Now accepting projects — Q3 2026

Financial Services: Custom AI Solution

Financial services AI should not start as a chatbot project. The valuable work is to turn documents, service history, product rules, and internal knowledge into searchable, traceable, evaluated workflows inside a compliance boundary.

// AI use cases to start with

Policy and product RAG

Chunk, embed, permission, and search SOPs, product manuals, compliance policies, FAQs, and service knowledge so employees can ask without crossing access boundaries.

Document and contract extraction

Extract fields from applications, policies, contracts, financial statements, or KYC follow-up documents and route them into review workflows.

Service / RM assistant

Connect call summaries, client context, next-step reminders, risk prompts, and ticket creation while keeping human confirmation in place.

// How EKel would deliver it

01Define compliance red lines: what cannot enter the model, what needs source citations, and what requires human review.
02Create a reference dataset and eval rubric for hallucination, access leakage, and wrong advice.
03Ship one KPI-bound sprint first, such as summary time, KYC handling time, or internal-policy lookup time.
04Operate with logging, sampling, alerts, and version control after launch.

// Best fit

→Financial teams with heavy documents, policies, product information, or service history that is slow to search and summarize.
→Financial organisations that want to convert manual unstructured work (documents, lookups, summaries) into measurable AI workflows.
→Teams that want to prove ROI in a small sprint before buying a large AI platform.

// Sources

Salesforce Agentforce for Industries Salesforce Financial Services Cloud

Back to industries See financial Salesforce →

// Custom AI architecture

AI is not a tool — it is a four-layer stack.

// LAYER L4

User layer

Bankers, service agents, customers — interacting via web, mobile, internal tools, Slack, or email. Every AI surface must keep a human-review escape hatch.

Web · MobileInternal toolsSlack · Email

// LAYER L3

Application layer

Built with **Vibe Coding** — AI embedded into existing business flows (KYC, contract review, service triage, reporting). Multi-step tasks (e.g., “KYC follow-up → AI extraction → risk rules → officer review”) run as **agentic workflows**, with the AI agent chaining multiple tools and data sources. One flow, one KPI — not another dashboard.

Vibe CodingAgentic WorkflowAuth · RBAC

// LAYER L2

AI layer

LLM Gateway (multi-model routing) + RAG retrieval + eval pipeline + guardrails + hallucination monitoring. Models are swappable; guardrails are not.

LLM GatewayRAG · EmbeddingsEval · Guardrails

// LAYER L1

Data layer

Vector DB (policies / docs embedded) + structured database + document store + full audit log. Data classification decides what may leave the data centre.

Vector DBDocument storeAudit log

// Finance AI compliance checklist

Every box ticked — that is what “ready to ship” means.

Data classification

Classify what cannot enter any model, what can go to commercial LLMs, and what stays on private deployment. Without this classification, do not start.

Deployment choice

Private deployment / compliance LLM gateway / on-prem inference — chosen by data classification, not by what is easiest.

Retention & audit trail

Every prompt, retrieval context, and LLM response is retained with timestamp, user, model version. Auditors can pull what they need.

Hallucination monitoring

Continuous sampling after launch (auto-eval + human review). Alert thresholds for high-stakes flows. Full regression test on every model upgrade.

Human-in-the-loop

AI prepares the proposal; humans decide. KYC outcomes, contract clauses, lending recommendations are all human-signed. AI removes the repetitive work — it does not remove judgement responsibility.

Eval & reference dataset

Before launch, run 100+ real finance scenarios (compliance, leakage, wrong advice). Re-run on every model version change. No pass, no ship.

// FAQ

Five questions clients ask most.

01Data cannot leave on-prem. Can AI still work?

Yes. On-prem open-source models (Llama, Mistral, Gemma) + self-hosted vector DB + local inference is fully viable. Trade-off: model quality lags frontier by 1–2 generations — usually fine for structured document extraction and policy-lookup RAG, noticeably weaker for natural service dialogue. The decision is driven by data classification: what cannot leave the data centre, what can go through a compliance gateway, what can reach a commercial LLM. Without that classification, do not start.

02After a model upgrade, do existing evals still hold?

No. Model upgrades are exactly the moment behaviour can change — the same prompt may yield different output. The healthy practice is to wire the eval pipeline into CI/CD, re-run the full reference dataset on every model version (including major bumps like GPT-4 → GPT-5). Hallucination rate, accuracy, and latency must clear thresholds before cutover; if not, roll back. Eval-dataset maintenance cost is consistently underestimated, but it is the line between production AI and demo AI.

03RAG or fine-tuning — which?

Pick RAG for 90% of cases. Reasons: (1) policies change — RAG just updates the document, fine-tuning requires retraining; (2) RAG cites sources, useful at audit; (3) RAG is 1–2 orders of magnitude cheaper. Fine-tuning fits style consistency (e.g., service reply format) or domain vocab (e.g., institution-specific acronyms), but is rarely non-negotiable in finance. A common production stack is RAG primary + a small fine-tuned model for query reformulation or reranking.

04Why do many AI projects take six months to ship?

Usually three reasons: (1) data, permissions, and integration points were not mapped before kickoff; (2) scope too wide (“build an AI platform” instead of “ship one flow”); (3) the eval rubric was added after launch instead of being defined upfront. Invert all three and a vertical flow ships in 4–6 weeks — the precondition is one flow / one KPI, not a platform.

05Why not just use off-the-shelf AI service / document AI products?

If an off-the-shelf SaaS fits, buy it — that is the engineer’s call, not a consulting line. Custom is appropriate when one of three holds: (1) compliance constraints rule out the SaaS (data cannot leave the country, cannot be multi-tenant, etc.); (2) the flow needs deep integration beyond webhooks; (3) AI behaviour itself is the competitive edge (“our service voice is distinctly ours”). The first two are common in finance; the third is rare.

Financial AI has to be controllable before it scales.

We can start with one process, one sprint, and one real eval set to measure time saved and risk introduced in your financial workflow.

Book Free Consult→

We use cookies

We use strictly necessary cookies to run this site, plus optional analytics cookies (Google Analytics) to understand how visitors use it. See our Cookie Policy and Privacy Policy.