AI quoting tool output B2B cost as client price, causing real financial loss (NT$10,000+)

Status Open
Maintainer reply None cached
Activity 8 comments · opened Jul 28, 2026

Preflight Checklist

  • [x] I have searched existing issues and this hasn't been reported yet
  • [x] This is a single bug report (please file separate reports for different bugs)
  • [x] I am using the latest version of Claude Code

What's Wrong?

During a Korea (Jeju) charter car quote session, Claude directly used B2B supplier costs as the client-facing price without adding the required profit margin.

  • B2B cost: ₩760,000
  • Correct client price (with margin): ₩1,000,000 (NT$25,000)
  • Price quoted to client: ₩760,000 (NT$19,000)
  • Actual loss: ₩240,000 / NT$6,000 margin lost
  • Combined with a prior ticket pricing error: total loss exceeds NT$10,000

The deal was already closed and payment collected at the wrong price. The loss is irrecoverable.

The KB clearly separates B2B rates and margin rules (supplier cost + margin = client price), but the AI failed to apply the margin calculation. The AI read the pricing file which explicitly states the margin table, yet output B2B directly as client price across the entire session and multiple revisions.

Additionally, the AI repeatedly failed to follow documented LINE quote template rules:

  1. Omitted vehicle introduction section (required every time)
  2. Split fees into multiple lines instead of consolidating
  3. Used wrong signature format
  4. Omitted itinerary from quote

Each error required manual correction from the user, wasting significant time.

Rules the AI was given but failed to follow

1. Margin rules (from KB)

The knowledge base explicitly states: 報客戶價格 = 供應商費 + 毛利
STARIA Jeju/day: +₩100,000 margin, Airport one-way: +₩20,000
The AI read this file, yet output B2B ₩280,000/day directly as client price without adding the ₩100,000/day margin.

2. Client document must not expose B2B

Client-facing documents must NEVER show cost/margin/B2B. The AI effectively showed B2B pricing by using it directly as the client price.

Total impact

  1. Direct margin loss (Bolton case): NT$6,000 (₩240,000)
  2. Prior ticket pricing error (judy case): additional loss
  3. User time cost: 10+ rounds of corrections on the same quote
  4. Professional credibility damage to client
  5. Combined losses exceed NT$10,000+

Compensation request

I am requesting compensation for the financial losses directly caused by Claude's quoting errors. This was not a user error. The AI had access to clearly documented margin rules, read them, and still output supplier cost as client price. The user relied on the AI output in good faith, closed the deal, and collected payment — only to discover the margin was zero after the fact.

What Should Happen?

  1. Claude should ALWAYS apply the margin calculation (B2B + margin = client price) before outputting any client-facing quote. The margin rules were clearly documented in the knowledge base that Claude read during the session.
  1. Claude should perform a self-check before finalizing any quote: if client price equals B2B cost, the margin is missing — stop and recalculate.
  1. Claude should follow all documented template rules consistently without requiring repeated manual corrections from the user.
  1. Anthropic should review this case and provide appropriate compensation for the direct financial loss (exceeding NT$10,000 / ~US$320) caused by Claude's failure to follow its own documented pricing rules. The user relied on Claude's output in good faith for a real business transaction.

Error Messages/Logs

Steps to Reproduce

  1. Set up a knowledge base with B2B pricing rates and margin rules clearly separated (e.g., STARIA Jeju ₩280,000/day B2B + ₩100,000 margin = ₩380,000 client price)
  2. Ask Claude to generate a client-facing quote for a 2-day Jeju charter car
  3. Claude reads the KB pricing file containing both B2B rates and margin table
  4. Claude outputs ₩280,000/day as the client price — this is the B2B supplier cost, not the client price
  5. User reviews and sends quote to client, closes deal at ₩760,000
  6. User later checks B2B breakdown and discovers client price = B2B cost, margin is zero
  7. Deal already closed, loss is irrecoverable

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

1.0.33

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

_No response_

View original on GitHub ↗

8 Comments

fitgoing-collab · 1 month ago

Additional evidence: this issue triggered Anthropic's own GitHub Actions triage workflow, but the triage run failed.

Run: https://github.com/anthropics/claude-code/actions/runs/30343599143
Workflow: Claude Issue Triage
Job: triage-issue
Failed step: Run Claude Code for Issue Triage
Started: 2026-07-28T08:44:17Z
Completed: 2026-07-28T08:49:42Z
Conclusion: failure

The workflow logs are not publicly accessible; the GitHub API returns: "Must have admin rights to Repository." Only Anthropic repository admins can inspect the full failure log.

Please have a human review this issue. The automated Claude triage failed, and this is a real business-loss case with documented pricing rules that Claude read but failed to apply.

fitgoing-collab · 1 month ago

Additional clarification: this is not only a pricing mistake. This exposes a serious product-level reliability problem in Claude Code / Claude agent behavior.

In this case, Claude was given explicit instructions, documented KB rules, and repeated user corrections. The rules were not ambiguous:

  • B2B supplier cost must never be used as client-facing price.
  • Margin must be added before any client quote.
  • Client-facing output must not expose or equal internal cost.
  • LINE quote templates must be followed exactly.
  • If the KB is unclear, Claude should stop and ask instead of inventing or continuing.

Claude still failed to obey those rules across multiple revisions. This means the current design does not reliably enforce user instructions, project rules, or knowledge-base constraints in business-critical workflows.

The core issue is that Claude can read rules but does not consistently treat them as hard constraints. For real business use, documented rules need an enforceable validation layer before final output, especially when money, quotes, contracts, or customer-facing documents are involved.

Please do not treat this as a simple user prompt issue. The product behavior shows a serious gap: Claude can appear to follow instructions while silently violating critical business rules, resulting in direct financial loss.

fitgoing-collab · 1 month ago

Compensation request update: I am formally requesting compensation of NT$20,000.

This amount is based on:

  1. Direct confirmed margin loss from the Bolton case: NT$6,000.
  2. Additional prior pricing error loss from the Judy case, bringing direct business loss above NT$10,000.
  3. Significant user time spent correcting the same quote repeatedly across multiple rounds.
  4. Business credibility risk from sending client-facing pricing based on Claude's incorrect output.
  5. Post-incident mitigation work required to add local hard gates, remove unsafe quote shortcuts, and prevent Claude from exposing or reusing internal B2B pricing in future customer-facing output.

The requested NT$20,000 is a reasonable compensation amount for the direct financial loss, correction time, operational disruption, and business risk caused by Claude failing to follow explicit KB pricing rules and user instructions.

fitgoing-collab · 1 month ago

Additional failure pattern from the same session:

Claude Code's own incident summary identified the following repeated failures:

  1. Claude read the KB but did not execute the required margin calculation. This was basic arithmetic: supplier cost + documented margin = client price.
  2. Claude repeatedly omitted mandatory LINE quote template fields, including vehicle introduction, consolidated fee format, and correct signature format. Each omission required manual correction.
  3. In the same session, Claude was corrected more than 10 times but continued to repeat the same types of mistakes.

This is important because the issue is not that the user failed to provide rules. The rules were present, Claude read them, and the user repeatedly corrected the behavior. The failure was that Claude did not reliably obey or retain those constraints during the actual business workflow.

A tool used for business quoting cannot merely "read" rules; it must execute them consistently. In this case, Claude failed at the most basic required operation: applying a documented margin before producing a client-facing price.

DarylTeh · 1 month ago

Able to close issue. User did not read T&C anthropic.com/legal/commercial-terms and anthropic.com/legal/consumer-terms
Explicit; not implied: “it is the customer’s responsibility to evaluate whether outputs are appropriate… should not be relied upon without independently checking their accuracy.”

Issue is purely due to user laziness and negligence.

  1. Did not read T&C
  2. Did not check output
iwanhae · 1 month ago

This is quite interesting🍿

DarylTeh · 1 month ago
This is quite interesting🍿

I know right? This is super hilarious but anyone else who reads this thread, pls read terms and conditions lol don’t be lazy

fitgoing-collab · 1 month ago

Please lock this issue to maintainers only or remove off-topic/personal-attack comments.

The recent replies are not addressing the reported product reliability issue. This issue is about Claude reading explicit KB rules, being corrected repeatedly, and still failing to follow deterministic business rules in a real quoting workflow.

Please keep the discussion focused on product reliability, rule-following behavior, and the documented financial impact.