AI for finance & business

AI is the best research intern you will ever manage: fast, tireless, widely read, fluent in accounting, and completely confident even when it is wrong. It will draft a company analysis in minutes and structure an investment memo better than most juniors. It will also invent revenue figures, mix up fiscal years, and quote a market-share number to one decimal place that it simply made up. The whole discipline of using AI for finance is taking the speed and the structure while trusting it with exactly zero numbers.

Market and industry research: fast landscape, then verify

Start any new company or sector with a landscape pass. Ask ChatGPT, Claude, Gemini, or a search-grounded tool like Perplexity to map the industry — the main players, the value chain, the regulators, the demand drivers, the KPIs analysts watch. Prefer the search-grounded modes, because they cite sources you can open. But treat everything that comes back as a set of hypotheses to check, not as findings; the model is giving you a map of where to look, not the territory. A worked example: ask for the competitive landscape of Kenya's cement industry and you get a fast, mostly-correct structure — Bamburi, East African Portland, Mombasa Cement, National Cement (Simba), Savannah, the clinker-versus-imports dynamic, the economics of a 50kg bag. That structural map is genuinely useful. What you must not trust is the numbers bolted onto it: installed capacity in tonnes, market-share percentages, plant utilisation. Those are stated confidently and are frequently stale or invented. Use the map to know what to pull from primary sources — company annual reports, Capital Markets Authority (CMA) filings, KNBS production data — and get every number there.

Reading financial statements with AI

Modern models with document upload — Claude, ChatGPT, Gemini, NotebookLM — will read a 200-page annual report and orient you fast. Attach the audited income statement, balance sheet, and cash-flow statement and ask the model to walk you through them, flag unusual movements, and compute the ratios: liquidity (current, quick), leverage (net debt/EBITDA, interest cover, debt-to-equity), profitability (gross, operating and net margin, ROE, ROA, ROIC), efficiency (asset turnover, receivable days, inventory days), and the per-share figures. Ask it to interpret each ratio against the company's own history and its sector, not merely compute it. The mechanics of these ratios and what they actually tell you are the substance of LeadAfrik's Accounting Fundamentals, Equities, and DCF Valuation courses — this module is about making AI a faster lens onto them, not a replacement for knowing them.

You are an equity-research analyst. I have attached [Company]'s audited FY20XX annual report.
Using ONLY the audited financial statements and their notes:
1. Give me a 3-year table of revenue, gross profit, EBIT, net profit, and operating cash flow.
2. Compute net debt/EBITDA, interest cover, ROE, and gross margin for each year, and show the formula and the exact line items you used.
3. Flag every line that moved more than 20% year on year and quote the note that explains it.
4. For every figure, cite the statement, the page, and the line item.
If a number is not stated in the financial statements, write 'not disclosed' — do not estimate, infer, or fill it in from memory.

Two rules make this safe. First, recompute by hand every ratio that drives your conclusion — the model transcribes numbers from tables incorrectly more often than you would believe, especially from scanned or multi-column PDFs, and a single misread digit changes the story. Second, never accept a figure the model 'remembers' rather than reads from the document you gave it. If it did not come off the page you attached, with a page and a line you can point to, it does not go into your analysis.

The investment-research memo

An investment-research memo has a standard skeleton, and holding a standard structure while drafting is exactly what AI is good at. Feed it your verified numbers and your view, and let it organise the argument, tighten the prose, and keep the sections consistent. The rule reverses the usual instinct: you supply every fact and figure, the model supplies structure and language — never the other way round.

• Recommendation and thesis — buy, hold, or sell, and the argument in one or two sentences
• The business — what the company does and how it actually makes money
• Thesis drivers — the two or three things that must be true for the thesis to work
• Financials — the trend in revenue, margins, cash flow, and balance-sheet strength
• Valuation — where it trades versus peers and versus its own history, plus a rough intrinsic sketch
• Risks — what would break the thesis, organised by category
• Catalysts and timeline — what moves the price, and when you expect to be proven right or wrong

You are a sell-side equity analyst writing for an institutional audience. Draft a two-page investment-research memo on [Company, NSE ticker] using ONLY the figures I provide below — do not add, estimate, or recall any number that is not in my inputs. If a section needs a figure I have not given you, insert [VERIFY: describe the figure needed] and continue.

Structure the memo exactly as:
1. Recommendation and one-sentence thesis
2. Business overview — how it makes money
3. Thesis drivers — the 2-3 things that must be true
4. Financial summary — trend and quality of earnings
5. Valuation — multiples versus peers and a brief intrinsic view
6. Key risks — operational, financial, regulatory, FX, governance
7. Catalysts and timeline

Write in plain, confident, non-promotional prose. For every claim, make clear whether it is a fact from my inputs or my judgement. End with the single most important thing that would make this thesis wrong.

Verified inputs:
[paste your traced, sourced figures here]

For the valuation section, AI is a useful setup tool and a poor source of truth. Ask it to build a comparable-multiples table — EV/EBITDA, P/E, price-to-book against a peer set (on the NSE, for a bank you might line up Equity Group, KCB, and Co-operative Bank; for a consumer name, EABL against regional brewers) — and to lay out a simple DCF or reverse-DCF skeleton showing the formulae. But every input is yours: the peer multiples come from verified market data, and the discount rate, growth, and margin assumptions come from your judgement and the filings. A valuation the model produces from its own remembered figures is worth nothing, however tidy the table looks.

Forecasting: scenario partner, not oracle

Here the line is sharp. AI is a strong partner for framing forecasts — building bull, base, and bear scenarios, listing the drivers to flex (subscriber growth, ARPU, margin, capex intensity, FX), and pressure-testing whether your assumptions hang together. It becomes dangerous the moment you ask it for a point forecast. 'What will revenue be in FY2027?' returns a confident, precise-looking number with nothing underneath it — the model is pattern-matching to what a forecast sounds like, not building one from drivers. Use it to structure the scenarios; you own the numbers that go into each one, built from verified inputs and stated assumptions.

Risk and strategy as a thinking partner

AI's best financial use may be as a disciplined sceptic. Ask it to argue the bear case against your own thesis, to list what would have to be true for you to be wrong, and to enumerate risks by category — operational, financial, regulatory, FX, governance, customer concentration. For African listings, prompt it explicitly for the ones that bite here: currency depreciation against reporting in shillings, tax and regulatory shifts (a KRA ruling, a new levy, a rate cap), political and policy risk, related-party and governance concerns, and thin trading liquidity. The same partner works for strategy — a market-entry plan, a pricing decision — where it will surface frameworks (five forces, unit economics, willingness-to-pay), checklists, and blind spots. It will not give you the market size or the right price; it gives you the questions, and you go and answer them with real data.