For Investment Bankers

How Investment Bankers Can Use AI Without Losing Deal Judgement

AI can build a three-statement model in minutes, but it cannot tell you whether the buyer's synergy thesis will actually materialise or whether management is lying about working capital. When junior bankers outsource the foundational modelling work to Copilot, they never develop the feel for when a number is wrong. The risk is not that AI will replace you. The risk is that it will replace your edge before you realise it is gone.

These are suggestions. Your situation will differ. Use what is useful.

Download printable PDF

Own the Assumptions in Your Financial Models

When you generate a DCF or LBO model using Capital IQ AI or ChatGPT, every assumption is a choice someone made. AI will pick inputs that are mathematically sensible but economically hollow. A model that assumes 3 per cent terminal growth because that is the historical average may be correct for a mature industrial business and ruinous for a SaaS company losing market share. Before you send a model to a client or put it in a deck, you must be able to articulate why each assumption exists and what would break the deal if it was wrong.

›Spend fifteen minutes challenging the WACC, growth rates, and margin assumptions before you let a model leave your desk, even if AI generated it
›Force a junior banker to explain what happens to the IRR if revenue growth drops two percentage points or EBITDA margin compresses by fifty basis points
›Keep a one page sheet of your assumptions for the specific subsector you are working in so you can spot when AI has used a generic comparable instead of your market knowledge

Use AI for Speed on Due Diligence Mechanics, Not Judgment

Capital IQ AI and Bloomberg AI are excellent at finding financial statements, pulling transaction multiples, and spotting obvious red flags in regulatory filings. They fail at the work that matters: understanding whether the management team is incentivised to hide something, whether the customer concentration is a hidden risk that the CFO downplayed, or whether the working capital position is sustainable or a sign of deteriorating operations. Let AI do the data gathering. You do the interrogation. A sell-side management presentation drafted by ChatGPT will be polished and internally consistent. It will also have no original insight into what actually makes this business work or what could break it in a downturn.

›Use AI to compile regulatory filings and historical financials, then spend your time in the data asking the questions AI cannot ask
›When AI flags a potential risk, do not assume the risk is real until you have traced it back to the actual financial statement line and understood the operational driver
›Before you sign off on a due diligence report, make sure you have a hypothesis about the one thing that could destroy deal value that has nothing to do with the spreadsheet

Preserve Your Deal Instinct by Doing the Work AI Could Do

A banker who has manually built fifty LBO models and stress tested each one develops an intuition for when a deal is marginal. You feel it before you prove it. When Kensho or ChatGPT builds the model, your brain skips the work that would have built that instinct. You become dependent on the tool to tell you whether the deal is investable. The junior bankers on your team will never develop the pattern recognition that lets them walk into a pitch meeting and spot the fatal flaw in the sponsor's logic before anyone asks a question.

›Build the first model yourself, even if it takes four hours and AI could do it in twenty minutes
›Have junior bankers construct models from scratch on the first deal they work on, before they are allowed to use AI as an accelerant
›When you review a deal memo that AI drafted, rewrite the deal rationale from your own reasoning and see where it diverges from what the AI produced

Differentiate on What AI Cannot Commoditise

Every bank now has access to the same AI tools. Capital IQ AI, Bloomberg AI, and ChatGPT are not proprietary. Within two years, a competent junior at a mid-market boutique will produce a financial model that is structurally identical to one from a bulge bracket bank. The client will not pay for the model. They will pay for the judgment that came before the model and the conviction that comes after. Your advisory edge now lives in the work that cannot be automated: knowing the sponsor's playbook so well that you spot when they are deviating from it, understanding the subsector dynamics so deeply that you see a consolidation coming before it happens, or building a relationship with a CFO over five years so that they tell you the real constraint when you ask.

›Spend time in the subsectors you cover actually talking to management and customers, not just building models about them
›Keep a deal journal for each major transaction where you document what you got wrong in your initial hypothesis and why
›Invest in the relationships that lead to proprietary deal flow rather than competing on the speed at which you can model a syndicated deal

Know When AI Will Mislead You in M&A

ChatGPT and Copilot are confident in ways that should alarm you when applied to deal work. An AI can generate a plausible seller's note structure or a tax opinion that sounds sophisticated and is completely wrong in the specific jurisdiction your deal is in. It can write a deal memo section on integration risks that reads like it came from a best-practice textbook and misses the operational complexity that will sink the acquisition. In M&A, confidence without nuance is dangerous. Before you use an AI output as the foundation for your advice to a client, you must personally verify that it accounts for the particular constraints and opportunities of this specific deal.

›When AI generates legal or tax analysis, treat it as a starting point that requires review by someone with actual expertise in that jurisdiction, not as a finished output
›Compare AI-drafted deal rationales to past deals you worked on that looked similar but failed for reasons that would not appear in a generic model
›Build a checklist of deal-specific risks for each engagement that AI could not possibly know and use it to review any AI-generated client deliverables

Key principles

1.Interrogate every assumption in an AI-generated model because mathematical coherence is not the same as economic reality.
2.Preserve deal instinct by doing the foundational work yourself before you use AI as an accelerant.
3.Differentiate on subsector knowledge and relationship depth because financial modelling itself will become commoditised.
4.Use AI to gather and organise financial data, not to replace the judgment that determines whether a deal will work.
5.Always verify AI-generated legal, tax, and operational analysis against the specific circumstances of your deal because confidence is not correctness.

Key reminders

Build your own model on the first two deals you work on in a subsector before allowing AI to draft them for you.
When you review a financial model that AI produced, change one assumption and trace the impact through all three statements to verify it is right.
Keep a subsector-specific assumption guide so you can spot when AI has used a generic comparable instead of your market knowledge.
Spend time with management and customers in your target sectors so your sector knowledge stays ahead of what AI can gather from public sources.
Require junior bankers to articulate what would break the deal before they are allowed to present the financial model to a client.