A general-purpose language model cannot produce a development appraisal that a board, the Regulator or Homes England could rely on, and Proval Vantage is built so that it does not have to. Vantage's calculation engine produces every figure. Any buyer can reasonably ask whether a language model could simply be asked for the appraisal. This article compares the two approaches against the four things an appraisal has to deliver: accuracy, repeatability, traceability and calculation at scale.
| Requirement | General-purpose AI | Proval Vantage |
| Accuracy | The best model scored 34.52% on financial modelling tasks in a published benchmark (18 October 2026). | The engine calculates every figure from the appraisal inputs. |
| Repeatability | Output varies between runs; benchmarks average several runs to report a score. | The engine reproduces its own stored figures run after run. |
| Traceability | Returns a figure with no inspectable calculation chain. | Each figure follows from named inputs through named formulas, and a reviewer can walk the chain. |
| Calculation | Predicts text; the strongest models hand arithmetic to a calculator. | Executes the full monthly cashflow in seconds. |
| Role of AI | Asked to produce the numbers. | Figures come from the engine. |
Social Rent schemes start from a viability gap, so a single assumption can decide whether a scheme proceeds. Social Rent homes do not pay their way: their capital value sits well below total development cost. The subsidy needed to deliver 90,000 Social Rent homes a year in England has been estimated at about £18.83bn annually, set against the government's announced £39bn over ten years.
When the gap is that wide, the appraisal decides which schemes proceed. An organisation's appraisal assumptions are the numeric expression of its appetite for risk and its desire to build, and development teams under pressure to win sites can drift towards making projects work on paper. A tool that returns a confident figure on request, with no visible working, makes that risk harder to detect.
The published evidence shows general AI is still well short of the accuracy an appraisal needs. On a leading benchmark that tests AI models on the work of entry-level financial analysts, the best model scored 65.40% overall as of 18 October 2026. On the financial modelling tasks, which cover DCF and NPV frameworks, the best score was 34.52%.
The grading mirrors how appraisal errors behave: each question carries load-bearing figures, and one wrong figure scores the whole answer at zero. Each of those tasks is a single self-contained question. An appraisal is thousands of linked calculations, where a wrong figure in an early period flows into every period after it.
Proval Vantage does not estimate its figures. Its engine calculates each one from the inputs the appraiser sets, so accuracy depends on those inputs and the model logic, both of which the appraiser can inspect. AI scores rise with each model release, but an estimate that improves is still an estimate.
A board paper, a funder return and an audit trail all need the same scheme to give the same NPV every time it is opened. Language models can return different answers to the same question on different runs, which is why AI benchmarks run each model several times and report an average with a margin of error.
For registered providers this is a compliance matter. The Regulator of Social Housing requires detailed and robust stress testing across a range of scenarios, and boards must certify compliance in their annual accounts. A stress test is only meaningful when every difference between two runs comes from a changed input.
Proval Vantage's engine is deterministic: identical inputs return identical outputs. Its run timings were taken with the engine reproducing its own stored figures run after run.
Every number in a Proval appraisal follows from named inputs through named formulas, and a reviewer can walk the chain. A language model produces a plausible number with no chain behind it, so a wrong figure looks exactly like a right one.
That difference decides whether a figure survives scrutiny. Homes England's compliance audit checks that grant-funded providers kept to their funding conditions, and public subsidy cannot exceed actual development costs. S106 negotiations routinely come down to both sides trading spreadsheets until they agree. Local authorities preparing business cases are rated partly on whether their models have been checked for errors, and analysis without quality assurance falls into the lowest robustness category. Each of these depends on a reviewer being able to see how a figure was reached.
The arithmetic is the product. Take a 45-year appraisal of a 213-unit scheme. A monthly cashflow over 45 years has 540 periods, which across 213 units is 115,020 unit-months, each carrying loan, grant, rent and inflation assumptions, with interest and repayment feeding back on each other every period.
The mechanism is easiest to see in a single loan: £45,000 at 6% over 30 years needs a flat annual payment of £3,269, with each year's interest charged on the balance the year before left. Timing alone moves the answer a long way. In one worked private-sale example, moving the same build costs later in a ten-month programme raises the NPV at a 10% discount rate from £16,432 to £170,448.
A calculation of that kind has to be executed. Language models predict text, and the best-performing AI agents on financial tasks rely heavily on a calculator tool to do their arithmetic. Vantage's engine executes the full model directly and returns the result in seconds.
The evidence points to a clear division of work: the calculation engine produces the figures, and AI is confined to reading and summarising. On the same financial benchmark, the leading models scored above 70% on retrieval and summary tasks such as earnings and disclosure analysis, against 34.52% on financial modelling. On that benchmark, AI models performed best when finding, reading and summarising information, and weakest when building the financial model itself.
That division keeps every figure in a board paper deterministic and traceable, because the numbers come from the engine. Any AI involvement is limited to tasks that do not generate figures.
Does the same scheme return the same NPV every time it is run? Without that, stress test results cannot be separated from run-to-run noise.
Can every figure be followed back to a named input and a named formula? Compliance audit, S106 negotiation and public-sector robustness ratings all depend on it.
Is the cashflow calculated or predicted? A 45-year monthly model has to be executed period by period.
Which part of the tool produces the numbers and which writes the words? The numbers should come from the engine.
Could the board explain the figures to the Regulator without the AI in the room? If the explanation depends on the AI, the board is certifying figures it cannot defend.
Regulator of Social Housing (published by the Homes and Communities Agency), Governance and Financial Viability Standard, April 2015, paragraphs 2.4–2.8.
Homes England, Capital Funding Guide: chapter 6, Programme Management, paragraph 3.4.2, and the Glossary entry for Compliance Audit (GOV.UK, captured 1 September 2026).
MHCLG, The Appraisal Guide, fourth edition, paragraphs 3.68–3.69.
National Housing Federation, Understanding Development Appraisal (2015): foreword, chapter 1, section 5.3, chapter 8, chapter 10 and section 12.3.
Homes for People We Need, Making Social Rent Homes Viable, edition one (2025): executive summary.
Vals AI, Finance Agent v2 leaderboard and methodology, updated 18 October 2026 (archived; superseded by Finance Agent v2.1).