The analyst workflow in 2026 has changed. AI tools are now table stakes for drafting, summarising, exploring data, debugging code, and reviewing each other's work. Used well, they multiply throughput. Used poorly, they degrade quality faster than you can repair it.
The daily-use stack
- A chat assistant (Claude, ChatGPT, Gemini) for drafting, summarising, brainstorming
- An AI-native IDE (Cursor) or in-editor assistant (Copilot) for code
- An AI search (Perplexity, Phind, You.com) for synthesised answers with citations
- A document-grounded tool — NotebookLM, Claude Projects, or your own RAG pipeline — for working with private documents
Drafting with AI
Use it for the first 60% — outline, structure, draft. Edit ruthlessly for the last 40%. The voice should be yours. The structure can be the model's. The facts must always be verified — every number, every citation, every quote.
Coding with AI
Cursor and Copilot have changed engineering productivity materially. The catch: it's now easier than ever to ship plausible-but-wrong code. The skill is reading AI-generated code with the same scepticism you'd apply to a stranger's PR — does it compile, does it match the codebase's idioms, does it handle the edge cases, does it leak resources?
Building a prompt library
Most of your AI use is repeated tasks. Save the prompts that work. Examples:
- Board-meeting summary: 'Summarise these notes into a 200-word board-meeting recap, structured as: outcomes, decisions, risks flagged, next steps. Use second-person plural ('We decided...').'
- Market update: 'Read this CBK MPC statement and extract: rate decision, vote split if disclosed, inflation guidance, growth guidance, risks. Format as a markdown table.'
- Code review: 'Review this PR. Flag: bugs, security issues, performance regressions, code-style departures from the rest of this file. Don't praise — only flag.'
What to delegate, what to keep
The senior-analyst test
Keep: the question definition, the judgment calls, the verification, the final voice. Delegate: the first draft, the formatting, the literature scan, the routine code, the rewording. The hard skill is knowing the boundary — and it shifts every six months as models improve.
The Brynjolfsson finding
The 2023 NBER paper by Brynjolfsson, Li and Raymond showed that AI assistance lifted the productivity of customer-service agents by 14% on average, with the largest gains for the least experienced (35%). The takeaway for any team: AI assistance compresses the gap between novice and expert. The expert's edge — judgment — is what doesn't compress.
Exercise
You manage a 4-person equity-research team covering East African listed stocks. You are deciding which parts of the research workflow to support with AI tooling. For each of the following tasks, decide whether to delegate to AI (with what verification), keep with the human analyst, or use AI as a first-pass assist. Justify each decision: (1) Reading a 200-page company prospectus and extracting key financial metrics. (2) Building the company's three-statement financial model in Excel. (3) Writing the buy/sell investment recommendation paragraph. (4) Drafting the equity-research note's commentary sections. (5) Translating the investment-note summary into Swahili for retail distribution.