Here is a scenario that is happening at UK field service companies right now, more often than anyone is comfortable admitting.

An estimator uses an AI tool to draft a remedial quote. The output looks reasonable. The sections are correct. The language is professional. The pricing looks plausible, a CCTV crawler unit at one figure, labour at another, materials in the right ballpark. The estimator is busy. They have seventeen more quotes to get through before end of day. They review it quickly and send it.

Three weeks later, the job completes. The actual cost of the crawler unit from their supplier was forty percent higher than the AI had quoted. The margin evaporates. The client is not at fault. The contract is not at fault. The AI had generated a confident, well-structured quote using pricing data sourced from somewhere on the internet. Not from the company's supplier agreements. Not from their historical job data. Not from any data source the business had ever authorised or verified.

This is not a hypothetical. It is the predictable, structural consequence of deploying a generic AI tool into a pricing-sensitive workflow it was never designed for. And it is costing UK contractors real money.

Why Generic AI Gets Prices Wrong: Every Time

The answer to why AI keeps generating inaccurate prices is not a mystery. It is a direct consequence of how large language models work, and understanding this makes the failure mode obvious, predictable, and preventable.

A general-purpose AI like Copilot, Gemini, or ChatGPT is not a database. It does not retrieve stored facts the way a search engine retrieves pages. It is a prediction engine, trained to generate the most statistically plausible next word given everything that came before. When it produces a price, it is not looking up what your supplier charges for a riser second stage unit this quarter. It is predicting what a price for that item plausibly looks like, based on patterns in the data it was trained on, which is broadly the internet, frozen at a point in time, and containing no knowledge of your supplier agreements, your contracted rates, your regional material costs, or your margin requirements.

The financial cost of this structural flaw is now measurable at scale. According to Forrester Research, AI hallucination-related errors cost enterprises an average of $14,200 per employee annually in verification, correction, and rework. For a field service business processing twenty to thirty quotes per day, the cumulative exposure from pricing errors that slip through, or the time cost of the checks required to catch them, represents a significant and largely invisible operational risk.

What makes this particularly dangerous in a quoting context is a finding from MIT research published in 2025: AI models are 34% more likely to use confident language when generating incorrect information than when generating correct information. The quote that is most wrong often reads as the most certain. An estimator under time pressure reviewing a plausible-looking output has no signal that the pricing is fabricated. It looks exactly like a quote that got the prices right.

The Difference Between AI That Guesses and AI That Knows

The problem is not that AI cannot be used for quoting. It is that most AI being used for quoting has not been given the information it would need to do the job accurately. There is a fundamental difference between an AI that predicts prices from its training data and an AI that retrieves prices from your actual supplier relationships, your SOPs, and your historical job records.

The architecture that makes this distinction possible is Retrieval-Augmented Generation, commonly known as RAG. Rather than relying on a model's frozen training data, a RAG system retrieves relevant, curated information from a trusted knowledge source now it generates each output. Studies from Vectara's HHEM benchmarks show that RAG-grounded AI reduces hallucination rates by between 40 and 71% compared to standard LLM outputs, a reduction that, in a pricing context, translates directly into more accurate, more defensible quotes.

In practical terms for a UK drainage or M&E contractor, this means the AI is not guessing what a CCTV crawler unit costs. It is retrieving the price from a pricing table the business has uploaded, a table that reflects the company's actual supplier agreements, updated to current rates, not scraped from a trade forum three years ago. It is not guessing what the correct output structure should look like. It is retrieving the mandatory sections from the company's own SOP documentation and populating them with evidence from the engineer's job sheet.

This is the distinction between intelligent quoting and expensive guessing. One is grounded in the business's own operational reality. The other is grounded in probability, which in pricing means plausibility, not accuracy.

What Intelligent Quoting Actually Requires?

Building a quoting workflow that produces accurate, defensible outputs consistently requires three things that generic AI tools do not provide by default.

The first is a grounded knowledge base. Pricing tables built from your actual supplier agreements. SOPs that define what a compliant remedial report must contain. Historical job data that captures how your estimating team has priced similar work in the past. Compliance standards, SFG20, Building Safety Act documentation requirements, embedded as validation rules rather than left to the AI to infer. This knowledge base is the difference between an AI that knows your business and one that is guessing about it.

The second is a structured, validated output format. Generic AI produces outputs that vary in structure from job to job, depending on how the input was framed. A properly designed quoting system generates outputs in a mandatory format, the same sections, in the same order, every time, and validates that every section is populated with evidence from the source document, not with fabricated content. Missing information is flagged with a placeholder. Contradictions in the source documents are surfaced as ambiguities. Nothing is invented to fill a gap.

The third is full traceability. Every line item in the output should trace back to a specific source: a job sheet section, a supplier price table entry, or a remedial form field. Every material, every labour estimate, every equipment reference. Traceability is not just a compliance requirement. It is the mechanism that allows an estimator to review an AI-generated quote in five minutes rather than twenty-five, because they can see immediately where each figure came from and assess whether it is correct.

What This Looks Like in Practice?

A mid-sized UK drainage contractor processing remedial quotes through a properly grounded AI workflow sees a fundamentally different process to one using a generic tool.

An engineer submits a job sheet: a brief set of notes and photos from a reactive attendance. The system ingests it, applies OCR and visual reading where the document is image-based, and extracts the relevant content. It then generates a structured remedial report across a mandatory seven-section format, populated entirely from evidence in the source documents. Labour hours that are not stated are flagged as missing rather than estimated. Equipment that is not explicitly mentioned is not included. Materials are listed using the exact terminology from the job sheet, not synonyms the AI prefers.

Alongside the report, a pricing CSV is generated with every line item traceable to a specific source reference. Labour rates come from the company's own rate card. Material prices come from the uploaded supplier pricing table. Where a price cannot be confirmed from the knowledge base, a placeholder is inserted, not a fabricated figure that will erode margin three weeks later.

This is the approach built into PromptX, VE3's AI platform for field service document workflows. PromptX uses a workspace model, each job isolated, each output grounded in the company's own uploaded knowledge: pricing tables, SOPs, compliance standards, historical job data. The system enforces a zero-hallucination framework: evidence-led only, no invented quantities, no fabricated pricing, no assumptions substituted for stated facts. Estimator review takes five to ten minutes rather than twenty-five, because every figure is sourced and every gap is flagged. The AI does not guess. It retrieves, structures, and validates, then hands a defensible draft to the human who approves it.

The Commercial Case for Getting This Right?

For UK contractors operating on tight margins, which describes most of them given rising NLW rates, increased employer NI contributions from April 2025, and sustained material cost inflation, a quoting tool that occasionally gets prices wrong is not a minor inconvenience. It is a margin risk that compounds daily across a high-volume workflow.

At twenty quotes per day, five working days per week, fifty working weeks per year, a contractor processes approximately five thousand remedial quotes annually. If even two percent of those contain a material pricing error that slips through review, a conservatively low estimate for a generic AI tool working from internet data, that is one hundred quotes per year where the job is priced incorrectly before it is even accepted. At an average error value of a few hundred pounds per quote, the annual exposure is material.

The choice between AI that guesses and AI that knows is not a technology decision. It is a commercial one. The cost of deploying an intelligent, grounded quoting system is a known, bounded investment. The cost of deploying a generic tool and absorbing the margin erosion from pricing errors is an unbounded, ongoing liability, one that is easy to miss quarter by quarter and impossible to ignore when you total it up annually.

A quoting tool that doesn't know your supplier rates isn't saving you time. It's redistributing your margin into your clients' pockets.

PromptX is VE3's AI platform built specifically for field service document workflows: remedial report generation, pricing extraction, and quote automation, grounded entirely in your own operational data, not the internet. To see what supplier-grounded AI quoting looks like on a real UK contractor job sheet.