Blog/AI & Innovation
AI & Innovation

ChatGPT for Financial Services: Designing Banker Workflows With Morgan Stanley & Evercore

i
iFynx StudioSep 19, 2026 · 8 min read
chatgpt financial services morgan stanley evercore

Junior banker work is becoming a design brief

On 10 September 2026, OpenAI launched ChatGPT for Financial Services, a GPT-6 Astra-powered ChatGPT Work experience developed with Morgan Stanley and Evercore as design partners (OpenAI, CNBC, Reuters, Fortune). The product targets investment banking and equity research: company research, financial modeling, comps, LBO analysis, and pitchbook drafting — with built-in data from partners such as Daloopa, PitchBook, LSEG News, and Crunchbase, plus granular citations and subscription integrations.

What product teams should steal

1. Design partners who live in the workflow. Naming Morgan Stanley and Evercore is more than PR. It signals that prompts, templates, and review loops were shaped by people who ship pitchbooks under deadlines. Vertical AI without practitioner partners becomes a demo.

2. Citations as UX, not compliance theater. Granular source links turn model output into audit-ready draft. Regulated MENA banks and investment firms should require the same pattern before any “AI analyst” feature reaches production.

3. Bring-your-own data subscriptions. Integrating customers’ existing PitchBook/LSEG seats reduces procurement friction and clarifies who owns the license. Your fintech agents should model the same: agent layer + licensed data plane.

4. Scope to junior-task leverage, not partner replacement. Positioning around research assembly and modeling scaffolding is politically and ethically clearer than “replace bankers.” Ship escalation paths to human sign-off.

iFynx takeaway

Vertical ChatGPT experiences win when workflow, citations, and licensed data are co-designed. For GCC capital markets builders, copy the partnership model — then localize disclosure and Sharia/governance review into the same canvas.

Related reading


Banker-workflow acceptance criteria for financial-services AI

ChatGPT-style assistants in wealth and investment banks only ship when the workflow respects supervision, citations, and disclosure — not when the demo impresses a steering committee.

  • Every answer that touches a client recommendation cites the source document version and retrieval time
  • Human approve / edit / reject is mandatory before anything lands in a client-facing channel
  • Prompts and tool calls that touch PII or account numbers are redacted in logs by default
  • Model outputs that include forward-looking statements trigger a disclosure template the bank already approved
  • Arabic and English banker UIs expose the same supervision controls — no English-only compliance gate
  • Session export produces an audit pack: prompt, tools, citations, human edits, final send

If a junior banker can paste a model answer into email without a gate, you do not have a financial-services product.


Failure modes Morgan Stanley–grade buyers will ask about

  1. Stale research retrieved as current; client gets last quarter’s thesis dressed as today
  2. Cross-client context leak because the retrieval index was not tenant-isolated
  3. Over-confident summaries omit risk factors that were in the source PDF footnote
  4. Supervisors rubber-stamp because the UI dumps walls of text with no diff against sources
  5. Weekend model updates change tone; compliance has no pinned eval suite for top intents

Product acceptance should include a “reconstruct the advice” drill from the audit pack alone.


Ops playbook: pilot without theater

  • Week 1: pick three intents only (e.g. summarize deck, compare filings, draft internal brief) — freeze scope
  • Week 2: wire citation + human gate; disable free-chat in the pilot cohort
  • Week 3: run bilingual evals on golden questions with compliance in the room
  • Week 4: publish an internal one-pager — what the assistant may and may not draft for clients
  • Success metric: supervisor edit rate falling while policy exceptions stay at zero

Enterprise logos are not a substitute for a supervised workflow. Ship the gate first.


MENA banking constraints on the same canvas

Regional desks often need Sharia / governance review and Arabic client communication on the same thread.

  • Allow a parallel review lane for Sharia or local compliance without forking the whole product
  • Store operator rationales in both languages when the client relationship is Arabic-primary
  • Keep data residency switches visible to the admin — default to in-region retrieval indexes
  • Do not invent product claims about specific bank deployments beyond what those institutions publish

Design the canvas so localization and governance are modes, not separate apps.

What banker workflows actually look like

Investment banking and equity research work is a chain of tasks, each with its own sources and quality bar. Designing AI for it starts with mapping that chain:

  1. Research: gather filings, news, transcripts and market data on a company and its peers.
  2. Modelling: build or update financial models, comparables tables and valuation analyses.
  3. Synthesis: turn numbers into a narrative: investment thesis, risks, scenarios.
  4. Production: create pitchbooks, memos and reports in the firm’s templates.
  5. Review: senior bankers and compliance check accuracy, consistency and disclosures.

AI can speed up every step, but review never disappears. The product’s job is to make review faster and more reliable, not to skip it.

Citations are the core feature

In financial work, a number without a source is a liability. Every figure an AI assistant puts into a model or a slide should link to its source document and location, and the interface should make checking that link one click away. When sources conflict, the assistant should show both rather than choose silently. Teams building similar tools should treat citation coverage, the share of generated figures with a verifiable source, as a primary quality metric.

Integrating with firm data and templates

Banks run on their own templates, models and data subscriptions. Useful AI tools connect to those: they fill the firm’s model template rather than inventing a new layout, use licensed data feeds the firm already pays for, and respect information barriers between teams. Access control matters especially in banking, where deal teams must not see each other’s confidential material.

Lessons for Gulf banks and investment firms

Banks and investment firms in Saudi Arabia and the UAE can apply the same approach with local needs in mind: Arabic and English sources side by side, regional market data, Sharia-compliant structures where relevant, and data residency requirements. Under Saudi Arabia’s PDPL, enforceable since 14 September 2024, and the UAE’s federal Personal Data Protection Law (Federal Decree-Law No. 45 of 2021), any personal data in deal documents must be handled carefully, and firms should know where AI processing takes place.

Designing review into the product

  • Highlight AI-generated content until a person has reviewed it.
  • Show changes between versions of a model or deck clearly.
  • Let reviewers comment on and reject specific figures, feeding corrections back into the workflow.
  • Keep an audit trail of who approved what, for compliance and regulators.

Data security and confidentiality

Banks handle material non-public information, client data and deal documents under strict confidentiality rules. Any AI deployment must answer basic questions before going live: where prompts and documents are processed and stored, whether the provider uses them for training, how long they are retained, and how access is logged. Enterprise agreements, regional hosting and single sign-on with role-based access are common requirements. Information barriers between advisory, research and trading teams must be enforced in the AI layer as well, so that a model cannot surface one team’s confidential material to another.

Measuring value in financial services AI

Banks should measure results in terms that matter to the business: hours saved per analyst on research and production tasks, turnaround time for client materials, error rates caught in review, and adoption across teams. Quality measures matter as much as speed. Track how often reviewers correct AI-generated figures and why, and use those findings to improve sources, prompts and workflows.

Common mistakes

  • Allowing AI-generated numbers into client materials without source links.
  • Rolling out a general chatbot instead of tools built for specific workflows.
  • Ignoring information barriers in shared AI workspaces.
  • Measuring usage instead of time saved and quality.
  • Skipping training on how to review AI output critically.

A practical starting checklist

  • Pick one workflow, such as first-draft company profiles, with a clear owner.
  • Connect only approved, licensed data sources and the firm’s templates.
  • Require a source link for every figure and show it next to the number.
  • Run a pilot with a small deal or research team and compare time and error rates with the previous process.
  • Review access controls and information barriers with compliance before wider rollout.
  • Train staff to check AI output, and collect their corrections to improve the setup.

Expanding step by step, with measured results at each stage, is slower than a firm-wide launch but far more likely to deliver lasting value.

How iFynx can help

iFynx is a product and engineering team based in Gaza, Palestine that works with clients across the Gulf. We help companies in Saudi Arabia and the UAE design, build and test digital products like the ones discussed here, with Arabic and English interfaces from day one, careful attention to local regulation and data protection, and a focus on measurable outcomes. If you are planning a similar product, we are happy to discuss scope, risks and a realistic first release.

Frequently asked questions

Can AI build financial models?

It can draft and update models from source data, but professionals must review assumptions, formulas and outputs before use.

Why are citations so important in AI for finance?

Because every figure may be relied on in decisions or client materials; reviewers need to verify it quickly.

Will AI replace junior bankers?

It changes their work, shifting time from manual assembly to checking, analysis and client-facing tasks.

How should regional banks start?

With a focused workflow such as company research summaries, connected to approved data sources and reviewed by experienced staff.

Related: AI agents for banks · ruya × Magure production agentic AI. Building AI tools for financial teams? See AI development for Saudi Arabia and AI development for the UAE.