axrea.

The thinking

How experts and AI decide well together.

Too much trust in AI is a risk. Too little is a waste. Getting between the two is not luck. It is design, grounded in decision science and peer-reviewed AI research.

Hybrid intelligence

The idea, in three sentences.

  1. Hybrid intelligence is experts and AI working on the same decision, each doing what it does best.

  2. The choice is not expert or AI, and it is not AI bolted on to the old way of working.

  3. For every kind of decision there is a right split between the two, and it can be designed.

Putting the two together is not enough. A meta-analysis of 106 experiments found that on average a human and an AI together performed worse than the better of the two alone, with losses on decision tasks [1]. And AI is uneven: in one field experiment, consultants did better with it on tasks inside its reach and worse on one outside it [2]. How they are put together is the work.

Calibrated confidence

Rely on the AI where it has earned it. Challenge it where it has not.

Calibrated confidence is trust in the AI that matches the evidence [3]. It can fail both ways, and at three levels: the expert relying on the AI, the firm's leaders investing in it, and the committee, the lender or the client relying on the decision. It is settled task by task, not once.

Illustrative: one kind of decision, task by task

Trusted more thanit has earnedA risk Better than your people thinkA waste How reliable the AI is at the task → How much your people rely on it → 1Reading every lease2Extracting terms and dates3Reconciling the rent roll4What is missing5Drafting the paper6Choosing the yield7Citing its sources

Trust and reliability rarely match.

One kind of decision, task by task. The AI is strong at some things and weak at others, and your people's trust does not follow the difference.

  • Above the line. Polished drafts and silent gaps, taken on trust.
  • Below the line. Sound work by the AI, redone by hand.

Confidence that follows the evidence.

Calibration is not one setting. It is task by task: delegate what the AI does well, and put your experts where it does not.

  • On the line. Each task relied on as far as it has earned, and no further.
  1. 1Reading every lease
  2. 2Extracting terms and dates
  3. 3Reconciling the rent roll
  4. 4What is missing
  5. 5Drafting the paper
  6. 6Choosing the yield
  7. 7Citing its sources

From confidence to a decision

Confidence is the foundation. Four steps are built on it.

Knowing where the AI can be relied on is the start. A team still has to decide: price the asset, approve the loan, sign the valuation. We work one decision at a time, in four steps.

  1. Measure the decision

    On your own past deals, loans or valuations, where your senior people can agree what a sound decision looked like: what your team decides with AI, and without.

    Did the decisions improve?

  2. Design the way of working

    Who does what, what must be opened and checked, and what is recorded when someone commits the firm.

    Where can we rely on it, and where must we look?

  3. Show what it is worth

    In price, loan to value, loss or claim, worked on your own figures.

    What is a better decision worth to us?

  4. Re-test

    Tools and habits move. The same past decisions, run again, show when the way of working has stopped being right.

    Who tells me when it drifts?

What each step gives you

What happens to judgement

Seven things change when AI joins the work. None is a fault in the tool.

Each is still there when the AI is accurate. Some have been studied for decades. Others we name ourselves, because property work makes them plain.

  1. Reliance

    People follow a system's suggestion, even when the evidence in front of them says otherwise. [4]

    LendingA credit analyst accepts a pre-filled lease term that the tenancy schedule on the next page contradicts.

  2. Anchoring on the draft

    The first number or the first draft pulls every later judgement towards it. [5]

    InvestingA deal team handed a draft paper with an exit yield already in it starts from that yield, and tests the evidence for it too lightly.

  3. The default

    The option already selected wins more often than it should, because changing it takes a reason. [6]

    ValuingA valuation arrives with every comparable already chosen and weighted, and leaving one out becomes the choice that needs defending.

  4. The unread source

    Checking a summary replaces reading the document. The better the summary looks, the less often anyone goes behind it.

    ValuingA tenancy schedule is adopted from the lease summary, and the break option in the side letter is not in it.

  5. Agreement without accuracy

    People who start from the same draft agree more, and the agreement is taken for quality.

    InvestingTwo analysts reach the same rental growth assumption because both began from the same market summary.

  6. Accountability without authorship

    The person who commits the firm stays accountable for a view that something else formed.

    LendingA credit officer approves the loan and could not say which statements in the paper were tested and which were taken on trust.

  7. Skill no longer practised

    A skill that is not used fades, and with it the ability to catch the AI when it is wrong.

    ValuingSurveyors who no longer read a lease in full lose the eye for the clause that changes the value.

What can be designed

Six things a team can design.

Each is aimed at particular effects. Each has a price, and each can backfire. That is why the choice belongs to your experts.

Lever 1

Who commits first

The expert reads the key terms and records a view before the AI's draft is shown. [7]

Aimed at
Reliance and anchoring. A view formed first is harder to overwrite.
It costs
Expert minutes on the items read first.
How it can backfire
The first read shrinks to a token entry. Or a correct draft is overruled by a wrong first view.

Lever 2

What the AI shows, and when

The AI shows the facts first and holds its conclusion back. [8]

Aimed at
Anchoring, and agreement without accuracy. The expert meets the evidence before the answer.
It costs
One more step on every case.
How it can backfire
The conclusion is a click away, and the click becomes the first thing anyone does.

Lever 3

Marking what was not found

The AI marks each required item found or not found, and the expert follows up every gap.

Aimed at
The unread source. Silence stops looking like "nothing to report".
It costs
A required list that your experts own and keep current.
How it can backfire
Too many "not found" marks, and people stop looking at them.

Lever 4

No default on the decision

The AI fills in the facts and leaves the recommendation blank. The expert chooses, and says why. [6]

Aimed at
The default. Nothing is accepted by leaving it alone.
It costs
A sentence of reasoning on every decision.
How it can backfire
The reason becomes a phrase typed by rote.

Lever 5

Reasons on the record

Before committing the firm, the expert lists what they relied on.

Aimed at
Accountability without authorship. The person who decides can say what the decision rests on.
It costs
A short list, written on every decision.
How it can backfire
The list becomes a checklist, ticked unread.

Lever 6

Where the expert must stop and choose

The work pauses at the items your experts have named. The expert opens the source and decides each one.

Aimed at
The unread source, and skill. Judgement is practised where it carries the decision.
It costs
Time, on purpose, on a few items in every case.
How it can backfire
Too many stops, and each one turns into a click-through.

What we design and build

Built around your experts.

Where the work needs it, we design and build what helps your experts decide. Each rests on an established method.

Every statement tied to its source

Every figure in a paper or a report linked to the lease, the letter or the evidence it came from, and the gaps shown before anything is read. [9]

Your own past decisions, ready to compare

Similar deals, loans or valuations from your own record: what was assumed, what was decided, and what then happened. [10] [11]

A clear view of where the AI can be relied on

Task by task, so your people know where to rely and where to look. [3] [12] [11]

A way to re-test when anything changes

Your own past decisions, run again whenever the tool, the data or the work changes. [13]

What we build for you is yours.

How we know

Research is a reason to try something. Your own work is the proof.

Much of this research comes from other fields. Very little has been run on valuers, lenders or deal teams. So we test today's way of working on your own past deals, loans and valuations, change it, and test again.

Your cases

Past decisions, from your own record

Routine and awkward, with their source documents. They stay in your environment.

Your senior people

They say what a sound decision looked like

They fix the answers before the AI is run on a single case, and they judge the results.

Your choice

You see what changed, and what it cost

Every design has a price. If a change does not help, we say so.

What a property decision is worth

The sums, and what the published record says.

The first two figures here are arithmetic. The third is published, and so are the findings beneath them. Each is given as its publisher states it.

Investing

The price you pay.

£2.4mA quarter point on the exit yield. Net income of £2.5m is worth £50.0m at 5.00% and £47.6m at 5.25%.

Arithmetic, not a survey. Put your own income and yield in its place.

Lending

Whether to lend, and how much.

60% to 70.6%The loan to value on a £30m loan, if the property is worth 15% less. Arithmetic.

Around 6% of UK banks' stock of commercial property debt was written off between 2008 and 2012, cumulatively. [14]

Valuing

The figure others rely on.

7.0%UK sale prices differed from the valuation, rolled forward for the market, by 7.0% on average in 2025 and 8.8% a year since 2000. [15]

MSCI's figures, on MSCI's basis: value-weighted, with gaps of 50% or more left out.

Supervisors
Bank supervisors found headline rents used with rent-free periods ignored, evidence from 2021 used unadjusted in 2022 and 2023, reliance on valuations of property not properly inspected, and over-reliance on automated models. European Central Bank, August 2024. [16]
Bought-in AI
24% of firms rely on deep knowledge of their important third-party AI systems. 81% rely on the vendor's assurance and contracts. A survey of 196 UK financial services firms by the Bank of England and the FCA, published on 26 October 2026. It is self-reported, and it covers financial services as a whole, not property lending alone. [17]
The duty
The professional standard on the responsible use of AI in surveying practice has been in force since 9 March 2026. Where an AI output has a material impact, it requires a written decision on its reliability. It gives no method for reaching one. [18]

References

Where this comes from.

The work draws on published research in AI assurance, calibrated confidence and architectures for expert decision support, including our founder's own [11] [13]. The property figures come from the publishers named in entries 14 to 18. The numbers in the text point here.

  1. When combinations of humans and AI are useful: a systematic review and meta-analysis Vaccaro, Almaatouq and Malone (2024). Nature Human Behaviour, 8, 2293 to 2303. A meta-analysis of 106 experiments. On average, human and AI combinations performed worse than the better of the two alone, with losses on decision tasks. doi.org/10.1038/s41562-024-02024-1
  2. Navigating the jagged technological frontier: field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality Dell'Acqua, McFowland, Mollick, Lifshitz, Kellogg, Rajendran, Krayer, Candelon and Lakhani (2026). Organization Science, 37(2), 403 to 423. A field experiment with 758 consultants. With AI they did better on tasks inside its reach, and worse on one outside it. doi.org/10.1287/orsc.2025.21838
  3. Trust in automation: designing for appropriate reliance Lee and See (2004). Human Factors, 46(1), 50 to 80. On trust that matches what a system can do. doi.org/10.1518/hfes.46.1.50_30392
  4. Complacency and bias in human use of automation: an attentional integration Parasuraman and Manzey (2010). Human Factors, 52(3), 381 to 410. On reliance. doi.org/10.1177/0018720810376055
  5. Judgment under uncertainty: heuristics and biases Tversky and Kahneman (1974). Science, 185(4157), 1124 to 1131. On anchoring. doi.org/10.1126/science.185.4157.1124
  6. Do defaults save lives? Johnson and Goldstein (2003). Science, 302(5649), 1338 to 1339. On the default. doi.org/10.1126/science.1091721
  7. To trust or to think: cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making Buçinca, Malaya and Gajos (2021). Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1). On who commits first. doi.org/10.1145/3449287
  8. Does the whole exceed its parts? The effect of AI explanations on complementary team performance Bansal, Wu, Zhou, Fok, Nushi, Kamar, Ribeiro and Weld (2021). Proceedings of the CHI Conference on Human Factors in Computing Systems. On what the AI shows. Explanations made people more likely to accept the AI's answer, right or wrong. doi.org/10.1145/3411764.3445717
  9. Knowledge graphs Hogan, Blomqvist, Cochez and colleagues (2021). ACM Computing Surveys, 54(4). On tying statements to their sources. doi.org/10.1145/3447772
  10. Case-based reasoning: foundational issues, methodological variations, and system approaches Aamodt and Plaza (1994). AI Communications, 7(1), 39 to 59. On reasoning from past cases. doi.org/10.3233/AIC-1994-7104
  11. Layer-appropriate augmentation for human-AI risk assessment: integrating case-based reasoning with fuzzy calibration McConomy, Bond, Glass, Liu and Toner. HHAI. On case-based structure and calibration tested together. Read the paper: layer-appropriate augmentation
  12. Effect of confidence and explanation on accuracy and trust calibration in AI-assisted decision making Zhang, Liao and Bellamy (2020). Proceedings of the Conference on Fairness, Accountability, and Transparency. On showing confidence. A confidence score helped people calibrate their trust, and that alone did not improve the joint decision. doi.org/10.1145/3351095.3372852
  13. Knowledge-generated test suites for LLM components: specification-derived targets with human-certified stimuli McConomy, Bond, Glass, Liu and Toner. AICS. On test cases built from structured knowledge, with people checking the cases. Read the paper: knowledge-generated test suites
  14. Commercial property and financial stability Bank of England (2013). Quarterly Bulletin, 2013 Q1, page 49. Around 6% of UK banks' stock of commercial property debt was written off between 2008 and 2012, cumulatively. bankofengland.co.uk: commercial property and financial stability
  15. Valuation and Sale Price Comparison Analysis 2025 MSCI (June 2026), with an MSCI note of 14 August 2026. UK sale prices differed from the valuation, rolled forward for the market, by 7.0% on average in 2025 and 8.8% a year since 2000. The comparison is value-weighted, sets each sale against a valuation rolled forward for the market, and leaves out gaps of 50% or more. msci.com: the MSCI note
  16. Supervision Newsletter, 14 August 2024 European Central Bank, banking supervision (2024). Supervisory findings: headline rents used with rent-free periods ignored, evidence from 2021 used unadjusted in 2022 and 2023, reliance on valuations of property not properly inspected, and over-reliance on automated models. bankingsupervision.europa.eu: the newsletter
  17. Artificial intelligence in UK financial services 2026 Bank of England and Financial Conduct Authority (26 October 2026). Section 7.3. A survey of 196 firms, self-reported. 24% of firms rely on deep knowledge of their important third-party AI systems, and 81% rely on the vendor's assurance and contracts. bankofengland.co.uk: artificial intelligence in UK financial services 2026
  18. Responsible use of artificial intelligence in surveying practice RICS. In force from 9 March 2026. Section 4.2. Asks for a written decision on reliability where an AI output has a material impact, and gives no method for reaching one.

Worked examples

See all three decisions.

A price, a loan and a valuation, with a review to try in each. The pack for your team adds what the decision is worth and the way of working we would design.

Try the worked examples

Thirty minutes with our founder, not a salesperson. Bring one kind of decision, with no client details, and we will sketch how it would be measured.

Arrange the 30 minutes