Compliance said no to ChatGPT — but KYC queues and credit memos keep growing.
Applications arrive as piles of PDFs that analysts read page by page.
Research, client reports and policy questions over documents that must not leave the firm.
Regulators and partners ask where data and models run — the answer must be “with us”.
Wall Street closed the door on public AI tools in 2023 — while onboarding and compliance work kept getting slower.
JPMorgan Chase restricted staff use of ChatGPT; Bank of America called it prohibited for business use; Citi, Goldman Sachs, Deutsche Bank and Wells Fargo reportedly did the same.
for a corporate KYC review on average in 2023, up from 84 days the year before; each review costs about $2,598.
of banks say they lost clients because onboarding was slow or inefficient.
average cost of a data breach in financial services — the second most expensive industry.
of organisations have banned public GenAI apps at least temporarily over privacy and data security.
Put your monthly volume into the calculator and see hours and dollars.
The path of one onboarding or credit file. The model runs on your infrastructure, behind your SSO, and every decision stays with an analyst.
A new onboarding or credit case in your CRM or case tool starts a run with all attached documents.
IntakeOCR reads scans and PDFs; tables from financial statements are kept as tables.
IntakeA local model sorts IDs, registry extracts, statements and contracts into the case structure.
IntakeOwners, directors, addresses, turnover, debt and covenants are pulled out with the page they came from.
AnalyseThe file is compared with your checklist: missing documents and expired IDs are listed.
AnalyseResults from your sanctions and PEP screening provider are attached — the screening itself stays with that provider.
AnalyseMismatched owners, unusual structures or adverse hits stop the run and go to compliance.
AnalyseA case summary or credit memo in your template, every statement linked to its source page.
DraftKey ratios are calculated from the extracted statements and checked against policy limits.
DraftAnalysts ask about credit policy or procedures and get the paragraph and version it comes from.
DraftThe analyst reads, edits and approves the memo; approval rights follow your delegated authority.
DecideInputs, model version, outputs and approvals are logged for model risk and audit.
DecideA local AI for finance is a language model that runs inside the institution’s own infrastructure — on premises or in a private cloud tenancy — and works on documents that cannot go to a public AI service: KYC files, financial statements, credit applications, policies and procedures. It reads, extracts, checks and drafts, with a source link for every statement and a log of every run.
Banks were among the first to restrict public chatbots in 2023. At the same time, onboarding and review work kept slowing down. A private model lets compliance say yes to AI without saying yes to a new third party.
Files arrive read, sorted and summarised. Analysts spend their time on judgement and on the cases that need it, while every decision and approval stays with them.
Typical minutes of analyst work per corporate KYC file. Your own numbers go into the calculator below.
Based on 3 h 10 min by hand and 45 min with local AI per item, from the table above.
The model reads, extracts and drafts. Risk decisions, approvals and client communication stay with people.
Approval follows your delegated authority; the model never approves.
Screening stays with your provider; hits are judged by compliance.
Validation, monitoring and change control follow your model risk policy.
Requests and decisions to clients are sent by people.
Red flags always go to a person with the full source trail.
Flip a switch to hand a step to the model or take it back.
Analysts open files that are already read, sorted and summarised, with sources for every line — on infrastructure the bank controls and can show to its regulator.
We look at the documents and the work, agree what the model may and may not do, and size the hardware.
The model on a server in your network, on your real documents, with a small group of users.
Access rules, logging and integrations, then the whole team — with the old way kept as a fallback.
Model updates in maintenance windows, monitoring and a monthly report of use and time saved.
Regulators ask about data location, third-party risk, explainability and control. A model on your own infrastructure answers the first two directly; source links, logs and human approval answer the rest. We document it for your model risk and outsourcing reviews.
No. It prepares the file and drafts the summary; approvals follow your delegated authority and compliance judges every red flag.
Yes. Models, index and interface run without internet access; updates arrive as signed packages you install.
For a team of 20–100 analysts, one or two servers with 48–80 GB GPUs run a 70B model. We size on file volume and can start in your private cloud tenancy.
The model answers only from the case documents and policy texts it retrieved, and links every statement to its page. If something is not in the file, it says so.
You do. Weights, index and logs live in your environment; we keep no copies and no telemetry.
Describe your use case in two sentences — we reply within a day with a hardware and model recommendation.
hello@wireclad.com →