Why Generic AI Assistants Cannot Handle Construction VAT Compliance

Every major AI assistant — ChatGPT, Claude, Gemini, Copilot — produces output that is the most likely response given the words in your question, based on patterns in its training data. That output may be useful as general background. It is not a compliance decision.

The fundamental issue is structural, not a matter of which model you use or how you phrase the question. Generic AI responses on construction VAT are probabilistic (the most likely answer, not necessarily the correct one), stateless (the model has no knowledge of your project unless you re-state every relevant fact each time), and unverifiable (there is no mechanism to confirm which HMRC guidance the answer is based on, or whether that guidance is current).

The output is prose. It cannot be attached to an invoice, submitted to HMRC, or reviewed by an accountant as evidence that a compliance decision was made correctly.

HMRC Notice 708 is a layered ruleset. The correct VAT treatment of any supply depends on the intersection of multiple factors simultaneously: the project route (new build, conversion, long-empty, developer), the building's previous use, the vacancy period, the nature of the supply, the supplier's registration status, and whether the Domestic Reverse Charge applies. Generic AI models consistently fail on this domain in several predictable ways.

Context collapse: the model cannot hold your full project context — building type, prior use, vacancy period, client registration status — in a way that conditions every subsequent classification decision. Ask about scaffolding on a conversion and you get a general answer, not one conditioned on your building's specific history and your client's registration position.

Rule bleeding: the rules for new builds (zero-rated contractor services) and conversions (reduced-rated at 5%) are frequently conflated. Getting this wrong means accepting an overcharge on a conversion or incorrectly disputing a zero-rated invoice on a new build.

DRC blind spots: the Domestic Reverse Charge, introduced in 2021, is complex and interacts with CIS registration in ways that training data handles poorly. A model that misidentifies whether DRC applies creates VAT liability on both sides of a transaction.

Hallucinated citations: models will confidently cite 'HMRC Notice 708 Section X' for a proposition that section does not support. There is no runtime check to catch this, and the citation sounds authoritative.

No invoice-level resolution: a model can discuss VAT treatment in general terms but cannot process an actual invoice, apply project-specific rules to each line item, and produce a structured output that your accountant can review.

Construction VAT is one example of a broad class of compliance problems where AI assistants are genuinely valuable for conversation, but structurally insufficient for making the compliance decisions themselves. The defining characteristic of this class is that the rules are deterministic, not probabilistic: a supply is either zero-rated or it is not. A wrong answer has real legal and financial consequences.

The correct answer also depends on facts that exist only in a specific dataset — your building type, its prior use, the registration status of the parties, the nature of the supply. Without those facts, no general answer can be trusted.

The output must be in a form that can be reviewed by a third party — an accountant, a solicitor, an HMRC officer — and retained as evidence. An AI conversation is ephemeral. A structured, HMRC-referenced decision record is not.

VATBuild's architecture addresses this directly: a rules engine that encodes HMRC Notice 708 deterministically, a controlled workflow that ensures every decision is made in the right order, and targeted AI used only where it is well-suited — document understanding, natural language explanation — not for making the compliance decisions themselves.