axrea

The work, brought into focus

Design for better
human–AI decisions.

Three real estate workflows. Explore how evidence, AI capabilities and expert judgement can work together, and how Axrea would evaluate decision quality, confidence and reliance.

Fictional cases and figures, created to explain the method. No client results are claimed.

01 / Property lending

Use AI with confidence.
Keep the judgement that matters.

A useful architecture lets analysts use dependable AI preparation while recognising where the evidence still needs interpretation. The aim is proportionate reliance and more effective credit judgement.

Evidence + judgement
Valuation + tenancy + policy: AI capabilities and expert judgement contribute to the decision.Two coloured signals follow the evidence through AI and expert review, merge into a copper signal and briefly illuminate the recorded decision.VALUATION + TENANCY + POLICY62.5%PROVISIONAL LTVRenewal unagreed60% of rent · expiry 18 months£3.0mLOAN£4.8m assumed valueAI CONTRIBUTIONAI contributesConnect and compareEXPERT JUDGEMENTExpert judgesInterpret the basisRefer with a traceable reasonTEST HOW THE WORKFLOW SUPPORTS THAT JUDGEMENT

Illustrative workflow · signals show contribution, not live AI processing

01 / The task

The evidence is present. The dependency is unresolved.

A £3.0m loan is assessed against a £4.8m valuation that assumes a lease renewal. The principal tenant supplies 60% of rent; its lease expires in 18 months. A broker expects renewal, but agreement is absent. The analyst must reconcile the qualification, the broker’s account and the lender’s referral rule.

V1 · Valuation / T1 · Tenancy / B1 · Broker note / P1 · Fictional policy

02 / AI contribution

Make useful preparation dependable enough to use.

AI extracts figures, links documents and calculates 62.5% loan to value. A suitable architecture also preserves the distinction between an assumed renewal and an agreed one. In the starting design, “expected” and a pre-filled “Accept” blur that distinction. Axrea would retain the useful preparation and design a clearer handover to judgement.

Confidence in the arithmetic need not imply confidence in the unresolved valuation basis.

03 / Expert judgement

Use the support. Interpret the dependency.

The analyst uses the calculation and source trail to reason about the renewal, valuation basis and policy. One design shows an AI recommendation immediately; another asks for a provisional view first. Neither is automatically better. The useful design should support justified acceptance and challenge without making analysts redo dependable work.

Evaluate what the analyst believes, what they rely on and why those change.

04 / Recorded decision

A supported decision, with uncertainty in the right place.

Under this fictional policy, refer because the valuation depends on an unagreed renewal. Retain the useful calculation and source links, clarify the basis and record the analyst’s reason. A routine case with a supported basis may warrant a different path. Evaluation must establish whether the architecture helps distinguish them.

This illustrates a case decision, not evidence that the proposed workflow performs better.
£3.0mProposed loan
£4.8mValue on renewal assumption
62.5%Provisional loan to value

Starting point / the unresolved question

AI produces a useful calculation and a provisional recommendation, but does not make the boundary between supported facts and unresolved assumptions clear enough.

Candidate design / to evaluate

Connect extraction and calculation to source status, distinguish assumptions from established facts and give the analyst a clear role in interpreting the dependency. Evaluate whether that supports appropriate confidence and efficient reliance.

Inspect the source evidence

V1 + T1 / Valuation and tenancy

£4.8m assumes renewal on existing terms. Renewal is unagreed; no alternative valuation basis is supplied in this extract. The tenant supplies £240,000 of £400,000 annual rent (60%), with 18 months to expiry.

B1 / Broker note

The broker expects the tenant to remain. The note records an expectation, not an agreed renewal. The analyst needs to preserve that distinction when using the statement.

P1 / Fictional referral policy

Refer where the assessment depends on an unconfirmed renewal. Being below the illustrative 70% leverage cap does not remove that requirement. This invented policy makes the case assessable; it is not a recommended lending policy.

The question behind the example

Can analysts distinguish the AI work they can use from the judgement the evidence still requires?

The capability to enable

Use AI to connect documents, reconcile terms and prepare supported calculations, leaving analysts more capacity to interpret consequential dependencies.

Too much confidence

A correct figure or persuasive recommendation lends unwarranted certainty to an unresolved assumption. A sound initial judgement may be displaced.

Too little confidence

In a comparison case with a supported basis, the analyst repeats dependable work or refers unnecessarily. Evaluate whether doubt adds value or just consumes capacity.

Explore one architecture choice

Two plausible designs.
Different trade-offs.

These are choices within a wider decision architecture, not complete technical designs or measured results. Select an option to explore its contribution and trade-offs.

Recommendation first

AI prepares a draft and provisional outcome immediately.

The analyst receives “Accept, subject to review” alongside links to V1, T1 and P1. This may make routine work efficient, but the qualification competes with an already framed conclusion.

Watch for: acceptance of unresolved dependencies, quality of challenge and total review effort.

Judgement before suggestion

AI prepares evidence; the analyst forms a view before seeing the proposed outcome.

Show the renewal qualification, broker’s expectation and policy together. Ask what remains unresolved and whether it changes the decision. Then reveal the AI suggestion, capture any change of view and its reason.

Trade-off: potentially more deliberate challenge, but extra work. Test whether targeted use is preferable to requiring it on every case.

How Axrea would examine the contribution

Evaluate quality.
Understand confidence and reliance.

Evaluation examines the decision, how certain the expert feels and what they actually rely on. These are distinct. The criteria below would be agreed with your team, including the consequences of both overconfidence and underconfidence.

  1. Decision qualityAppropriate actions and defensible reasons across cases with and without a referral issue.
  2. Confidence in contextConfidence in the calculation, evidence status and final judgement assessed separately against agreed criteria.
  3. Appropriate relianceUse of supported AI work; challenge to unsupported conclusions; justified retention or revision of the analyst’s own view.
  4. Overconfidence and underconfidenceUnsupported certainty, unwarranted doubt and whether new evidence changes confidence appropriately. More confidence is not automatically better.
  5. Useful capacityPreparation, repeated checks, correction and senior review, weighed against decision quality and the consequences of each action.
Inside the evaluation design

Use a representative mix of routine and difficult cases, including cases with no referral issue. Agree expected actions and adjudicate professional disagreement before comparing designs. Use comparable reviewer groups or counterbalanced matched sets to limit differences in experience, learning and fatigue. Avoid showing the same person a case they can recall from the other condition. Record model, prompt and document versions. Choose the sample size for the question; report uncertainty and case-level failures rather than relying on an average improvement. Include an expert-only baseline when the question is whether AI adds value at all. Capture provisional judgements and confidence before and after AI support where appropriate to the comparison, while accounting for the effect of asking those questions. Include reliable and unreliable AI contributions. Assess confidence in specified claims against agreed reference criteria and report reliance separately from stated confidence. Do not label a necessary referral or a justified check as underconfidence merely because it takes longer.

Illustrative Axrea recommendation / what we would test

Design the evidence and judgement handover, then evaluate where it adds value.

Specify how AI retrieves the relevant clauses, reconciles the figures and preserves whether a statement is agreed, assumed or expected. Connect that evidence to the analyst’s reasoning and decision record. Within this architecture, compare recommendation timing while holding the evidence display and model constant. Use the results to decide where an independent first view helps and where supported preparation can be used without unnecessary repetition.

What your team would receive

An architecture and workflow specification

How sources, extraction, calculations and uncertainty connect to the analyst’s judgement and decision record.

A confidence and reliance evaluation

Representative cases, agreed criteria and a readout separating decision quality, expressed confidence and actual reliance.

An adoption and adaptation recommendation

What to delegate, where independent judgement adds value and what evidence would justify changing the arrangement.

As the technology and people evolve.

As AI improves at preserving qualifications, analysts may be able to delegate more preparation. Their own skills and expectations also develop. Revisit where confidence is warranted, whether useful capability is underused and whether familiar outputs are being accepted too readily.

Bring one workflow

Where could your analysts use AI more confidently, and where does the evidence still need their judgement?

In a 30-minute conversation, we will explore what your workflow could enable, how AI and expert judgement currently interact, and where architecture design or evaluation could help.

No client files needed. Prefer to get to know Gareth first? A LinkedIn connection is welcome.

02 / Valuation

A wider evidence base.
Stronger professional judgement.

AI can extend what the valuer sees and help compare competing evidence. The architecture should make those contributions usable while supporting the valuer’s interpretation, weighting and confidence.

Evidence + judgement
Comparable terms + source trail: AI capabilities and expert judgement contribute to the decision.Two coloured signals follow the evidence through AI and expert review, merge into a copper signal and briefly illuminate the recorded decision.COMPARABLE TERMS + SOURCE TRAILC1 / VERIFIED LETTING£32 headline rentHigher specificationSix months rent freeC2 / FOUND BY AICloser specificationOlder broker noteStatus to confirmAI CONTRIBUTIONAI contributesDiscover and compareEXPERT JUDGEMENTExpert judgesInterpret and weighRecord a defensible evidence viewPRESERVE DISCOVERY WITHOUT DELEGATING THE JUDGEMENT

Illustrative workflow · signals show contribution, not live AI processing

01 / The task

Recent evidence is not necessarily the closest evidence.

The valuer’s initial shortlist favours a recent £32 per sq ft letting. It has a higher specification than the subject and six months rent free over five years. An older broker note describes a lower-specification letting that could be more comparable, but its transaction status is unclear.

C1 · Recent letting / C2 · Older broker note / S1 · Subject schedule

02 / AI contribution

Extend the evidence available for judgement.

AI surfaces C2 outside the initial shortlist and brings C1’s incentive into view. Retrieval, source links and structured comparison can help the valuer explore material they might otherwise miss. The architecture must preserve that C2 is unconfirmed, so relevance is useful without being mistaken for reliability.

The contribution is broader discovery and comparison, with the limits of each source visible.

03 / Expert judgement

Use new evidence without surrendering the weighting.

The valuer investigates C2’s status, age and location, and interprets its closer specification. They may revise their initial view when the evidence warrants it, retain it with stronger reasons or seek confirmation. The architecture supports that reasoning; it should neither impose an AI ranking nor encourage dismissal of a useful discovery.

Confidence may rise or fall as evidence is confirmed. Both can be appropriate.

04 / Recorded decision

A wider evidence base with the uncertainty intact.

Retain C1 with its incentive and specification differences explicit. Seek confirmation of C2 before treating it as transactional evidence. Record the respective roles and weights, the unresolved questions and the evidence needed to conclude the subject’s rent.

No market rent or automatic adjustment is produced by this illustration.
£32.00C1 headline rent / sq ft / year
6 monthsC1 rent free over five years
£28.80Simple undiscounted average

Explore the incentive

Same headline. Different terms.

Change the rent-free period on this five-year lease.

0 months12 months
£28.80per sq ft per year

£32 × 54 ÷ 60 months

Simple undiscounted average only. It does not determine market rent or the appropriate valuation treatment. The specification difference still requires judgement.

Starting point / the unresolved question

The initial human shortlist favours recent evidence. AI could broaden it, but a ranked answer may replace one selection bias with another.

Candidate design / to evaluate

Use AI to search, reconcile and expose differences. Test whether the valuer can use newly found material well while retaining control of verification and weighting.

Inspect the source evidence

C1 / Verified comparable letting

Five-year lease; £32 per sq ft per year; six months rent free; no review, break or other incentive in this simplified case. The specification is higher than the subject’s.

C2 + S1 / Candidate evidence and subject

A broker note from nine months earlier describes a lower-specification property closer to the subject’s specification, in a different micro-location. It is unclear whether the terms were agreed. No rent from C2 is assumed here: its relevance and status need checking.

Calculation / Scope

£32 × 54 paying months ÷ 60 months = £28.80 per sq ft per year. This is a simple undiscounted average, not a formal net-effective rent, market rent or valuation. Appropriate treatment of timing, lease terms, location and specification remains a professional judgement.

The question behind the example

Can AI expand the valuer’s evidence and insight while supporting confidence that the evidence warrants?

The capability to enable

Broader retrieval and structured comparison give the valuer more relevant material and more capacity to interpret it. Evaluate what this adds beyond the existing shortlist.

Too much confidence

A relevance score is treated as proof of transaction reliability, or the familiar human shortlist is defended despite stronger contrary evidence.

Too little confidence

Useful AI discoveries are rejected simply because AI found them. In a comparison case where status is confirmed, repeated doubt can prevent justified use of the evidence.

Explore one architecture choice

Two plausible designs.
Different trade-offs.

These are choices within a wider decision architecture, not complete technical designs or measured results. Select an option to explore its contribution and trade-offs.

AI ranks the evidence

A ranked shortlist and suggested conclusion arrive together.

AI puts C1 first because it is recent, then includes C2 with a relevance score. The format is concise, but a score can obscure differences in source reliability and make the ranking an implicit instruction.

Watch for: anchoring to rank, unverified evidence gaining authority and relevant material being excluded.

AI searches; the valuer weights

AI presents candidate evidence, provenance and differences before proposing a ranking.

The valuer sees why C2 was retrieved, what is unconfirmed and how it differs from C1 and the subject. They form and record a provisional weighting before considering an AI critique or alternative view.

Trade-off: broader discovery and visible uncertainty may help, but a long evidence queue can consume expert time. Test filters as well as presentation order.

How Axrea would examine the contribution

Evaluate quality.
Understand confidence and reliance.

Evaluation examines the decision, how certain the expert feels and what they actually rely on. These are distinct. The criteria below would be agreed with your team, including the consequences of both overconfidence and underconfidence.

  1. Evidence and decision qualityRelevant discoveries, material omissions and defensible interpretation of incentives, differences and source status.
  2. Confidence as evidence changesWhether confidence responds appropriately to confirmation, contradiction or remaining uncertainty.
  3. Appropriate relianceUseful AI findings incorporated when justified; unsupported rankings challenged; the valuer’s own weighting retained or revised for sound reasons.
  4. Professional judgementReasoning and acceptable ranges assessed without treating legitimate differences between valuers as errors.
  5. Capacity after verificationRetrieval, confirmation, comparison and rework, including unnecessary repetition after evidence is sufficiently established.
Inside the evaluation design

Build a reference set with experienced valuers and preserve differences of opinion where evidence permits more than one defensible interpretation. Assess retrieval on both relevant findings and distracting candidates. Test cases where the familiar shortlist is adequate as well as cases where useful material lies elsewhere. For the presentation comparison, hold the evidence pack constant, separate test cases from development cases and limit repeated exposure. Compare experts using their existing approach with the proposed AI-assisted approach when estimating the overall contribution. Review explanations and evidence use alongside numerical conclusions; agreement alone does not establish quality. Include cases where AI adds relevant evidence and cases where its suggestions are weak. Compare provisional and subsequent weighting and confidence where the study design permits. Distinguish confidence in transaction status from confidence in the final valuation judgement. Use agreed evidence criteria and defensible ranges rather than imposing a single numerical answer. Examine underuse of confirmed evidence as well as undue reliance on unverified material.

Illustrative Axrea recommendation / what we would test

Design retrieval and comparison to support the valuer’s reasoning.

Connect retrieval to source provenance, transaction status and a structured comparison with the subject. Preserve the valuer’s role in interpreting differences and assigning weight, while allowing AI to offer alternatives and critique. Evaluate discovery separately from presentation, then examine the combined process: does it improve reasoning, confidence and use of relevant evidence after verification effort is included?

What your team would receive

An evidence and reasoning architecture

How retrieval, provenance and comparison support the valuer’s interpretation and weighting.

Evaluation of confidence and reliance

How judgement and confidence respond to evidence, alongside the quality of retrieval and the cost of verification.

A practical adoption recommendation

Which AI capabilities support useful work, where expert interpretation remains essential and when to revisit the design.

As the technology and people evolve.

New retrieval and comparison capabilities may broaden the evidence available. As valuers learn which contributions are useful, the division of work can change. Revisit confidence and weighting behaviour alongside retrieval quality, including both habitual deference and persistent underuse.

Bring one workflow

Where could AI broaden your valuers’ evidence and insight while supporting confidence in their own judgement?

In a 30-minute conversation, we will explore what your workflow could enable, how AI and expert judgement currently interact, and where architecture design or evaluation could help.

No client files needed. Prefer to get to know Gareth first? A LinkedIn connection is welcome.

03 / Investment

Connected evidence.
A more capable investment decision.

AI can help teams reason across delivery, leasing and income assumptions. A useful architecture connects that analysis to expert interpretation, proportionate confidence and the conditions that matter to the decision.

Evidence + judgement
Programme + leasing + cash flow: AI capabilities and expert judgement contribute to the decision.Two coloured signals follow the evidence through AI and expert review, merge into a copper signal and briefly illuminate the recorded decision.PROGRAMME + LEASING + CASH FLOWI1 / BASE CASEWorks complete M6 → Rent starts M7W1 + L1 / INDICATIVE, NOT COMMITTEDWorks M8 → Earliest occupation M9AI CONTRIBUTIONAI contributesMap and exploreEXPERT JUDGEMENTExpert judgesInterpret and decideConditions tied to the downsideTEST WHETHER THE CHALLENGE CHANGES THE PAPER

Illustrative workflow · signals show contribution, not live AI processing

01 / The task

The preferred sequence is not yet supported.

The base case assumes works complete by month six, with rent from month seven at £1.2m a year. A contractor’s indicative programme instead points to month eight, while the leasing note describes month nine as the earliest plausible occupation. Neither is committed. The team must reconcile the sources, not merely list delay as a risk.

I1 · Base case / W1 · Contractor programme / L1 · Leasing note

02 / AI contribution

Give the team a connected view of the decision.

AI links works completion to occupation and income, surfaces conflicting dates and helps explore coherent alternatives. This can give the team a clearer basis for reasoning across the case. The architecture must distinguish dependencies supported by sources from assumptions and generated possibilities.

A plausible scenario can inform judgement without establishing its probability.

03 / Expert judgement

Turn analysis into a view and actionable conditions.

The team interprets which assumptions can move together, examines a six-month delay scenario and identifies the delivery, leasing and funding conditions that matter. It uses the analysis to form or revise its view. Expert attention is directed to consequential uncertainty rather than every generated variation.

The rent timing calculation is one input. Costs, financing and the wider investment case still need analysis.

04 / Recorded decision

Make the conditions specific enough to act on.

Return the paper with source-backed timing assumptions, a coherent downside and explicit conditions for delivery, leasing and funding review. Record what would change the recommendation and who must resolve each open point. The committee retains the investment judgement.

The example proposes a better paper; it does not recommend buying, lending or proceeding.
Month 7Base-case lease start
£1.2mForecast annual gross rent
£600kTiming gap in a six-month delay

Explore the dependency

Move the letting. See the timing gap.

Hold annual gross rent at £1.2m and change the delay.

0 months12 months
£600kgross rent timing gap

£1.2m ÷ 12 × 6 months

Receipts absent from the original period, not necessarily permanently lost. Excludes incentives, costs, financing, discounting and any later recovery. No probability is assigned to the delay.

Starting point / the unresolved question

The AI-assisted paper follows the base case and adds a broad risk list. It does not reconcile the conflicting dates or distinguish connected assumptions from independent sensitivities.

Candidate design / to evaluate

Test a dependency-led challenge: AI maps sources and consequences, while experts select coherent scenarios and specify the conditions relevant to the recommendation.

Inspect the source evidence

I1 / Base-case model

Works complete by month six; rent begins in month seven at £1.2m gross per year. The simplified rent calculation holds annual rent constant after a delay.

W1 + L1 / Conflicting sources

The indicative contractor programme points to completion in month eight. The leasing note describes month nine as the earliest plausible occupation. Neither is committed. The six-month delay explored below is a separate scenario, not an inferred forecast or probability.

Calculation / Scope

£1.2m ÷ 12 × 6 = £600,000 gross rental receipts absent from the original six-month window. This timing gap is not necessarily a permanent loss, valuation change or investment return. It excludes incentives, costs, finance, discounting and any later recovery.

The question behind the example

Can the architecture turn connected evidence into useful reasoning and appropriately confident decisions?

The capability to enable

Connect source evidence and dependencies so the team can explore coherent alternatives, focus attention and specify meaningful decision conditions.

Too much confidence

A precise-looking output or familiar base case creates unjustified certainty. Generated scenarios are mistaken for forecasts with established probabilities.

Too little confidence

Useful sensitivity analysis is ignored because it cannot predict the future, or a supported decision is deferred while the team requests more analysis with little decision value.

Explore one architecture choice

Two plausible designs.
Different trade-offs.

These are choices within a wider decision architecture, not complete technical designs or measured results. Select an option to explore its contribution and trade-offs.

Generate a wider scenario pack

AI varies assumptions and produces an extensive sensitivity summary.

This may expose possibilities quickly. But independent changes to works and letting dates can duplicate the same delay, while a large pack can hide a conflicting source date behind polished charts.

Watch for: incoherent combinations, unsupported probabilities and scenarios that never inform a decision.

Start from the decision dependencies

AI maps source-backed dependencies; experts choose the challenges to examine.

Trace works completion to occupation and income. Resolve the conflicting dates, select coherent joint scenarios and identify which conditions could change the recommendation. Ask AI to search for omitted pathways as a further challenge.

Trade-off: a focused pack may use attention better, but expert selection can miss an unexpected dependency. Test omitted pathways and fresh case types too.

How Axrea would examine the contribution

Evaluate quality.
Understand confidence and reliance.

Evaluation examines the decision, how certain the expert feels and what they actually rely on. These are distinct. The criteria below would be agreed with your team, including the consequences of both overconfidence and underconfidence.

  1. Decision qualityConsequential dependencies understood, scenarios coherent and conditions relevant to the decision.
  2. Confidence and uncertaintyWhether certainty is proportionate to the support for key assumptions and changes as material evidence changes.
  3. Appropriate relianceAI analysis used where informative; unsupported probability claims challenged; expert views revised or retained with reasons.
  4. Useful actionWhether analysis informs conditions and choices, including cases where further analysis or deferral is genuinely warranted.
  5. Attention and effortPreparation, verification, discarded analysis and committee time, considered alongside the quality and consequences of the decision.
Inside the evaluation design

Agree what constitutes a material issue with experienced investment reviewers before testing. Use cases with different delivery and leasing structures, including cases where a simple base case is adequately supported. Hold source information and review time comparable, limit recall between conditions and keep development cases separate. Assess whether reviewers identify defensible decision conditions rather than whether they all reach the same recommendation. Review cases where a dependency map narrows attention too much. Record the model and workflow version so an improvement in drafting is not confused with a change in the underlying investment judgement. Ask for confidence in specific assumptions and decision conditions rather than a vague confidence score for the whole investment. Distinguish confidence in a decision under uncertainty from a probability forecast. Test responses to stronger and weaker evidence, useful and unsupported AI contributions, and cases where the existing approach is adequate. Deferral can be appropriate; assess its justification rather than treating every delay as underconfidence.

Illustrative Axrea recommendation / what we would test

Connect evidence, dependencies and expert reasoning before expanding the analysis.

Design a source-linked dependency model that separates assumptions, supported facts and unresolved conditions. Use AI to explore coherent alternatives and help the team identify what could change its view. Evaluate whether the resulting workflow improves decision reasoning and appropriate confidence, including whether useful analysis is acted on and whether extra scenarios justify the attention they consume.

What your team would receive

A decision architecture

How source evidence, dependency analysis, AI exploration and expert reasoning connect to the committee’s decision.

An evaluation of the combined process

Decision quality, confidence, reliance and attention measured across credible alternatives.

An adaptation recommendation

Which capabilities to use, what uncertainty remains and what changes should trigger further evaluation.

As the technology and people evolve.

As AI supports richer reasoning, the team may delegate more analysis and concentrate on interpretation and decision conditions. Revisit whether confidence remains justified and whether new capabilities are used effectively as models, transactions and committee experience change.

Bring one workflow

Where could better connections between AI analysis and expert judgement improve your investment decisions?

In a 30-minute conversation, we will explore what your workflow could enable, how AI and expert judgement currently interact, and where architecture design or evaluation could help.

No client files needed. Prefer to get to know Gareth first? A LinkedIn connection is welcome.

Research applied to the way work gets done

AI architectures.
Stronger expert decisions.

Axrea brings AI research and two decades of real estate advisory experience to architecture design, evaluation and continuing advice. We connect evidence, AI capabilities and human judgement, and establish when confidence in the combined process is warranted.

Founder Gareth McConomy is a chartered surveyor (FRICS) and AI researcher with several peer-reviewed AI papers.

Research informing the design questions

These examples illustrate design choices to test, not established results in real estate. Buçinca, Malaya and Gajos (2021) examined interventions to reduce overreliance, including usability trade-offs. Vaccaro, Almaatouq and Malone (2024) examined when human–AI combinations help. Schemmer and colleagues (2023) distinguish appropriate reliance on AI advice from overreliance and underreliance. These studies inform evaluation questions; they do not establish the outcomes of these fictional real estate cases.