AI vs. Human Quotes: The Data Governance Truth Behind 41% Inconsistency

Title

AI Quotes vs. Human Quotes Conflict: The Data Governance Truth Behind 41% Inconsistency

Meta Description

A commercial cleaning company case reveals: 41% inconsistency between AI and human quotes stems from data governance, not the model. Learn the five-step fix — establish a single source of price truth, enforce structured quotes, measure drift weekly, and cut inconsistency to 6%.

Article Content

A commercial cleaning company asked me to diagnose their AI receptionist because "it keeps quoting the wrong prices." I pulled a quarter's worth of their 143 quote records — every case where the AI gave a number first and a salesperson then quoted the same job. The results were startling: 59 quotes were inconsistent, 41%, with an average gap of 18%. In the most extreme case, the AI quoted $1,850 for an office cleaning of the same square footage at the same frequency, while the account manager quoted $2,400. The client received two written quotes within four days.

The Problem Isn't the Model — It's Data Governance

My first instinct was AI hallucination, but reading the conversation logs showed otherwise. The AI was reading a 2023 price list in a shared drive, sales were using a 2025 version in their own email, and the CRM system held a third set of numbers that matched neither. Three sets of "officially correct" data sources had never been reconciled:

Data Source	Used By	Last Updated	Monthly Minimum
Price List v3 (shared drive)	AI agent	March 2023	$480
Price List v5 (sales email)	Account manager	January 2025	$620
CRM Product List	CRM-generated quotes	Partially migrated	$550

The AI wasn't wrong — it was simply loyal to a document no one had taken down. This is both an easy problem to fix and a worse one — because when clients discover two prices, they don't think "their data is messy," they think "they're making things up."

Quote consistency is fundamentally a data governance problem wearing a prompt-engineering disguise. Tuning prompts first only makes the AI more confident about the wrong numbers.

AI hallucination typically refers to a model generating plausible but factually incorrect content, such as fabricating nonexistent citations or prices. But in enterprise settings, many errors mistaken for "hallucination" are actually the model faithfully reading incorrect or outdated internal data. Distinguishing the two matters: hallucination requires improving the model or prompt, while data source errors require governance. This article's case is the latter — the AI didn't invent numbers; it faithfully repeated an old price list that had never been taken down.

Step 1: Establish a Single Source of Price Truth

The fix begins before any prompt work. Designate exactly one artifact as the price authority, and give it a version number and expiration date. All other copies are allowed to be wrong because they aren't authoritative.

The structure I use is a single file, one sellable unit per row, no prose:

sku,unit,list_price,min_qty,tier_2_qty,tier_2_price,floor_price,effective_from,expires,notes
CLEAN-STD-SQFT,sqft,0.14,2000,5000,0.11,0.09,2026-01-01,2026-12-31,excludes consumables
CLEAN-DEEP-SQFT,sqft,0.31,1000,3000,0.26,0.22,2026-01-01,2026-12-31,first clean only
CLEAN-WEEKLY-VISIT,visit,180,1,4,155,140,2026-01-01,2026-12-31,min 4 visits/month

Three rules make it usable by the AI:

1. The expires field is mandatory. An expired price is worse than no price because the AI will quote it confidently. Instruct the AI: if today's date exceeds expires, state that confirmation is needed and escalate — no extrapolation.
2. floor_price is none of the AI's business. The AI quotes only list or tier prices. Floor prices are for humans. If the AI can see the floor, it will eventually use it to close a deal, giving margin away to a machine that's just trying to help.
3. notes contains exclusions only. Not sales guidance, not "good margin here." Anything that reads like advice will be paraphrased by the AI as a commitment.

Then delete or watermark the old tables. That company's 2023 table was still in the shared drive 18 months after it expired — and it was the first result when searching for "price."

Single Source of Truth (SSOT) is a core principle in data governance: designate one authoritative storage location for a given type of data, with all other copies being derived or read-only. In quoting scenarios, this means all systems (AI, CRM, sales tools) read from the same versioned, expiration-dated price file. SSOT doesn't require physically having one file; it requires logically having one source recognized as "correct," with all other copies explicitly marked as non-authoritative or synced from the authoritative source.

Step 2: Enforce Structured Quotes, Not Quote Sentences

Most AI quotes in prose: "A 3,000 sq ft office cleaned weekly runs about $420 per visit." Such sentences can't be audited, can't be compared, and invite clients to haggle over feelings.

Make the AI output a fixed block:

QUOTE
Service:        [service name]
Quantity:       [number] [unit]
Unit price:     $[x] per [unit]  (tier: [list/tier 2])
Line total:     $[x]
Assumptions:    [up to 3 — access hours, frequency, exclusions]
Valid until:    [date from expires field]
Not included:   [consumables / travel / after-hours]
Next step:      [book survey / confirm frequency / human review]

This achieves three things at once: all quotes are comparable in a spreadsheet; assumptions are listed explicitly — 80% of disputes hide here ("you didn't say there were three floors"); and the AI is forced to state what it's not quoting, the most common cause of angry post-invoice calls.

The Valid until field also quietly does its job: when a client digs up an old email, the AI has a legitimate reason to say "that price was valid through March 31" rather than argue.

Structured output means requiring a model to return results in a predefined format (e.g., JSON, fixed field blocks) rather than free text. Structured output is easier to parse, store, and compare programmatically, and it reduces the model's room for improvisation. In quoting, a fixed block forces the model to explicitly list quantity, unit price, assumptions, and exclusions, making every quote an auditable record. This can be enforced via prompt templates, JSON Schema constraints, or function calling.

Step 3: Measure Drift Weekly, Not by Anecdote

You can't manage this by reading conversation logs. Build a small report: for every job that received both an AI quote and a human quote, calculate the gap.

Week of: ____
Jobs quoted by AI then by human: ____
Quotes that matched (±2%): ____%
Median gap: ____%
Largest gap: ____% on [service]
Gap direction: AI higher ____% / AI lower ____%

Two patterns are worth watching:

• AI consistently lower than humans: usually expired tier pricing, a missing volume discount, or the AI applying a discount it shouldn't know about. The business risk is clients bringing AI numbers to humans, who eat the difference to close — training everyone to distrust the AI and quietly destroying margin.
• AI consistently higher than humans: usually the AI quoting list price while humans habitually discount 10–15% off the books. That's not an AI defect; it's an unrecorded discount policy. Write it down, or stop doing it.

After that cleaning company consolidated to one price list and added the quote block, inconsistency dropped from 41% to 6% in five weeks. The remaining 6% was almost all legitimate: site surveys uncovering issues that couldn't be caught over the phone.

Quote drift refers to price differences for the same job across different times, channels, or quoters. Drift isn't necessarily an error, but persistent drift erodes client trust. Measuring drift requires pairing AI quotes with human quotes and calculating match rate, median gap, and gap direction. Continuously monitoring drift reveals systemic issues — such as expired tier pricing, unrecorded discount policies, or unsynced data sources — and lets you fix them before clients notice.

Step 4: Publish Discount Authority

When two parties can quote, you need a table showing who can move numbers and by how much. Otherwise the AI quotes list price, the human quotes list minus 15%, and clients learn to always ask for a human.

Role	Discount Authority	Must Record
AI agent	0% — list price or published tier price only	Full quote block in conversation log
Front desk	5%	Reason code
Account manager	12%	Reason code + competitor reference
Sales director	20%	Written approval
Below floor price	Owner only	—

Give the AI a one-liner for pressure: "I can't adjust the price, but I can connect you with someone who can look at the whole project." That's honest, protects the floor, and routes negotiation to someone authorized to negotiate. An AI that improvises a discount to be helpful is the most expensive kind of helpful.

Step 5: Reconcile Before Contract, Not After

One checklist item prevents most problems. Before any quote becomes a contract, someone compares the AI quote block with the final human quote and records the difference. 30 seconds per job.

[ ] AI quote exists and is attached to the deal
[ ] Final quote price = AI price, or difference has a reason code
[ ] Assumptions in AI quote match survey findings
[ ] Final quote's valid-until date is in the future
[ ] Final quote's exclusions match what the AI stated
[ ] If discount applied: within role authority, reason recorded

The second-to-last line catches a lot. The AI said "not included: consumables," and the human's final quote dropped that line to look cleaner. Six weeks later the client gets a consumables invoice and calls it bait-and-switch.

What Actually Costs You

No one loses a deal because the AI quoted $1,850 instead of $1,900. You lose deals, referrals, and the presumption of good faith because a client can prove you gave two different numbers for the same job and can't explain why.

Here, consistency is cheaper than accuracy. An AI that quotes slightly high but consistently is more trustworthy than one that quotes perfectly but changes its mind when a different person answers the phone.

Start with the price file. Add the expires column today, take down the old tables this week, and make the AI output quote blocks instead of sentences. Within a month you'll find your own drift number — and it'll be higher than you expect.

One-line Summary

AI quote inconsistency stems from data governance, not model hallucination; unifying price authority and structuring quotes can cut inconsistency from 41% to 6%.

Paragraph Summary

A commercial cleaning company's AI receptionist kept quoting wrong prices. After analyzing 143 quote records, the author found 41% of AI quotes were inconsistent with human quotes, averaging an 18% gap. The root cause wasn't AI hallucination but three unreconciled price data sources: the AI read a 2023 price list, sales used a 2025 version, and the CRM held a third set of numbers. The fix included: establishing a single authoritative price file with a mandatory expires field; forcing the AI to output structured quote blocks instead of prose; measuring quote drift weekly; publishing a discount authority table; and reconciling before contract. After implementation, inconsistency dropped from 41% to 6% within five weeks. The article emphasizes that consistency is cheaper than accuracy, and client trust comes from explainable, stable quotes.