal:

Independent subtotal = A + B + C + D + E

Then account for any explicitly documented adjustments. Keep a discount or tax separate if the source treats it separately. Do not invent an adjustment merely because it would make the totals agree.

Finally:

Unexplained difference = AI total − independently calculated comparable total

These expressions describe the method. Without source amounts, they cannot produce a numerical result.

Preserve the arithmetic trail when you fill them in. Another reader should be able to reproduce the calculation without asking what a cell means.

Give each mismatch a reason

An omission means an eligible source charge is absent from the comparison or its calculation. Mark the source identifier, the omitted amount, and whether that amount was already included elsewhere.

A duplicate means the same charge is counted more than its source permits. Similar wording alone is insufficient evidence: implementation and handover could be separate charges. Trace the underlying source entries before removing either.

A unit difference means the compared quantities describe different bases. A recurring rate and a full-contract amount need a common period before they can support a purchasing comparison.

Multiple issues can affect the same entry. Preserve separate labels when necessary.

Matching totals do not establish that the underlying rows are correct.

An omitted charge and an extra charge could offset each other. That is a mathematical possibility, not an observed failure in this assignment. It explains why the review must inspect the rows even when the final amounts agree.

Keep uncertainty visible through the final decision

Do not describe a suspected failure as a finding. “Possible duplicate; source clarification needed” carries a different meaning from “duplicate confirmed and removed.”

Rounding also needs an explicit rule. Record whether the source rounds individual line amounts or only the final amount. If the source is unclear, retain that uncertainty instead of declaring a small discrepancy acceptable by intuition.

Use this compact sign-off record alongside the table:

Review fieldRequired record
EvidenceSource quote reference and preserved AI table
Comparison basisCurrency, scope, billing period, and adjustment treatment
CalculationIndependent line amounts and reproducible total
ExceptionsAffected source identifiers, issue labels, and explanations
DecisionUnverified, awaiting clarification, or reconciled

The decision for this article’s worksheet is unverified. No completed calculation or failure receipt is available. For an actual quote, reconciliation requires both supported arithmetic and a consistent comparison basis; it does not establish that the offer is suitable or the source quote itself is accurate.

Complete the worksheet against the original quote before using the AI table to make a buying decision.

TL;DR

Rebuild the total from source rows, label every mismatch, and keep the comparison unverified until the arithmetic and units reconcile.

The next episode examines how to preserve unresolved quote conditions without turning them into confident comparison claims.