A two-second decision on credit joins several distinct tasks: verifying data, estimating risk, a legal assessment of creditworthiness and telling the applicant the result. An accurate model can work with a faulty input, and a low share of non-performing loans says nothing about the people the system refused in advance.
What we know: A lender must assess creditworthiness from information that is necessary, reliable, sufficient and proportionate. Automated decision-making may fall under the GDPR, and some systems for assessing creditworthiness fall into the high-risk regime of the AI Act.
What we do not know: Public sources do not show a complete list of variables, thresholds or human-review rules for individual Czech lenders. The use of an address, a device or geolocation therefore cannot be attributed to the market as a whole.
1. A score that is not yet fairness
Picture a meeting in a mid-sized Czech bank. The head of retail wants to cut consumer-credit approval to under two minutes, because competitors already do so. The risk manager insists on the model that best predicts twelve-month default on historical data. Compliance recalls Article 22 GDPR, the anti-discrimination act and the European Artificial Intelligence Act. IT wants to connect Open Banking, because movement on an account says more than a payslip. Marketing promises approval on the first click. The customer, who is not in the room, wants to know why they received a no.
At first glance this is an argument about speed. In fact each person is describing a different problem. The director compares application conversion. Risk compares loss from non-payment. Compliance compares procedural defensibility. IT compares the availability of data. The customer compares the intelligibility of a verdict that will close the door to housing, a car or a roof repair.
A model can predict default well and still fail at justification or at handling a faulty figure. Assessment must therefore separate the prediction of default, the quality of inputs, the risk of discriminatory proxies, the possibility of correction, the legal basis for automation and genuinely available human review.
This is not an abstract contest between a human and a machine. What is being compared is a concrete prediction, a decision rule, an explanation to the applicant and a path to remedy; each part has a different test.
A model is not fairness. A score is not character. A threshold is not the law. And the speed of a decision is not proof that the decision will stand.
— Jiný Kontext
2. What a credit verdict is made of
When “approved” or “rejected” appears on the screen, it is easy to form the impression that the model delivered the verdict. The result usually arises in a chain of several layers. Each can raise quality, lower it or change the kind of error. Two lenders with the same foundation therefore need not decide alike, and the same lender may behave differently by channel, product and whether a human still sits beyond the threshold.
Five layers of one credit verdict
- Data: registers, the account and the form; in some systems a device, geolocation or Open Banking may be added. Their actual use must be documented at the particular lender.
- Model: a statistical or machine prediction of the probability of default.
- Threshold and policy: from which score lending occurs, at what rate, with what exception.
- Permissions and channel: what the system may see, trigger, write and to whom it may show it.
- Governance and accountability: audit, human review, justification, correction of data, complaint.
In the scoring chain the model creates only part of the result. An application may insert a business rule before the score, add a channel surcharge after the score, apply currently valid regulatory or internal availability limits and send the file for manual approval. The same number in another product means a different rate. The same rate in another month may mean a different commercial policy.
This is precisely where the first error of public debate arises. A headline announces a contest between a human and AI, but in fact it compares two entirely different chains: a manual file with incomplete supporting documents and an automated system with thousands of variables, of which the applicant sees three. Or, conversely, it changes only the marketing word “AI” and presents an old logistic regression as new intelligence. An honest comparison must state not only the name of the model, but also the version, the channel, the available data, the threshold, the time limit and the method of control.
This layering is not an academic detail. It determines where to turn when something fails. If the model punished a postcode, the problem is in the data and the proxy. If it punished an evening time of submission, the problem is in telemetry. If it punished correctly, but nobody could say so, the problem is in the justification layer. The sentence “the algorithm rejected it” is the start of a diagnosis, not its end.
Illustrative scheme: The same model can, at a different threshold, channel and supervisory regime, create a different human fate.
3. The best system is the one whose error you can afford
Alongside predictive performance it is necessary to watch the type of error, its detectability and the possibility of review. An unfavourable rate can sometimes be changed by refinancing. An unexplained rejection may exclude the applicant from the next process and at the same time hide them from the statistics of customers to whom credit was granted.
NIST, in its framework for managing artificial-intelligence risk, notes that bias can appear in the data, in the objective, in the threshold and in whom the system measures at all. Risk does not grow only with the frequency of error. It also grows when the professional form of a verdict invites disproportionate trust.[13]
| Type of failure | What it looks like | How to test it | What limits the harm |
|---|---|---|---|
| Proxy discrimination | The verdict punishes an address, a device, a contract type or a time instead of the ability to repay | Comparison of groups with the same income; proxy ablation | A ban on selected features, a disparity audit, human review |
| Data error | Outdated income, someone else’s address, an uncorrected register | A list of the data used and the date of their validity | A right to rectification, short retention, an owner of the datum |
| Opacity | “Internal scoring” without the decisive factors | Whether the applicant can challenge the premise | A meaningful justification, documentation for supervision |
| Threshold error | A correct score, an arbitrary boundary, a channel exception | The same score in another channel and month | A written policy, a ban on silent exceptions |
This also changes the question we are asking. We do not ask only which system hit default on average. We ask which system will fail in a way that our process will catch. A risk team may tolerate false-positive caution, because it reduces loss. The applicant pays for that caution with a closed market. An accounts department cannot tolerate a hidden change of a decimal sign. A credit model cannot tolerate a hidden penalty on a neighbourhood if the company claims that it must not discriminate.
AUC itself does not tell whether the applicant will receive an intelligible justification, whether an input can be corrected and whether limited proxies lead to indirect discrimination. A somewhat weaker prediction may be more acceptable in a particular deployment if it has demonstrably effective control mechanisms.
4. What the share of non-performing loans does not show
Market indicators are needed, because they replace part of impressions with a repeatable number. The problem arises at the moment when one percentage begins to stand for the entire fairness of the system.
According to the banking statistics of the Czech National Bank, the volume of loans granted to resident households in the Czech Republic in June 2026 reached CZK 2,713 billion. Housing loans accounted for CZK 2,081 billion, that is 77 per cent of this volume. Consumer loans reached CZK 428 billion. The volume of non-performing loans remained at CZK 32.9 billion and their share of the total volume stayed at 1.2 per cent.[2] This is the state of the banking sector for one month, not the sum of all contracts in the CBCB and NRKI registers and not an audit of the fairness of individual models.
| Segment (CNB, June 2026) | Volume | Note |
|---|---|---|
| Loans to resident households | CZK 2,713 bn | Banking statistics, not registers |
| Housing loans | CZK 2,081 bn | 77 % of the volume to households |
| Consumer loans | CZK 428 bn | The highest month-on-month dynamism |
| Non-performing loans | CZK 32.9 bn | Share 1.2 % |
Every test contains hidden decisions. Who is in the population? Only banks, or non-bank lenders as well? What is the definition of default and how long is the observation window? How many applications were rejected before a contract arose? An answer to any of these questions can change the meaning of the number.
Five figures without which the NPL share is only an advertising number
- The precise population of contracts, the period and the source of the statistic.
- The definition of a non-performing loan and the length of the window.
- The share of rejected applications that never entered the portfolio at all.
- The difference between groups with the same income and a different data trail.
- The share of rejections that passed through human review.
At a large number of applications, automation can reduce time and costs substantially. Public statistics do not, however, show that full automation is economically inevitable for every product and every application. Portfolio statistics, moreover, see those who received credit, not all who were refused. A public indicator helps in understanding the size of the market; fairness of entry requires a different test.
5. Compare chains, not the “AI” mascot
The table separates the form, the registers, the model and human review according to their role in the credit verdict. For each layer it states the context and the failure that a test of the particular lender should look for. The state of models and policies changes; what decides is always the current configuration of the given product.
| Layer | Main working position | What to verify in a pilot | Typical reason for the choice |
|---|---|---|---|
| Registers (CBCB, NRKI, Solus) | A shared memory of contracts | Correction of an entry, age of the datum, coverage | Reducing information asymmetry |
| Form and income | What the applicant can document | Date of validity, individualisation | Statutory creditworthiness |
| Open Banking | Movement on the account with consent | Scope of items, retention, withdrawal | A fresher picture of solvency |
| Telemetry | Manner of completion, device, time | Whether a person can change the feature without changing repayments | Fraud and impulse |
| Threshold and rate | Commercial policy | The same score in another channel | Managing portfolio risk |
| Human review | Exception and complaint | Whether a human may say yes | A high price of error |
The rows are not a judgment on the quality of banks. The same lender offers several products, thresholds and channels, and each of them may change within a few months. The map is meant to prevent another error: that an institution buys an excellent model into a process that does not see the particular person, cannot explain the verdict or locks it into an environment from which it is later hard to leave.
How to read the scheme: Neither the position nor the size of the elements expresses a score. The diagram separates input data, the model, the decision rule and review.
6. Proxies: a ban on a feature is not a ban on information
The anti-discrimination act prohibits, in designated areas, different treatment on enumerated grounds; the use of a particular feature or its substitute variable must be assessed in the precise legal and factual context.[14] A model can, moreover, derive from other data information that correlates with a protected characteristic. The risk is called proxy discrimination.[4]
A postcode can, in particular data, correlate with social status, income or the composition of a neighbourhood. The kind of purchases, the time of transactions, the sector or the type of contract can function as proxies and create differences between groups. Whether this actually happens and whether the difference is legally permissible must be shown by a test of the particular model; this article does not document the use of these features by Czech banks.
A model can, when learning from historical data, reproduce or amplify existing differences if design, testing and supervision do not catch them. That is not an automatic result of every algorithm, nor proof of the programmer’s intent.[4][12]
Barocas and Selbst, in the study Big Data's Disparate Impact, showed that a discriminatory result need not arise from a programmer’s bad intent. It is enough for the model to learn from data that already contain inequality, and for the target variable — here default — itself to be socially biased.[4] Fuster, Goldsmith-Pinkham, Ramadorai and Walther showed on credit markets that a shift to machine learning can increase predictive power and at the same time widen differences between groups. This is empirics of a particular market and particular models, not proof that every Czech bank decides in this way. It does show, however, that accuracy and equality are not the same thing.[5]
Three tests of proxy discrimination
The first is a test of comparable abilities: do two groups with the same documented income, the same repayment history and the same requested amount have the same chance of approval? The second is an ablation test: does the disparity change when we remove the postcode, the device type and the time of submission from the model? The third is a substitution test: if we remove the proxy, does the model replace it with another variable that carries the same information? A product that “does not use race” need not be fair. It may merely be using an address, a school and a telephone.
An illustration, not reportage: applicant A, 32 years old, employment of indefinite duration, net income CZK 42,000, a new-build in the centre, an iPhone, an application in working hours, a loan of CZK 300,000, a rate of 5.9 per cent in 90 seconds. Applicant B, 32 years old, a self-employed person (OSVČ) with a net income of CZK 45,000 and five years without delay, a prefab on the periphery, an older laptop, an evening application, a rejection without a reason. Both have the same ability to repay. What differs is the data trail. If after the removal of the postcode and the device the verdict aligned, it was not character. It was a proxy.
7. Telemetry: typing speed is not character
Alongside registers and income, some systems may work with behavioural data about the completion of an application: time on the page, number of corrections, the device or the time of filing. Whether a particular Czech lender collects these data and uses them for the credit verdict must be documented separately.
A behavioural feature may in training data correlate with fraud or risk, but correlation itself does not explain the reason. An evening filing may be connected with shift work or care just as much as with financial distress. It is therefore necessary to measure the contribution of the feature, differences between groups and the possibility of correction.
Device type is not a direct measure of solvency. It may carry side information about income, age or digital behaviour and become a proxy. Use itself does not yet prove discrimination; what decides are the data, the weight of the feature, the purpose and the impact on comparable applicants.
The Basel Committee and related BIS work note that machine learning in financial services changes not only accuracy, but also the way institutions understand their own decisions. The more complex the model, the larger the gap between predictive power and explainability.[1] The gap is not cosmetic. It is the place where accountability is lost.
The practical minimum is a map of telemetry. Which behavioural features does the model actually use? Which does it only collect? Which can the applicant change without changing their ability to repay? If a feature can be changed by swapping a telephone, it is not character. It is a costume.
8. Registers: the market’s memory is not a person’s memory
CBCB, NRKI and Solus perform a useful function: they reduce information asymmetry between creditor and debtor.[3] Without them part of the market would guess in the dark and the price would rise for everyone. Registers are not, however, a person’s memory. They are a memory of the contracts the market has concluded about the person.
Memory has gaps. A person who has never applied for credit has a thin trail. A thin trail is often read in a model as risk, even though it may mean caution. A person who years ago had a dispute with an operator carries a mark longer than their present solvency lasts. A person who repaid on time, but in another regime — abroad, a partner’s account, cash — “does not exist” in the register.
An error in a register is especially expensive because it spreads. One bad entry enters several models. Several models produce several rejections. Several rejections thin the future trail. A loop arises: the market’s bad memory becomes the person’s bad future.
That is why the right to rectification of historical data is a practical, not a ceremonial, question. An applicant who does not know what the registers claim about them cannot challenge the verdict. An applicant who knows, but has no channel for correction, is challenging a wall. The Consumer Credit Act requires the lender to assess creditworthiness; it does not require it to let the applicant look into every weight of the model.[6] The gap between the duty to assess and the duty to explain is precisely the place where fairness is lost.
Registers are therefore usefully understood as infrastructure. Infrastructure needs an audit, retention periods, a right to rectification and an intelligible extract. Without that, “internal scoring” is only a politer name for someone else’s memory that you are not allowed to read.
9. GDPR and SCHUFA: an explanation is not open source
Article 22(1) GDPR gives the data subject the right not to be subject to a decision based solely on automated processing if it has legal or similarly significant effects. It is not an absolute ban. Paragraph 2 permits necessity for the conclusion or performance of a contract, authorisation by Union or Member State law, and explicit consent. For contractual necessity and explicit consent, paragraph 3 requires at least the possibility of human intervention, the expression of a view and a challenge to the decision; a legal exception must itself lay down suitable safeguards.[8] The Court of Justice of the EU in Case C-634/21 (SCHUFA) held that the calculation of a score can itself constitute such a decision if a third party attributes a determining role to it in its own decision-making.[7]
That is sharper than it sounds. The verdict need not be delivered by a robot with a stamp. It is enough for a human employee in practice merely to copy the score. The Court thereby shifted attention from the theatre of “we have a human in the loop” to the question of whether a human actually decides. This is an interpretation of one provision on one type of scoring, not an automatic ban on every banking model in Czechia. It does show, however, where the law looks for a decision: where the score in practice becomes the verdict.
The European Data Protection Board, in its guidelines on automated decision-making, emphasises meaningful information about the logic involved, the significance and the envisaged consequences.[16] This need not be a printout of all coefficients, but the sentence “internal scoring” itself usually does not make it possible to challenge an input or a decisive factor — for example outdated income, someone else’s address or a variable with an impermissible discriminatory impact.
In practice an open question remains: how to explain the output of a non-linear model so that it is intelligible to a lay applicant and at the same time does not disclose a trade secret? That is not a reason to abandon explanation. It is a reason to separate an explanation for a person from an audit for supervision. The customer needs the decisive factors and the possibility of correction. Supervision needs documentation, disparity tests and evidence that threshold values are not arbitrary.
10. The AI Act: high risk is not finished fairness
Annex III of the AI Act ranks among high-risk AI systems those intended to evaluate the creditworthiness of natural persons or to establish their credit score; an express exception covers systems used to detect financial fraud.[9] The regime includes requirements for risk management, data, documentation, logging and human oversight. Article 26 imposes obligations on deploying entities according to their role and conditions of use. Precise applicability to a given system and date must be verified against the regulation’s timetable; transparency under Article 50 is not a substitute for the high-risk-systems regime.[20][21]
Classification is a map of duties. It is not automatically fairness at the branch. A bank can have documentation and still punish a postcode. It can have human oversight and still let a person merely click “I agree”. It can have logs and still not show them to the applicant.
The difference between a consumer chat and a credit model is precisely the price of error. A bad slogan can be deleted. A bad mortgage rejection is written into a life course. That is why it is not enough to say “we have the AI Act”. It is necessary to say which layer of the chain is under supervision, who may return the verdict and how it is measured that it is also returned to people who are not the average of the portfolio.
Cathy O'Neil described the same logic as a scalable, opaque model that punishes people for the statistics of their surroundings and at the same time presents itself as neutral science.[12] The book is not an audit of a Czech bank. It is a description of a mechanism that the AI Act therefore placed among the high-risk ones.
11. The CNB and the EBA: prudence is not individual fairness
The European Banking Authority, in its guidelines on loan origination and monitoring, requires institutions to understand their models, to assess creditworthiness on the basis of proportionate information and to supervise the quality of newly originating contracts.[10] The aim is that new loans be prudent and that practices be aligned with consumer protection.
Act No. 257/2016 Coll. requires the lender to assess creditworthiness on the basis of necessary, reliable, sufficient and proportionate information.[6][11] For housing loans it is also necessary to separate currently effective CNB measures from the creditor’s internal limits; LTV, DTI and DSTI cannot, without a date of effect, be presented as one permanently binding set.[22]
Regulation thereby protects the market from systemic collapse and the consumer from a loan that will destroy them. It cannot, however, by itself guarantee individual fairness for every rejected applicant. Prudential supervision sees risk in the portfolio. Anti-discrimination logic sees risk in disparity between people with the same ability to repay. Both axes are legitimate. They are not interchangeable.
The EBA, in discussions of machine learning in internal rating models, notes that complexity must not deprive the institution of understanding.[18] The OECD, in work on artificial intelligence in finance, emphasises governance, auditability and human accountability, not a ban on prediction.[17] The CNB Financial Stability Report maps the credit cycle and household vulnerabilities; it does not measure whether a model punishes the self-employed for fluctuating income, or for a genuine inability to repay.[19]
The bounds of supervision are therefore necessary and insufficient. Necessary, because without them the market would optimise only profit. Insufficient, because the systemic health of a portfolio can coexist with a quiet unfairness at the entrance.
12. What a justification is for, and where it is not enough
Before the duty to explain, the applicant had to check whether the bank would even tell them no. With a justification they check something more complex: whether the stated factors exist, whether they are theirs, whether they are current and whether they actually carry the verdict to which they are attached. A justification is navigation to the evidence. It is not automatic evidence.
One rejection can combine three true figures into an untrue conclusion. The first describes a postcode, the second self-employment, the third an evening time. Each datum works, but the synthesis punishes the environment, not the ability to repay. Elsewhere the system uses a high-quality register even though a fresh statement is available. Or it cites income that applied two years ago.
A four-step check of a justification
- Existence: the stated datum actually entered the decision.
- Authority: the datum is primary for the given claim, not a proxy for a prohibited characteristic.
- Support: the particular passage actually supports the particular no.
- Scope: the person, the period, the definition of income and the limitations fit.
When explaining a rejection, a table of decisive factors, their sources and their limitations is assembled first. The text for the client should arise only from this check. An untraceable datum should not become a reason merely because the model formulated it persuasively.
A credit system is safer when it shortens the path to a check, not when it replaces it. For a decision on housing, meaningful human review is a suitable organisational rule especially in disputed or high-impact cases; it is not, however, by itself a universal legal command that a human must always issue the last verdict.
13. Privacy is not a switch. It is the entire data flow
The question “is it trained on my data?” is important, but too narrow. The applicant and the institution must also know where forms and statements are stored, for how long, who has access to them, whether audit records are kept, which third parties process scoring and whether a connected tool sends data onward. “We do not use it for training” does not mean “we store nothing” nor “no subcontractor will process anything”.
The differences between a consumer click, a bank identity and Open Banking consent are marked in the legal texts.[8][15] It does not follow that every enterprise regime is suitable for every sensitive datum. It follows that security cannot be judged by a logo without the name of the product, the contract and the configuration.
In the European Union the AI Act framework is added. An ordinary chat for drafting a slogan and a system affecting access to credit are not the same. Concrete classification requires a legal assessment.[9][20]
The practical minimum is a map of the data flow. It begins with the applicant, continues through the form, the registers, the model, the threshold, the logs, the exports and ends at the place where a datum can be corrected, deleted or a verdict returned. Without this map a claim about privacy is only a description of one part of the pipe.
14. The price of default is the price of a raw material, not the price of a verdict
The market uses several price languages at once. The bank measures loss from non-payment. Marketing measures conversion. The customer measures the rate. Compliance measures the fine. A direct comparison of these price lists without knowledge of the working procedure is therefore misleading.
A cheap automated system can be economically excellent for classifying a million short applications for a small limit. The same automated system can be expensive for a task that it repeats three times, consumes an Open Banking trail and in the end requires half an hour of human correction — or, which is worse, no correction, because the rejected person disappears. An expensive manual file can be pointless for a predictable card. It can, however, be cheaper for a one-off decision with a high price of error if it reduces the number of invisible nos.
The true price of an accepted verdict
Model + data + human assignment + verification + the price of default + the price of a false rejection + the impossibility of a complaint + future dependence on one chain.
A verdict is cheap only when it has passed a check and can actually be used.
Costs also include fluctuation. If the same application is decided differently in another channel, the average price of a correct result is not the price of one run. If the system exhausts the human-review limit in the middle of the day, the cost is waiting. If thresholds, telemetry and automated systems are firmly bound to one environment, the cost is also a future exit.
That is why it is useful to measure the price per accepted verdict, not the price per application. Accepted means that the result met predetermined criteria, passed a check and did not cause subsequent harm that portfolio statistics do not see. This metric can reverse an order that looked unambiguous from NPLs.
15. What to choose for which application
A pilot should start with a description of the decision, the data and the control mechanism, not with the brand of a model. The table determines which two regimes to compare on the same data, with the same applicants and with the same budget of time for review.
| Need | First group of candidates | Decisive test | Unacceptable shortcut |
|---|---|---|---|
| A quick small limit | Automation with a clear threshold and justification | Repeatability for identical applications | Swapping speed for fairness |
| Housing and a high amount | Automation with documented meaningful human review of disputed cases | Individualisation of income, current limits and exceptions | Watching only portfolio NPLs |
| The self-employed and a thin trail | A manual file with statements and taxes | Whether fluctuation is not automatically default | Punishing the form of income |
| Fairness of entry | A disparity audit, proxy ablation | The same income, a different address and device | The sentence “we do not use race” |
| Compliance with the GDPR and the AI Act | Documentation, a log, a changeable verdict | Whether a human may say yes | Human-in-the-loop theatre |
| Sensitive movement on an account | Narrow Open Banking, short retention | A map of the flow and withdrawal of consent | Consent as a blank cheque |
For Czech use a separate test is necessary. A foreign study of machine learning on mortgages will not measure whether the system preserves the significance of a trade, correctly reads a tax return, distinguishes a decimal comma, understands parental leave, or does not translate a Czech postcode into a non-existent analogy. The test set should include the self-employed, a thin trail, a register correction and an application in which the correct answer is “the datum is not enough”.
Some institutions ultimately need several regimes. One for a small limit, a second for housing, a third for complaints. That is not a failure of standardisation. It is the same decision as using a different tool for accounting and a different one for graphics. The condition is that the division be intelligible, that data not pass over uncontrolled and that people know where final authority arises.
16. A two-week test that says more than a single AUC
A choice can be made without a six-month tender and without an impulsive purchase. A short, disciplined pilot is enough. Its aim is not to prove that a candidate can do something impressive. The aim is to find situations in which it fails, and to establish whether the organisation can catch them.
Fourteen days from impression to decision
- Days 1–2: Write down 20 to 40 anonymised applications, their frequency, the price of error and the present time.
- Days 3–4: Prepare the same inputs, mandatory points and an evaluation sheet. One version with postcode and device, one without them.
- Days 5–7: Repeat each important application at least three times and keep a register of verdicts, rates and justifications.
- Days 8–10: Verify correction of data, human review, export, logs, deletion of the Open Banking trail and a safe return.
- Days 11–12: Calculate the price of an accepted verdict including the check and the disparity between comparable groups.
- Day 13: Stress tests: a bad register, old income, evening submission, an older device, a prohibited characteristic.
- Day 14: Decide according to the matrix, document the limitations and set a date for the next review.
What needs to be tested
The first examination concerns consistency between channels. The same application from the web, from an app and from a branch is not an identical experiment if telemetry and permissions differ. The second concerns the threshold: the marketing word “instant decision” may mean automation only up to a certain amount. The third concerns data. Open Banking, registers and the form do not measure the same person. They measure three different trails of the same person.
The evaluator should if possible not know which regime created the verdict. They watch substantive correctness, completeness of the justification and the possibility of correction. Every failure is classified: proxy, data, opacity, threshold, tool or interruption.
A pilot should also have conditions for exclusion. For example: a single unannounced penalty of a prohibited characteristic, a non-existent justification, the impossibility of correcting a register or the impossibility of exporting a record. A high average AUC should not outweigh an error whose impact the organisation cannot accept.
The last output is not only the name of a model. It is an operating rule: what scoring is used for, what it is not used for, which data it may see, what a human must approve, how the result is verified and when the decision is opened again. Models, tariffs and conditions change. Without a date of revision even a good pilot gradually turns into a corporate superstition.
17. Open Banking: consent is not a blank cheque
The Payment Services Directive allows a provider to see movement on an account with the client’s consent.[15] That can be fairer than a static statement. It can also be a deeper probe into privacy: rent, a pharmacy, maintenance, betting, gifts. Consent to access to an account is not consent for every item to enter a black box without justification. It is consent to a data flow that must be possible to describe.
Two similar accounts can have a different data regime. An employee with a regular wage looks clean in Open Banking. A self-employed person with the same net year looks like noise. A partner who pays the rent from their own account “does not exist”. A person who cares and has irregular benefits looks riskier than they are. The system sees items. It does not see the household.
Before switching on a fully automated verdict
- Separate the retrieval of a score from the action that concludes or rejects a contract.
- Determine the owner of each decisive datum and the date of its validity.
- Introduce a threshold above which a machine must not decide alone: a high amount, a thin trail, a contradiction in the data.
- Verify that the applicant sees from which source the verdict drew.
- Introduce an exclusion condition: an uncorrectable register, a missing justification, a silent proxy.
The greatest competitive advantage of enterprise scoring may therefore not be the model. It may be order in the data. And the greatest weakness need not be hallucination. It may be a true answer based on an item that was true two years ago.
18. An invisible rejection outside portfolio statistics
The first generation of credit scores sold speed. The next sold accuracy. Today’s products sell something larger: a place in which a person applies, documents, waits and accepts a verdict on a life chance. That is why lenders do not compete only for the best prediction. They compete for the data flow. For Open Banking, for telemetry, for the habit that two seconds are enough.
The more comfortable the ecosystem that arises, the higher the price of exit can be. Thresholds are hard to export, automated systems rely on proprietary features and people stop distinguishing their own process from the product in which they built it.
That does not mean automation is bad. It is often precisely automation that reduces the arbitrariness of fatigue, mood and the prejudice of a particular officer. It means that it should be bought consciously. An institution needs a transferable justification, an audit of disparities, a right to rectification, an emergency procedure with a human and a regular test of whether the original reasons for the model still hold.
Default is visible in indicators of non-performing loans and in impairment allowances. A false rejection does not appear in the portfolio of granted loans. If a model learns predominantly from the results of previously approved customers, a risk of selection bias arises; the particular impact must be shown by an audit of the data and of rejected applications.
The best system is therefore not the one that never errs. We do not have such a system today. Nor is it the system with the highest AUC, the fastest click or the longest list of variables. It is the system whose abilities match the actual work and whose errors are visible, reversible and economically and humanly bearable.
At the moment when the model only ranks files, we are choosing an assistant. At the moment when it reads an account, selects sources, sets a threshold and itself says no, we are also choosing a new structure of accountability. And then the question is no longer: “Is the algorithm more accurate than a human?”
What decides is which error we will see in time, who can review it and who will bear the price of an invisible rejection.
— Jiný Kontext