Ask-Your-Portfolio
Will let you ask the loan book a question in your own language and get a figure you can trace.
In most lending institutions every portfolio question becomes a request to the one person who knows how to build the export, and that person becomes the bottleneck for the board, the oversight body, the auditor and the finance team at the same time. Monthly briefings are rebuilt by hand from the same five reports each cycle, with commentary retyped and variances recalculated in a spreadsheet outside the system. The cost is not just the hours: it is that questions which would be useful simply do not get asked, because asking is expensive.
The pipeline, stage by stage
Each stage names what goes in and what comes out, because a capability you cannot inspect is a capability you cannot audit.
- 01
Question intake and scoping
A plain-language question would be asked in English or an Indian language. It would be bound immediately to the asker's role and data scope, so a question asked under a regional permission runs against that zone, and a borrower's question runs against its own loans, using the same scoping that governs the built views.
- 02
Mapping to a named query plan
The question will be mapped to a structured plan over a reviewed set of measures and dimensions: outstanding, disbursed, repaid, accrued interest, penal, overdue position, by scheme, zone, entity, financial year, product. The plan would be drawn from an approved catalogue. No free-form query will be generated against the production database.
- 03
Deterministic execution
The plan is designed to run against the same reporting layer that produces the five built reports, so a figure in an answer is the same figure in the export. There will be no second calculation path that can drift from the first.
- 04
Answer composition with citation
The answer will state the figure, the query plan used, every filter applied, the as-of timestamp, and a link to the record-level list behind it. A number without its working will not be shipped as an answer.
- 05
Briefing drafting
On demand or on a schedule, a briefing would be assembled from the lender's template: portfolio movement for the period, largest variances with the loans driving them, overdue position by bucket, scheme-wise and zone-wise position, with drafted commentary naming specific loans rather than describing trends abstractly.
- 06
Review and issue
The draft would go to a named human who edits and issues it. Nothing will be circulated to a board, an oversight body or a borrower automatically.
What it proposes, and what you approve
This is the architectural rule of the whole AI layer, made specific to this capability. Nothing a model produces reaches a balance on its own.
| Ask-Your-Portfolio would propose | A named human would approve |
|---|---|
| An interpretation of the question, echoed back before answering | The interpretation, where the question was ambiguous |
| A query plan drawn from the approved measure catalogue | Additions to that catalogue, which are a configuration change |
| A figure computed by the deterministic reporting layer | No approval step: the figure would be the report figure, read-only and identical to the export |
| Drafted commentary explaining a movement | Every word of that commentary before it reaches a board pack |
| A complete briefing document | Issue of the briefing, by a named officer |
Shaped to your book, not shipped as one fixed version
These are the decisions you make and we build against. They are the reason two deployments of the same capability do not look alike.
- The measure and dimension catalogue
Which measures your institution recognises and what each one means precisely: whether "outstanding" includes accrued-but-unpaid interest, whether "overdue" counts from due date or grace expiry. Definitions are yours and would be stated on every answer.
- Languages
Which languages officers and board members actually ask in, including transliterated input, and which language briefings are issued in.
- Role scope and masking
Which measures each role may query, and whether any field is masked at role level, using the zone and entity isolation already running in production.
- Briefing template and cadence
Section order, which comparisons appear (month on month, year on year, against sanction), the variance threshold above which something earns commentary, and the schedule.
- Answer policy
Whether a figure may be quoted without its record-level export attached, and whether answers carry a watermark or an as-of stamp when used outside the system.
- Question log retention
How long questions, plans and answers are kept, which matters because the log is itself an audit artefact showing who asked what and when.
Inputs
- Your briefing template and the commentary conventions your institution follows
- the measure catalogue and the precise definition of each measure, agreed with you during configuration
Outputs
| Artefact | Format |
|---|---|
| Answer card: figure, filters, query plan, as-of timestamp | In-app |
| Record-level list behind every answer | Excel and PDF |
| Drafted board or oversight briefing | DOCX and PDF |
| Saved question library, per role | In-app |
| Question and answer log with plan and timestamp | Excel, for audit |
| Scheduled briefing pack with attachments | PDF bundle |
The duties this creates
Duties, not job titles. Each one below is a permission you grant, so you map them onto the roles you already have. Preparing, approving and configuring are separate by design, which is how you stop one person holding two of them.
- Asking segment questions
Whoever is accountable for a slice of the book asks about it and receives a standing briefing. This is the highest-value use of the capability.
- Asking operational questions
Day-to-day lookups by anyone working the portfolio, answered within what their permission already lets them see.
- Cross-checking an answer
Reconciles what the assistant said against the underlying export before a figure is used externally.
- Defining scope
Sets which data each permission level can reach. The assistant inherits that boundary rather than having one of its own.
- Reviewing after the fact
Reads the question log, which records who asked what and what they were shown.
Where the lender enables it, the borrower can ask questions about its own loans and nothing else.
What it will do when the data fights back
Any vendor can describe the happy path. These are the cases that decide whether a capability is safe to put near a loan balance. Designed behaviour in each case:
The question cannot be mapped to an approved measure
It would say so plainly, state what it can answer, and offer the nearest supported question. It will not produce an approximate answer to an adjacent question and present it as the answer to the one asked.
The period, scheme or entity is ambiguous
One clarifying question, then the answer. It will not silently apply a default financial year and let the asker discover the assumption three slides into a board pack.
The asker's scope excludes part of the true answer
The scoped figure would be returned with the scope stated on the answer card in the same weight as the number. A zone figure must never be readable as a portfolio figure.
A question is asked while an accrual run is in progress
The answer would carry the as-of timestamp and a note that accrual is running for a stated month, so a figure taken mid-run is not mistaken for a closed position.
A measure definition changed part-way through the year
The same question then returns different figures for the two halves, so the answer card would split the period at the change date and show a figure for each definition with the date it took effect. It would state plainly that a single combined figure is not comparable, and it will not apply the current definition backwards over a period that definition did not govern.
Why it sits where it does
A second-wave build. It has no data prerequisite of its own: the reporting layer, the five reports and the role scoping all run in production today, which is why it is the capability with the shortest path from question to answer. It is nevertheless sequenced after the first-wave correctness work, on purpose. Making a balance easy to ask for before that balance is continuously proven would mean distributing an unverified number faster, which is a worse outcome than making people wait for the export.
What has to exist first
- Reproducible outstanding computation
- the Anomaly & Reconciliation Guard, so a quoted figure sits behind a live integrity state
- an agreed measure catalogue per lender. Already in place: the five built reports and their agreed definitions, zone and entity data scoping
What buyers actually ask about this
The figures would be computed inside your deployment by the reporting layer, so the loan data stays there, and a language component would need only the question text and the schema of the measure catalogue. Where that component runs, including entirely inside your boundary, is fixed in the deployment agreement, and the security page sets out the detail.
Because the number will not be generated by a model. The question would be mapped to a query plan from an approved catalogue, and the plan would execute against the same reporting layer as your Excel exports. Every answer will show the plan, the filters and the as-of time, and link to the record list. If the answer and the export disagree, that is a defect, not an interpretation.
The officer who issues the briefing, which is why the design refuses to circulate anything automatically. Drafts are drafts, and they would carry their working so the reviewer can check it in seconds rather than rebuilding it. The commitment we make is that the figure in a draft will be the figure in the report, traceable to records.
Yes, and most institutions will. Access would follow whatever roles you have configured and the existing zone and borrower scoping, and the measure catalogue itself would be filterable per role. A lender will be able to let a regional permission query its zone freely while restricting a borrower to its own loans, or switch the capability off for a role entirely.
No. The question log would stay in your deployment for the retention period you set, and it exists mainly as an audit record of who asked what. Any tuning of the measure catalogue or phrasing would happen inside your own deployment, as part of the co-build. The same isolation applies here, and the security page sets out the detail.
The design target is a genuine answer in the asked language, with the interpreted question echoed back first so a misreading is caught early. Measure names and figures would carry their canonical form alongside the translation, so a board member reading in Gujarati and an auditor reading in English are provably looking at the same measure.
Other capabilities
Migration Copilot
Will read your legacy loan book, rebuild every timeline and reconcile every balance.
See the designAnomaly & Reconciliation Guard
Will prove continuously that every balance reproduces from the event log, and flag what does not.
See the designDocument Intelligence
Will read sanction orders and resolutions into draft records, and draft your letters back.
See the designShape this one around your book
Bring your document formats, your languages and your thresholds. We will show you what this looks like against your own data.
45 minutes | On the live deployment | A straight answer on sequence