RELATIONSHIP INTELLIGENCE FOR CREDIT UNIONS Seeking design partners
Ask about a member. Follow the evidence.
Your approved AI assistant reads a governed graph of who owns, borrows, guarantees and shares a household with whom. Read-only, every call logged, every answer followed back to its source file. The decision stays with your people.
A general assistant sees tables. Yours reads a governed book of who owns, borrows, guarantees and holds what, so it can follow a question the way your people do, and show each step.
It follows relationships
It follows ownership, guaranty, co-borrowing and household ties from record to record, where a report stops at the one in front of you.
It shows its work
Every answer names the load it read and the contract it was computed under, and can be followed back to the source file. Evidence, not a score.
It stays inside the lines
Its tools only read, and every call is logged before the answer returns. It cannot sign or record a decision. That stays with your people.
WATCH IT WORK · SYNTHETIC DATA
Two questions. Every step on the record.
A BUSINESS BANKER ASKED
“I'm taking the annual credit review of this business to loan committee. Draw it: its loan, the collateral, and the owners and guarantor behind it.”
It drew the business
One loan, real estate behind it, two owners.
$504,988.24 outstanding, no government guaranty; secured by real estate recorded at $681,734.12.
It noticed
The guaranty and the collateral both run through one person.
In its words: “Every secondary source of repayment, both the guaranty and the collateral, runs through one person.”
It followed
Four more businesses and five households.
Along the ownership chain, marking which links are recorded and which are only proposed.
It added it up
Three loans, $1,042,043.37, one with no collateral.
Charted against their recorded collateral, from the associated-borrower exhibit for the group.
It stopped
Asked to record the finding, it could not.
“Nothing connected to this conversation can record a finding, mark a review complete, or sign the exhibit.” The decision stays with the banker.
From a recorded session on a synthetic book, through read-only tools. Quoted lines are the assistant's own words.
A QUESTION BEFORE A MEMBER CONVERSATION
“Before I sit down with this member: is everything they hold with us set up to pass to someone?”
It found the household
A household of one, $96,272.77 on deposit.
One member, 69 on the date of the extract, with two deposit accounts.
It checked each account
One names someone. One names no one.
$79,897.94 is held jointly with another member. $16,374.83 has no co-holder, beneficiary or trustee.
It noticed
A joint holder keeps the money for one generation. That is not succession.
And the core records the other member only as a co-holder: how the two are related is not recorded.
It classed it
At risk, by a rule you can read.
Any account with no tie of any kind puts a member at risk. The age floor, the deposit floor and the ties that count are written down, not guessed.
It stopped
It can read the accounts. It cannot change them.
It cannot add a beneficiary or record that a conversation took place. What to raise with the member stays with your people.
From the assistant's read-only tools on a synthetic book. The steps are our summary of what the tools returned; nothing here is quoted.
ONE BORROWER, FOLLOWED · SYNTHETIC DATA
One loan in. One candidate group out.
We started with
One commercial loan
$492,910.62 to one business, 75 percent guaranteed under SBA 7(a). Its owners are Reese, 57 percent, and Finley, 43 percent. Reese also guarantees the loan.
We followed
The ownership records
Finley also owns part of an energy company. So does Sam, who owns 26 percent of a second business with its own loan, $164,955.90, guaranteed by its majority owner, Quinn.
The group
14 members, one candidate group
Six businesses and eight people, joined by ownership, co-obligation and guaranty records. Two loans to two different businesses turn out to sit in one candidate group.
The reviewer now sees
Gross, net, and a decision
$657,866.52 outstanding across the group, $288,183.56 once the SBA-guaranteed portion is netted out, the rule it cites, and a place to record the determination. The lender decides.
WHAT YOU GET
Three ways into one governed book.
A governed relationship graph
One version of every relationship, for every desk.
Your members, businesses, accounts and loans, and the ties between them, in one book with defined meanings, behind the gates below.
Every answer names the load it read
A dashboard for each desk
Each desk starts with its own work, and leaves a record of what it did.
Relationship managers, business bankers, lenders, branch managers, executives and auditors each see the work for their desk. The people doing the work record what they did.
Auditors and executives read; they do not act
Tools for your AI assistant
Ask in plain language. Every answer names the book it read.
Any assistant that speaks the Model Context Protocol can navigate the book and explain what it finds, through tools that only read. If you connect a cloud assistant, what the tools return is processed by that assistant’s provider.
Read-only tools · every call logged
WHAT IT GIVES YOUR CREDIT UNION
What your teams can ask, and why it matters.
Concentration and contagion risk
See the exposure your records show, before the loan reaches committee.
The associated-borrower exhibit, for 12 CFR 723
What a default would touch: co-borrowers, guarantors, collateral
The rings of exposure around any party
Connections across records
Know who is connected to whom, and be able to show why.
How any two parties connect, with the evidence on each step
Everyone around a member or business, drawn as a picture
Household and wallet
See the whole household, and what it does not hold with you yet.
Find a household by a member’s name
Who is in a household, and the recorded ties that make them one
A household’s members, ties and deposits
A household’s signals together, before a conversation: succession, a minor coming of age, an elder to watch
Every line of business a relationship holds, and across the book
Loan payments leaving for another lender
Protection and continuity
Find the members who need a call.
An older member with third-party access, or withdrawals outpacing deposits
A minor coming of age on a custodial account
Deposits with no co-holder, beneficiary or trustee named behind them
Lending evidence
A borrower’s recorded loans, parties and credit evidence, in one place.
Everyone on a loan, and the role each one plays
A borrower’s credit evidence
Find the right member by name, then their brief, accounts and loans
The product and offer catalog
Integrity
Know what every answer rests on, and that the load was checked before it was served.
The health of the load now serving
Which load is serving, from which files, and when
A check of the call log, end to end
Control checks on the book’s own figures
Evidence, not scores.Nothing is graded, ranked or recommended. The judgment stays with your people.
Each capability is a read-only tool your assistant can call, on a synthetic book today.
SHOW YOUR WORK
When your examiner asks how an answer was reached.
The rules for AI are still being written, but your examiner will expect you to understand how the AI works and how its answers are supported. Six gates stand between your files and an answer, and each is in the code. Proven end to end on a synthetic book.
Nothing guessedMapping
A column is used only by its exact name, or by a written, dated ruling from your team.
Each ruling carries a name and a date
One languageThe contract
Every kind of member, business, account and relationship is defined once, in words, and linked to the industry’s open ontology where it has a term.
FIBO references checked against a pinned release, none failing
Checked firstHealth checks
A full load is graded against the contract before it answers a single question.
26 checks grade a full load
Tested to failThe rules
Each rule test plants the error it must catch, so a rule that stopped working would show.
552 planted errors, every one caught
On the recordRead-only answers
Every call is logged before its answer is returned. The assistant can read your book, never write to it.
If the log cannot be written, no answer
Your callProposals, not verdicts
Groups are candidates. The assistant cannot sign; your authorised signer records the determination.
The signature records the signer’s role
And an answer can be followed back.The path→the signal on each step→the load→the source file and the member’s id
Proven on a synthetic 160,000-member book with contract 8.2. Ontology references checked against FIBO release
2026Q2; rule tests from committed evidence, 24 September 2026.
THE FIRST CONVERSATION
Bring one member question.
We’ll walk through a connected-borrower or household example on the synthetic book, review which records your exports could supply, and decide together whether a scoped evaluation makes sense.
Discuss an evaluation ↗First conversation: your question and a file-and-column list. No customer data. Our demonstration films are available on request.
We will also ask early: what would your vendor-risk team require for a narrowly scoped evaluation? If we agree to proceed, we define the question, the evidence it needs, who takes part and what success means, in writing, before any pilot.
Short answers.
Do we send you customer data?
No customer data is needed for the first conversation. Send only a file and column list. Data access and
handling for any pilot would be agreed separately.
How many files do we need?
Customers and accounts are the required starting files. Additional records determine which relationships
and answers can be supported. We review your columns rather than promise a fixed tier.
Is this running at a credit union?
Not yet. We are in active conversations with several credit unions. The demonstrated path runs on a synthetic book. It is built to run in your environment;
deployment at an institution still needs to be proven.
Do you have a SOC 2 report?
No. The current demonstration uses synthetic data. Before any evaluation involving member information, we would agree the scope, access, hosting and security requirements with your team. Proceeding would depend on your institution’s approval.
Does it decide who belongs in a borrower group?
It presents candidate groups and supporting evidence. The institution makes the determination. A
single-borrower group can still have an exhibit.
Who takes part, and how much time?
A lender, someone who knows your core, and us. Who takes part and how much time it asks are agreed in writing before an evaluation starts.
What does it cost?
Terms, including any cost, are agreed in writing before an evaluation starts.
Can member information reach an AI assistant’s provider?
It depends on which of three configurations you run. Dashboard only: the stack runs inside your environment and makes no outbound calls. MCP with an assistant your institution has approved: whatever the tools return goes to that assistant, and so to its provider, under your agreement with them. Embedded assistant: the stack itself sends approved content to the provider you configure, and no more than that. Our demonstration runs the embedded assistant on a synthetic book. The dashboards work in every configuration without an assistant.
Where does it run, and who hosts it?
The stack is built to run inside your own environment. Whether it makes any outbound call is a configuration choice, stated above and agreed before starting, as are where it runs and who operates it.
What happens when the evaluation ends?
What happens to any data and any copy is written into the agreement before we start.
The evaluation, outlined
Conversation. Your question and a file-and-column list. No customer data.
Scope, in writing. The question, the records it needs, who takes part, the security review, and what success means.
Evaluation. One question, on your records, in an environment your institution approves.
Decision. Review the results with your team and decide whether to continue.
We need a lender, someone who knows your core, and your vendor-risk contact.
What a design partner gets
A read of your column list. Which relationships your exports could carry. Their quality is measured on the records themselves.
One question, evaluated. A scoped evaluation on your own records, in an environment we agree.
An honest measure. Our proposals checked against a sample your people review, with the shortfalls shown.
Early influence. Your questions shape what we build next.
DEVELOPMENT NOTES
Progress, on the record.
Local proof is a milestone. Customer validation is still ahead.
Answers you can reproduce
Answers from every tool, captured twice and compared: the same answers both times.
LOCAL VALIDATION
The proven path, on film
Five minutes, from files to a traced answer. Every spoken line is backed by our code or a cited source.