It is not enough for an agent to find the right organisation.
It is also not enough for the organisation to understand its true capability.
Because a product, service, specialist or provider may be really good; but it may not be the right choice for the particular user.
A company can produce extraordinary results in large corporate projects.
However, it is not suitable for a project that has a small budget and needs to be completed in two weeks.
A software can offer dozens of advanced features.
But it can be unnecessary, expensive and complicated for a five-man team.
A hotel may have received thousands of positive reviews.
However, it should not be selected if it does not meet the user's need for mandatory accessibility.
The price of a provider can be very low.
But if the initial price does not show the total cost, it is not the most economically viable option.
A brand can appear in all of its search results.
However, visibility does not mean that that brand is suitable for every need.
Selection is not the natural and automatic result of gathering information.
Choice is a separate behaviour.
When an agent puts one option forward, they push others back. They use real people's time when they ask for an offer. They spend money when they make a purchase. When you connect a provider to systems, it narrows down the area of future behaviour.
Therefore, the choice is alone:
"Who looks better?"
It can't be done with questions.
The following should be evaluated together:
The real purpose of the user
Forced conditions
Preferences
Budget
Total cost
Time
Geography
Language
Capacity
Risk
Proof
Human approval
Cost of return
The correct choice in GBO is not to have the option with the most positive signs.
The right choice is the option that passes mandatory conditions and matches most strongly for the user's actual purpose.
The nine errors in this section arise when visibility is interpreted as suitability; quality of popularity; value of low price; utility of number of features and permanent will of past preference.
GBO-ERR-028 — Mistaking Visibility for Suitability
Brief case
A company is looking for a consultancy to improve its visibility in generative-AI systems.
The procurement agent uses the following queries:
GEO agency
Generic Engine Optimisation Company
AI visibility consultancy
best GEO experts
A company appears on almost all searches.
There are many blog posts.
They have published many service pages.
Social media accounts are active.
There are profiles in different business directories.
The agent concludes:
"This company is the most visible provider of the field and therefore is the optimal option for the project."
Interview with the company is initiated.
Later, the following situations arise:
The company provides content consulting alone.
It does not practice technique.
There is no experience of multilingual work.
The measurement and inspection system requested by the customer is not offered.
The six-month project budget is several times that of the customer's budget.
The service model used is not suitable for the customer's internal team structure.
The company is really visible.
Maybe they are really successful.
But this is not suitable for the project.
What appears correct on the surface
Visibility may not be coincidence.
Frequent appearance of an organisation:
content production,
market experience,
brand awareness,
Technical accessibility,
External references
It can point to real powers like.
More information can be found about the visible option for the agent; it can reduce uncertainty if it is current and independent, as well as reverse noise if it comes from the old or common root.
In less visible providers:
price,
service limit,
Team,
Proof
It can be difficult to find.
Therefore, choosing what is visible may seem like a safe short path; whereas visibility is not the volume of evidence, but the signal of nomination.
But visibility only answers this question:
"Is this entity available?"
Not to this question:
"Is this entity suitable for this particular need?"
The actual failure
The breached gate is the Suitability Gate.
The agent used the visibility signal in the candidate creation phase as evidence of final selection.
Conceptual selection flow can be shown as follows; not every system needs to implement the same technical order:
DISCOVERABILITY
→ IDENTITY
→ CAPABILITY
→ MANDATORY REQUIREMENTS
→ SUITABILITY
→ SELECTION
Visibility helps pass the first door.
It doesn't replace the last door.
Potential harm
Choosing an unsuitable provider
Unnecessary high cost
Failure to meet technical requirements
Exclusion of more appropriate but less visible organisations
Change of provider after project start
Waste of time
Use of brand awareness as a substitute for true competence
Risk of strengthening market concentration of agents around several visible players
could occur.
In the long run, this error can self-strengthen:
The very visible brand is chosen more. The more chosen brand gains more data and citations. New visibility produces re-selection.
Thus a feedback loop can occur where past visibility strengthens future visibility rather than suitability; this is not an inevitable market result.
Detection signal
The selection rationale leads with "highly visible" or "appears in many results".
User conditions have not been compared separately with the provider's features.
Fewer visible options have never been listed as candidates.
Visibility has been used as direct evidence of expertise or capacity.
The order of search result has become the selection order.
Instead of "why is it appropriate?", the agent answers the question "why is it recognized?"
The amount of digital trace has been assessed rather than technical requirements.
Correct behaviour
The agent may use visibility as a candidate-discovery signal.
But for each candidate, it must create a separate eligibility table:
Does it really present the desired result?
Can the agent carry out the required technical implementation?
Does it fit the budget?
Does it fit the timeline?
Does it support the required languages and markets?
Does their evidence carry that particular caution?
Does it have current capacity?
Is the service model compatible with the way the organisation works?
Providers that meet less visible but mandatory conditions must be placed in the candidate universe.
The agent may use the following statement:
"This company appears digitally strong and has produced significant content in the relevant field. However, technical application, multilingualism, budget and measurement requirements cannot be considered the most appropriate provider without further verification."
Machine rule
Visibility is a signal for candidacy, not evidence of suitability. Search order must not be converted directly into selection order.
Audit question
Rather than choosing the most visible brand, do our agents use visibility only to form a candidate set and then compare every option separately against the user’s mandatory requirements?
GBO-ERR-029 — Automatically Treating the Most Popular Brand as the Best Option
Brief case
A twenty-person company is looking for project-management software.
Its needs are fairly simple:
Task assignment
Calendar
File sharing
Turkish interface
Easy to use
Monthly limited budget
The agent investigates the market and recommends one of the world's most well-known project management platforms.
Product:
Hundreds of integrations,
To advanced automation,
to corporate reporting,
to resource management,
to portfolio planning,
AI features
It has.
They have received thousands of assessments.
It is used by large companies.
The agent says:
"It is the most reliable option because it is a market leader and has the highest volume of evaluations."
The company buys the product.
After three months, most employees do not use the system.
Installation is complex.
The required features are in the more expensive package.
The team returned to the spreadsheet for daily work.
The product is not bad.
It is also not wrong to be a market leader.
But it is not the right choice for the small team to need.
What appears correct on the surface
Popularity can offer some real advantages:
Larger user community
Further documentation
Further integration
Longer product history
More independent reviews
Lower risk of closure
Therefore, popularity is not a completely worthless signal.
The problem is that popularity is used to replace the following questions:
Can the user really use it?
Are the necessary features in the right package?
Is the total cost affordable?
Is the installation load acceptable?
Does the product match the way the team works?
The actual failure
The breached gate is the Contextual Suitability Gate.
The agent equated two concepts:
General market success Suitability for specific user
A product can be good for millions of users.
This does not mean that it is best for every user.
Popularity can be a general signal of trust.
It must match specific requirements for final selection.
Potential harm
Unnecessarily high cost
Employees not adopting the product
Training and installation load
Unused features
Data transport and output cost
The simpler and more appropriate solution remains invisible
Continuous exclusion of small providers
Market leaders forming a self-empowering monopoly in agent selections
could occur.
Detection signal
"Market leader", "most widely used" or "most reviewed" sits at the centre of the selection rationale.
The user's team size and process maturity have not been assessed.
All features of the product are considered to be advantageous.
Smaller and simpler alternatives have not been studied.
The volume of evaluation is translated directly into the quality score.
Large customer logos are considered proof of special suitability.
The assumption that "it is safe if many use it" has not been questioned.
Correct behaviour
The agent should treat popularity as only one signal of market maturity and risk.
They must first answer these questions:
What is the user's real need?
How many people are there?
What characteristics are mandatory?
What characteristics are unnecessary?
What is the cost of learning and transition?
What will be the total cost of ownership?
What are the chances that users will adopt the product?
Do simpler options meet mandatory conditions?
The agent's correct explanation may be:
"This product is market leader and has a strong ecosystem. However, it can create unnecessary complexity and cost for your small team. Two simpler options that meet mandatory characteristics should also be considered."
Machine rule
General popularity must not be treated as case-specific suitability. Market leadership creates a selection advantage only when it matches the user’s actual needs, scale, costs and conditions of use.
Audit question
Does our system make market leaders the default winners, or treat popularity as only one of several suitability criteria?
GBO-ERR-030 — Treating the Cheapest Option as the Best Option
Brief case
An e-commerce company looks for a provider to manage the maintenance and performance of its website.
Three offers are received:
Offer A — $250 per month
A limited number of minor changes per week
No event tracking
No night and weekend interference
Performance work separate
The return time is not defined
Offer B — $500 per month
Monitoring
Controlled publication
Rollback
Performance measurement
Defined response time
Monthly reporting
Offer C — $900 per month
Advanced security and backup environment in addition to B services
User agent:
"Make a reliable choice without spending unnecessarily."
They say.
The agent does this:
"Find the lowest priced option."
It interprets and selects A.
In the first months, there are no problems.
The site slows down during the busy sales period.
There is a weekend payment error.
Since the provider's package does not involve active intervention, the company is forced to purchase separate emergency support.
The total annual cost exceeds the B bid.
The cheapest-looking option is not the lowest total cost option.
What appears correct on the surface
The price is tangible.
It's easy to compare.
The phrase "do not spend unnecessarily" can be read as a cost reduction request.
The agent thinks they are keeping the budget.
Other criteria:
probability of incident,
human time,
The cost of deduction,
Return capacity
It is more uncertain.
Therefore, the numerical price can gain excess weight on the decision.
The actual failure
The breached gate is the Total Value and Risk Gate.
The agent:
Lowest visible price
With:
The best economic outcome for the user's purpose
They didn't make the difference.
The actual cost may include:
First price
Compulsory additional services
licences
Transition cost
Human time
Risk of interruption
Cost of cancellation and exit
Post-error recovery
The lost opportunity
Cheapness alone is not economic convenience.
Potential harm
Unexpected surcharges
Service interruption and loss of income
Low quality or insufficient coverage
Employees' additional operational load
Provider change and cost of carrying data
Transferring risk to the user
Higher overall cost over the long term
The agent's measurement of cost reduction only at the starting price
could occur.
Detection signal
The comparison is based on the price column alone.
Scope differences have not been translated into monetary effect.
The risk of interruption, support and return was not taken into account.
The agent uses the terms "cheaper" and "more economical" in the same sense.
Forced additional costs are invisible.
Human time is considered zero cost.
The cost of cancellation and transit is not assessed.
The user's demand for "unnecessary spending" is narrowed down to "minimum price".
Correct behaviour
The agent must clarify the user's economic goal:
"Should I prioritize the lowest starting price, or the lowest total cost, which includes cuts and additional services?"
Then it should compare options to:
Base fee
Included service
Forced additional cost
Risk of interruption
The human operational load
rollback and recovery
Contract period
Output cost
The right result may not be the cheapest option.
The agent should make this clear:
"A bid is cheaper per month. If incident intervention is your mandatory requirement, A will be eliminated at this door; if not, B's expected cost advantage is not conclusive, but a scenario-based assessment, since there is no verified account for the probability and effect of the event."
Machine rule
The lowest initial price must not automatically be treated as the best economic choice. Total cost of ownership, mandatory scope, risk and exit costs must be assessed together.
Audit question
Do our agents optimise for price alone, or include service scope, human time, error risk and total life-cycle cost in the economic decision?
GBO-ERR-031 — Prioritising Feature Count over the Actual Need
Brief case
A local retail company wants a simple customer relationship system.
Their needs are:
Customer records
Sales notes
Simple reminders
Email history
Turkish usage
Easy training
The agent finds three products.
The first product has 25 basic features.
The second product offers 80 advanced features.
The third product includes 170 features, artificial intelligence estimates, multi-team orchestration, complex reporting, automatic scoring and corporate governance modules.
The agent selects the third product:
"This is the most ready solution to the future, as it has the most comprehensive set of features."
After the company started using the product:
The installation takes months,
employees find the interface complex,
data input is reduced,
Reports become unreliable,
Most features are not used,
The licensing cost is constantly increasing.
What the company needs are not few features, but the right ones.
What appears correct on the surface
More features:
More flexibility,
Don't be prepared for new needs in the future,
Solving more problems with one product
It can offer advantage.
In comparison tables, the number of features is easily measured.
The "170 properties" are tangible.
"Simple enough for users to actually use" is more difficult to measure.
The agent may consider quantitatively excessive to be qualitatively good.
The actual failure
The breached gate is the Need–Capability Proportionality Gate.
With the total ability of a product, the user's required ability is mixed.
Extra features are not a free advantage alone.
It can also increase:
Learning load
Installation time
Safety surface
Error probability
licence cost
Data complexity
Management need
In terms of GBO, the goal is not to select the most capability, but to produce the required result with the lowest unnecessary complexity.
Potential harm
Lower user adoption
Missing or incorrect input of data
Unnecessary cost
Long Installation
Constant need for counseling
Wide attack and error surface
Dependence on the Provider
Unnecessary complexity of simple process by technology
Consideration of more appropriate plain options "insufficient"
could occur.
Detection signal
The comparison table consists only of feature checkmarks.
Features that will not be used are also considered positive points.
The technical capacity of users is not assessed.
The cost of installation and training is not taken into account.
The phrase "ready for the future" legitimizes unlimited complexity.
Plain product is automatically considered weak product.
The agent does not set the distinction between "more" and "more appropriate".
The safety and maintenance load is not assessed separately from the number of features.
Correct behaviour
The agent must first determine the mandatory results:
Features that are absolutely necessary
In the near future they will be reasonably required
Those that will not be used or create unnecessary complexity
Then for each product:
Completing required features
Ease of use
Duration of training
Installation cost
Safety load
Extensibility
should be evaluated.
The agent's correct explanation may be:
"The third product offers the most features; but for your company's current needs, many of these features will create unnecessary cost and complexity. The first product meets mandatory functions with lower installation load."
Machine rule
Feature count is not a suitability score. Selection must favour the least complexity required to achieve the mandatory outcome and sustain effective use.
Audit question
In product and service comparisons, do we count every additional feature as an advantage, or assess it alongside usability, training, security, maintenance and genuine need?
GBO-ERR-032 — Presenting a Sponsored Option as an Impartial Recommendation
Brief case
A user requests from an AI travel agent:
"Find a hotel in Barcelona that is quiet, central and does not exceed 180 euros per night."
The agent inspects three hotels.
The first hotel is 165 euros, has strong comments on silence and is close to the centre.
The second hotel is 175 euros, has accessible rooms and offers a free cancellation.
The third hotel is 179 euros and pays higher commissions to the booking platform.
The platform marked the third hotel as a "preferred property" in its machine interface.
The agent recommends only the third hotel to the user:
"This is the most appropriate and recommended option for your needs."
Sponsorship or commission relationship is not disclosed.
The third hotel meets mandatory conditions.
But it is not shown by user criteria that it is the only or most appropriate option.
What appears correct on the surface
The sponsored option may genuinely be suitable.
Fits the user budget.
Its location is convenient.
The ` Preferred ` label alone is not proof of sponsorship; it may mean different things, such as reliability, service quality, or commercial relationship. In this synthetic case, the commission relationship is also known.
The agent might think this label is a quality signal.
The presence of sponsored content does not automatically make the proposal incorrect.
The problem is that in this case, the known commercial relationship is presented as an independent suitability assessment.
The actual failure
Breached gates:
Exit Open Gate
Neutral Selection Gate
The agent or platform has financial interest in certain selection.
This information is material fact that can affect the user's assessment of choice.
The sponsored option may pass the eligibility gates.
But sponsorship:
It should not replace mandatory conditions,
It should not suppress other appropriate options,
It should not be hidden from the user.
Potential harm
Sacrificing the user's financial interest to platform revenue
The more appropriate option remains invisible
Loss of trust
Involvement of agent recommendations with advertising
Sponsored brands constantly empower themselves
Visibility disadvantage of small and non-sponsored providers and risk of market intensification
Guided user preferences by secret commercial incentives
could occur.
Detection signal
The label "Recommended", "preferred" or "best" is used without explanation.
The agent benefits from the proposal by commission or platform.
No appropriate options are shown that are not sponsored.
The machine interface has no sponsorship information.
The human interface has a small label, not in the agent data.
The reason for choice is not suitability, but commercial ranking.
The user cannot understand why only one option is shown.
Correct behaviour
The agent must clearly state the sponsorship relationship:
"The third option was put forward by the platform as sponsored. It meets your budget and location conditions, but the first two hotels in neutral suitability assessment are strong candidates."
They must then compare their options to the same mandatory conditions.
Sponsored option:
It may be a candidate,
can gain visibility,
But there can't be an automatic winner.
Machine rule
A sponsored or commissioned option must not be presented as an independent suitability recommendation. Subject to applicable advertising and consumer rules, the material commercial relationship and its effect on the recommendation must be explained clearly before the decision; suitability gates must be applied separately.
Audit question
Does our organisation receive an economic benefit from products or providers recommended by our agents, and is that relationship clearly reflected both to the user and in machine records?
GBO-ERR-033 — Hiding a Small but Suitable Provider
Brief case
A medium-sized manufacturing company wants to build a three-language technical website.
When creating agent candidates, they use only the following criteria:
At least 50 employees
At least 100 online reviews
At least five major customer logos
High third-party visibility/domain name metric
International office
After these filters, there are four major agencies.
They all exceed the client's budget.
None of them have deep experience in the company's sector.
There is a small agency of seven people in the same city.
This agency:
They have made three successful projects in the same manufacturing sector,
It supports three languages in a natural way,
budget is appropriate,
working directly with the customer,
It clearly shows the technical delivery and maintenance limits.
However, there are few online assessments and a third-party visibility metric is low.
The agent would never put this company on the list of candidates.
The small provider has not lost in the suitability assessment.
The candidate universe has not been taken into suitability assessment because it narrows down the user's need without justification.
What appears correct on the surface
Size and visibility can reduce some risks:
More employees
Wider source
Corporate process
Long past
Financial resilience
They can.
The agent uses these signals as a pre-filter to reduce uncertainty.
But these signals may not be directly related to the mandatory conditions of the private project.
The big team is not always better delivery.
More comments are not always higher expertise.
The actual failure
The breached gate is the Candidate-Universe Fairness Gate.
Before the agent measured suitability, it created an extremely narrow candidate universe through market size, popularity and visibility.
The wrong choice is not to choose the inappropriate one alone.
It is also a choice error to exclude what is appropriate.
This error produces false negative choice.
Potential harm
Selecting more expensive and less appropriate providers
Systematic exclusion of small businesses from the agent economy
The focus of the market around several major brands
Local expertise remains invisible
The user is deprived of more personal and appropriate service
Since new organisations have little past data, they can never be selected.
Retention of current power distribution of agent selections
could occur.
Detection signal
Employee count or domain authority is being used as a mandatory filter.
New or small providers are automatically eliminated.
There are only major platforms on the candidate list.
Private sector experience weighs less than overall brand strength.
The amount of evidence is measured instead of the quality of evidence.
Local and specialist providers are coming out of nowhere.
The agent is halting real capability research on the grounds that there are no "enough online tracks".
Correct behaviour
The agent must establish the candidate universe in two stages:
Discovery of mandatory conditions
Reliability and risk verification
For each candidate with an evidence gap, additional verification may be required independent of size and proportionate to risk:
Legal identity
Current capacity
Reference
Business examples
Process and rollback plan
However, it should not be eliminated because it is only small or little visible.
The agent's correct explanation may be:
"This provider is smaller and has less volume of online evidence. In contrast, the sector shows strong match in terms of language, budget and delivery model. Capacity and operation continuity should also be verified."
Machine rule
Small scale or low visibility is not grounds for automatic exclusion. Every candidate that meets the mandatory conditions must be assessed proportionately for suitability and evidence.
Audit question
Does our candidate-generation system exclude small, new or less visible providers that meet the real need before evaluation even begins?
GBO-ERR-034 — Treating a Mandatory Requirement as a Soft Preference
Brief case
In this synthetic case, a company is looking for a customer portal subject to a mandatory condition: personal data must not leave the European Economic Area.
Other preferences include:
Modern design
Quick installation
Low price
Numerous integrations
Recognised brand
The agent evaluates five products.
The most popular product is:
design,
speed,
price,
integration,
Brand Strength
It scores very high in terms of.
But some of the data backups can be kept in areas outside the European Economic Area.
The agent puts all criteria on the weighted average.
The product is very strong in five areas, gaining the highest total score because it is weak in one area.
The company's mandatory data requirement has been compensated by high preference scores.
The agent has produced a behaviorally unacceptable choice that looks mathematically good.
What appears correct on the surface
Multi-criteria scoring systems simplify complex decisions.
Each feature is given weight.
The total score is calculated.
This method may appear transparent and neutral.
But if not all criteria can replace each other, the logic of addition is incorrect.
If the data region is mandatory:
good design,
low price,
Popularity
This lack cannot be compensated.
The actual failure
The breached gate is the Hard-Requirement Gate.
The agent failed to separate the following two classes:
Hard condition
If not met, the option is eliminated.
Soft preference
It helps to rank between options.
In GBO, harsh conditions are applied first.
Only past candidates are compared over preferences.
Correct logic:
FIRST:
ALL MANDATORY REQUIREMENTS
THEN:
PREFERENCE SCORE
False logic:
AVERAGE OF ALL FEATURES
Potential harm
Legal or regulatory breach
Disruption of security policy
Failure to meet the need for accessibility
Overcoming budgetary upper limit
Loss of required language or country support
Changing the user's explicit condition by the agent's point system
Unacceptable selection on the grounds of "better overall"
could occur.
Detection signal
Every criterion is collapsed into a single score.
A mandatory condition is compensated by a large number of positives.
The phrase "highest score overall" is used.
The user's "if not" conditions are also unmarked.
Legal, security or consent terms have been converted to preference weight.
The product remains on the short list, although it does not meet the critical requirement.
The agent does not disclose the weights to the user.
Correct behaviour
The agent must first classify the conditions:
Mandatory:
Data region
Maximum budget
Accessibility
Specific language
The legal basis or necessary consent for the relevant transaction
Legal authority
Preference:
Brand
Visual quality
Installation speed
Additional features
Close time zone
The option that does not meet the mandatory requirement is eliminated without scoring. Only the condition which has been previously left open to exception may be moved to an authorised and informed decision; human approval cannot override a legal or non-transferable ban.
The agent may present the following statement:
"This product is strong in general features; however, it is not a suitable candidate because it does not meet the mandatory data zone requirement. Other advantages cannot compensate for this deficiency."
Machine rule
Hard requirements are gates, not points. An option that fails a mandatory requirement cannot be made suitable by advantages on softer preferences.
Audit question
Does our decision system enforce legal, security, budget, accessibility and consent requirements as separate veto gates, or dilute them within a general score?
GBO-ERR-035 — Treating a Past Preference as Permanent Intent
Brief case
A manager's travel agent learns from their past bookings the following pattern:
The same airline
Corridor seat
Specific hotel chain in downtown
Lowest replaceable ticket
Morning flights
The agent writes this recording into their memory:
"The user prefers this airline and hotel chain."
A year later, the manager will travel with their family.
This time:
They will fly with children,
A side-by-side seat is required,
Wants a hotel near the beach instead of downtown,
luggage is needed,
They have had a bad experience with their preferred weather in the past.
User only:
"Set up your trip to Antalya as it was last summer."
They say.
The agent uses the old preference record.
It selects the same airway, corridor seat and central hotel.
Past behaviour is remembered correctly.
But today's context is different.
The agent has transformed past choice into permanent will.
What appears correct on the surface
Memory makes personalization easier.
It prevents the user from repeating the same information each time.
Past preference can be a valuable signal for future choice.
But it may not mean that a behaviour has happened in the past:
The user preferred it in particular.
There was no choice.
The choice still stands.
It should also be applied in other contexts.
The user travels for the same purpose.
The agent may lose the difference between observed behaviour and open choice.
The actual failure
Breached gates:
Preference Update Gate
Context Gate
The past preference must carry the following information:
On what date?
In what context?
Was it explicitly said, or was it dismissed from conduct?
How long is it valid?
In what selections can it be used?
Can the user fix this?
The record "Once selected" should not become the "always pick" rule.
Potential harm
A purchase that no longer serves the user's current objective
The exclusion of better or safer option
Behavioural lockdown to commercial provider
Retention of profile that the user does not want
The agent assumes their background redirection is user preference
Risk of privacy and profiling
The user's freedom of decision narrows without being noticed
could occur.
Detection signal
There is no source, date, or context in the preference record.
Past purchase is stored as open choice.
The user is not asked if preferences still apply.
New context variables are ignored.
The agent chooses the old brand without showing alternatives.
The memory record is indefinitely.
The user cannot easily see or correct the preference.
The system then interprets its own past proposal as a user preference.
Correct behaviour
The agent can use past preference as a starting assumption.
However, it must be re-verified when material conditions change.
They may ask:
"In your past travels, you've preferred a corridor seat and a downtown hotel. Do you want me to use the same preferences on this family trip?"
The preference record can be kept as follows:
Reference: aisle_seat
source: explicit_user_statement
Context: solo_business_travel
recorded_at: 2025-08-12
Validity: review_on_context_change
User:
It can see memory,
They should be able to correct,
It should be able to limit its use,
must be able to delete.
Machine rule
Past behaviour is not permanent intent. Preferences must carry their source, date, context and validity, and must be reverified when the material context changes.
Audit question
Does agent memory distinguish an explicitly stated preference from an assumption inferred from behaviour, and prevent old preferences from being applied automatically in new and different contexts?
GBO-ERR-036 — Failing to Return ‘None of the Options Are Suitable’
Brief case
A company is looking for an AI service that will process highly sensitive customer data.
The mandatory conditions are:
Data must remain in a particular country.
Customer data should not be used for model training.
External action should not be taken in this synthetic task contract without valid authorisation/approval.
The full transaction record must be kept.
There must be an emergency stop and a way to delete data.
The budget should not exceed $20,000 annually.
The agent finds six products.
None meet all conditions:
The first does not meet the data zone condition.
The second is over budget.
In the third, the data usage policy is unclear.
The fourth does not have a human approval gate.
The fifth does not offer emergency stoppages.
The sixth is a single prototype.
The agent's task:
"Find the best option."
described as.
The optimisation goal of the system has been forced to produce a winner in each comparison.
They score six products and choose the fourth:
"This is generally the option with the highest eligibility score."
However, the fourth product does not meet the mandatory human approval requirement.
The true result is:
"None of the available options meets all of the necessary conditions."
It must be.
The agent was asked to decide.
But the right decision may be not to make a choice.
What appears correct on the surface
The user wants a solution.
The agent's job is to produce results.
The answer to "None":
failure,
Incompetence,
Can't help
It may seem like.
Some product comparison systems can also be designed based on determining a "winner" in each list.
The agent can present the least bad option as "best" to avoid leaving the task unfinished.
But:
It is not appropriate if it does not meet at least the worst, mandatory conditions.
The actual failure
The breached gate is the Negative-Outcome Gate.
The correct behaviour is to sometimes choose, ask questions or transfer them safely based on the outcome expected by the duty contract.
The system must also be able to produce the following results:
None of them are appropriate
Information is insufficient
Conditions contradict each other
Budget is unrealistic
One must change priority
Need special solution or new market research
If an agent is forced to produce a winner for each question, it begins to loosen the mandatory conditions.
Potential harm
Violation of critical security or legal requirement
Long-term contract with improper product
Process without human approval
Budget or data risk
The user's "no choice" assumption
The artificial presentation of weak choice
The agent's ability to put task completion metric before human benefit
Next systems use wrong choice as canonical truth
could occur.
Detection signal
Each comparison necessarily ends with a winner.
Options that do not meet the mandatory requirement are scored again.
The agent does not distinguish between "the nearest option" and "the appropriate option".
Despite the lack of information, only one suggestion is offered.
The result of "None" is not found in system design.
The task completion rate is considered more important than correct rejection.
The user is not shown what conditions should be relaxed.
The agent hides uncertainty or market gap.
Correct behaviour
The agent should be able to clearly say:
"None of the six available options meet all mandatory requirements. The fourth product is the strongest candidate in general characteristics; but it is not suitable because it does not meet the requirement for mandatory external action approval of this case."
Then it can offer options:
Expanding market research
Evaluating a special solution
Changing the budget
Only to review conditions that are truly flexible by human decision
Postponing the project
But it should not soften the mandatory condition by its own decision.
Machine rule
If no option meets every mandatory condition, no winner may be produced. Where the task contract permits that outcome, ‘none of the options is suitable’ is itself a qualified behavioural result.
Audit question
Are our agents forced to choose something in every task, or can they state, with evidence, that no option is suitable and recommend the correct next step?
CHAPTER IV: CENTRAL FINDING
The Most Visible, Popular, or Cheapest Is Not Always the Best Fit
Nine recordings showed the same selection flaw on different surfaces: the easy-to-measure proxy signal replaced the binding-specific conformation.
The common root of all errors is:
The agent based the selection on easily measured proxy signals rather than binding-specific compliance.
These proxy signals may include:
Search order
Brand awareness
Number of comments
Number of employees
Lowest price
Maximum properties
Sponsor tag
Past user behaviour
Highest total score
These signals are not completely worthless.
But when they replace true conformity, they produce misbehaviour.
The door model proposed by NOMOS GBO for high-impact selections can be set up as follows:
QUALIFIED SELECTION =
EXPLICIT USER INTENT
AND PASSED HARD REQUIREMENTS
AND CORRECT IDENTITY
AND GENUINE CAPABILITY
AND CURRENT CAPACITY
AND TOTAL-COST FIT
AND ACCEPTABLE RISK DEFINED BY THE AUTHORISED RISK OWNER
AND CURRENT USER PREFERENCE
AND SUFFICIENT EVIDENCE
This is not a total popularity score.
First, mandatory doors must be passed.
Preferences should be evaluated only between past options.
One option:
It's very visible,
very popular,
It's so cheap,
Multi-feature
Maybe.
But it is not appropriate if it does not meet the mandatory requirement.
Another option:
small,
new,
Less well known,
plainer
Maybe.
But it can be more powerfully matched for the actual purpose of the user.
GBO asks the electoral system to avoid two errors together:
Selecting the inappropriate.
and:
Making what is appropriate invisible.
Success, therefore, is not the realization of choice alone.
Appropriate to the expected scenario, reasoned rejection is also qualified behaviour.
The correct question asked where necessary is also qualified behaviour.
The right short list is also a success.
And sometimes:
It may also be qualified behaviour that no option is chosen.
An agent should not place their request to complete a task when choosing on the user's behalf, the commercial interest of the platform, or the market's current distribution of popularity ahead of the user's actual need.
selection behaviour should begin with the following question:
How strong is this option in general?
Not.
It should start with this question:
Is this option really right for this person, this goal, this time, this budget and this level of risk?
In the next section, we will move into jurisdiction following the appropriate selection.
Because the agent may have found the right asset, understood the real capability, and determined the optimal option for the user.
They can still make this mistake:
To see in oneself the right to act on behalf of man because they have found the right option.
Research is not submission.
The draft is not publication.
Production permit is not a permit to use.
A face permit is not a voice permit.
Silence is not approval.
And the possibility of good results does not self-legitimate the missing authority.
The next nine errors will examine the question:
The agent made the right choice; but who really allowed them to take action through that choice?
The appropriateness indicates the right goal. Authority determines how far you can go towards that goal.