Regulatory disclosure plain-language summary
LLM converts terms, FSCS, and MiFID disclosures into plain language calibrated to the customer's estimated financial literacy level.
Banking processAcquireOnboardOpenFundTransactServiceReviewClose
By Don, DoneThat’s AI coach · updated
The summary sits beside the terms, it does not replace them
A plain-language summary of regulatory disclosure is a reading aid generated from the bank's approved legal text. It is not a second contract, not a waiver, and not a substitute for the terms, the FSCS information sheet, or the MiFID costs and charges document the customer must still receive.
Legal owns the template. Product, digital, and any model that drafts wording work from that signed artefact. The customer sees the summary and the controlling document together. A tick that only acknowledges the summary is not acceptance of the terms.
If the statute does not give a protection, the summary must not imply one. If a fee or exclusion is in the signed terms, the summary must not drop it because the wording sounded harsh.
The opening journey already pulls identity data and routes the applicant through KYC. That is a different control from disclosure. form pre-population from ID extraction and agentic KYC orchestration can fill fields and sequence checks. They do not author the legal meaning of FSCS cover, MiFID costs, or account terms.
Generate from the signed instrument, then cite the clause
Start with the instrument legal has already signed for that product version: the terms and conditions, the FSCS information sheet for eligible deposits, and the MiFID information pack for investment business. Do not start from a product page, a chatbot script, or a sales deck.
- Bind the run to a product code, a document version, and a jurisdiction.
- Extract only the clauses in scope for that screen: eligibility, cover or the absence of cover, fees, cancellation, risks, and costs where MiFID requires them.
- Write a short summary in the approved register of phrases, then attach a citation to the controlling clause (document ID, version, clause or section number).
- Render the full document beside the summary, with a control that opens the signed text, not a paraphrase.
- Refuse to emit a summary if the citation cannot be resolved, if the product has no mapping for that disclosure type, or if legal has not signed the template for that version.
A current account and a stocks and shares ISA do not share an FSCS story. An illustrative walk-through, not a measured result: a UK retail applicant opens a standard current account and is also offered an optional stocks and shares ISA. For the current account, the model summarises the FSCS information sheet legal attached to that product, citing the sheet version and the paragraph that states whether deposits with the firm are eligible. The full sheet stays on the same step. For the ISA, the model does not reuse that deposit wording. It summarises the MiFID-required costs, charges, and risk information for that ISA share class, citing those sections, and shows the full key information and costs documents next to the summary. If cash in the ISA awaiting investment sits in a different protection category from the current-account balance, the summary must say so by citing the relevant paragraph, not by borrowing the current-account sentence.
Core banking and CRM stacks (Temenos, Salesforce, and the same class of platforms) can store product codes, document versions, and the screen the customer saw. They are systems of record for the journey, not sources of legal meaning. The model reads the approved text those systems point to. It does not invent a limit, a scheme, or a cooling-off right that the pointed-to document does not contain.
Literacy calibration changes the wording, not the bargain
Financial literacy calibration is a presentation choice. The bank may estimate whether the applicant needs shorter sentences, defined terms, or a glossary. That estimate must not select a different fee table, a different FSCS statement, or a softer exclusion.
Two applicants for the same product version are bound by the same signed terms. One summary may say you pay this fee when you use an unpaid-item facility. Another may say an unpaid-item fee applies if a payment is refused for lack of funds. Both must cite the same clause. Neither may omit the fee or add a cap the clause does not contain.
Treat literacy level as a rendering parameter on a locked template, similar to font size or language. If the model cannot express the clause in the simpler register without dropping a condition, it must keep the denser wording or fail the generation.
Do not let conversational lead qualification bleed into disclosure. A chat that scored the applicant as inexperienced is not a licence to hide complexity. Qualification decides whether the product is offered. Disclosure explains the product that was offered.
Failures that look helpful and fail legal review
These failure modes show up on first legal review. They usually appear when the model tries to be kind. Skipping the full terms because the summary "covers it" is also a fail: the full terms, FSCS sheet, and MiFID pack remain mandatory. The summary cites them. It does not retire them.
Promising FSCS cover the product does not have
The model has seen FSCS language on deposit products and applies it to an investment, an insurance wrapper, or an e-money balance that the attached information sheet does not treat as an eligible deposit, or treats only in part. The summary then reads as if the scheme stands behind market loss or behind a product the sheet excludes. FSCS sentences may only be generated from the FSCS artefact mapped to that product version. If that artefact says cover does not apply, the summary must say that, with the citation, and must still show the sheet.
Dropping a fee because it sounded harsh
Arranged overdraft interest, unpaid-item fees, foreign-transaction charges, and ISA platform or fund charges are easy for a model to soften or skip when the brief is plain language. The customer then consents to a summary that is quieter than the tariff. Every fee the screen is responsible for must appear in the summary or in a labelled list that still cites the tariff. Harshness is not a reason to omit.
Treating literacy calibration as a different contract
A simpler register is mistaken for a simpler deal: you can get your money back, without the actual cancellation window; protected, without the scheme's conditions; no charges, because the model summarised only the headline rate. The customer is shown two texts that do not match. Put the signed document beside the summary, and put legal's wording that the signed document prevails on the screen. The model does not invent that wording.
Template ownership and the opening journey
Legal signs the template: allowed phrases, citation format, literacy registers, the on-screen hierarchy of summary, citation, and full text, and the statement that the signed document prevails. Product may request a new product mapping. Digital may implement the layout. Neither ships a wording change without a new signed version.
Change control follows the document, not the model. When terms, the FSCS sheet, or MiFID costs change, withdraw the old template, bind the new version to the product code, and regenerate only after legal signs. Keep an audit record of document version, template version, citation set, literacy register used, and the identifier of the full text shown.
Place the step after product selection and before the applicant is asked to accept. If cross-sell propensity at account opening adds a second product on the same application, run disclosure per product. Do not merge two FSCS or MiFID stories into one paragraph.
The quality test is narrow. A reviewer can match every sentence in the summary to a cited clause, open that clause without leaving the step, and confirm the summary does not grant a protection the statute and the signed document do not give. If that test fails, the screen does not ship.
Is this worth automating for you?
Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other.
DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.
Measure the baseline first