AI in M&A: What Investment Banks Are Actually Doing, and What It Means for Deals Under $5M
Short answer: yes, AI is being used on live M&A deals today, but not in the way most coverage suggests. At the large end it assembles documents and models faster while a banker keeps every judgment call. Below roughly $5 million of EBITDA the advisor is there but the analyst bench behind him is not, and the buyer often has no analytical support of any kind. That changes what the tool has to do. Speed is not the useful thing. What matters is whether a number can be traced back to a document.
This page explains both, and what separates them.
What the investment banks are actually doing
In July 2026, RBC Capital Markets published an account of AidenBanker, the AI platform now in production across its Global Investment Banking group. It is a useful public record because most banks do not describe this work in detail.
Three capabilities are described. A pitchbook builder that assembles production-ready decks in days rather than weeks. A client meeting prep tool that compiles recent activity, priorities and relationship history into a brief. And an Excel-embedded modelling assistant that populates templates, wires formulas and runs scenarios inside the spreadsheet bankers already use.
Two design decisions are worth noting because they are deliberate and they are correct for that setting.
The first is that the human keeps the judgment. RBC states it repeatedly: the AI handles data assembly and production, and the banker retains strategic judgment. That is not modesty. It is the only defensible arrangement in work where somebody signs.
The second is that they did not build a new application. The tools live inside Excel, where those bankers have spent twenty years. Adoption is the hardest part of any deployment, and meeting people where they already work is the reliable way to win it.
Every benefit claimed is a speed benefit. Faster decks, faster prep, more scenario passes before a decision window closes. That is an honest description of what the tools do.
What is different below $5M EBITDA
The bank's arrangement is not one person. It is a senior banker whose judgment is backed by an analyst bench, and the bench is the part that does not exist further down.
At RBC, an analyst rebuilds the model. The senior banker reads the output and knows when something is off. When AI assembles a deck, a trained person checks it before a client sees it. The tool is an amplifier bolted onto expertise that already has a checking layer underneath it.
In the lower middle market the judgment is still there. A broker who has closed sixty deals in a vertical knows what a business is worth, knows which buyers are real, and knows exactly which questions kill a deal. What he does not have is anyone to rebuild the model at midnight, chase down which add-back came from where, or reconcile the statements line by line. That work either gets done by him, at the expense of the work only he can do, or it does not get done at all.
On the buy side the gap is wider. A first-time buyer looking at a $3 million manufacturer has no bench and no prior deals to reason from.
That changes what the tool has to be. Speed helps someone who already knows how to do the work and simply has too much of it. It does much less for someone doing it for the first time, and nothing at all where the checking layer is missing. You cannot accelerate a skill that is not present, and you cannot trust output that nobody is positioned to verify.
So the useful question below $5 million is not how fast the machine can produce a document. It is whether anyone can check what the machine produced.
Why that makes provenance the requirement, not a feature
Consider the ordinary case. A CIM states adjusted EBITDA of $1.28 million. Where did that number come from?
Ask the broker and it came from the accountant. Ask the accountant and it came from a spreadsheet the owner's bookkeeper prepared. Somewhere inside it is a stack of add-backs, one of which is a vehicle and one of which is a salary for a person who genuinely works in the business, and nobody wrote down which is which. The buyer discovers this in month four, having spent real money, and now distrusts every other number in the document.
That is the normal failure of a small deal. It is not fraud and it is usually not even carelessness. It is that the chain from claim to source was never recorded, and by the time anyone asks, the person who knew has forgotten.
Now add AI to that. A language model will read the CIM faster than any human and restate the number with complete confidence. It does not make the number more true. It gets you to an unverified figure sooner, formatted better.
The fix is not a better model. It is a record. Every figure in a document should resolve to the page it came from, and every human correction of that figure should carry the reason it was made. Adjusted EBITDA opens the normalization memo. The memo opens the reviewed financial statements. A number with no document behind it prints as incomplete rather than as a confident guess.
There is a second layer that matters as much and gets discussed less. Not all documents carry the same weight. An audited statement, a review engagement, a compilation and an owner-prepared spreadsheet are four different levels of assurance, and every lender prices them differently. A traced number is only as good as the authority of the document it traces to, so the grade of the source belongs on the face of the record alongside the trace.
That chain is ordinary practice at the top of the market, where an analyst holds it in a working file and a data room backs it up. It has never existed at the bottom, because nobody was paid to build it.
What this means depending on where you sit
If you are buying: the value of AI to you is not the summary. It is whether you can trace a claim to a document before you spend money, and whether the questions you carry into the first management meeting come from the gaps in the record rather than from a generic checklist.
If you are selling: the same chain works in your favour. Most price reductions late in a deal happen because a number cannot be substantiated, not because the business turned out to be bad. A seller whose figures resolve to documents has considerably less to negotiate away.
If you are a broker: this is the same move RBC made, at your scale. The machine does the assembly, you keep the judgment and the relationships, and your deals spend less time stalled while somebody hunts for a document a buyer asked for.
The honest limits
AI does not change the deal process. Teaser, NDA, CIM, management meetings, LOI, diligence, close. That sequence exists because it protects sellers from unqualified buyers and gives both sides a way to stage disclosure. Nothing here skips a step.
It does not replace judgment either. Whether a business can survive its owner leaving, whether a particular buyer can carry the risks this particular business holds, whether a price is right for you specifically: those remain human calls, and anyone selling you automation of them is selling something they cannot deliver.
And it does not make a document true. A traced number is a number you can check. That is all it is, and it is a great deal more than the alternative.
Where we sit
JCoBee builds the small-deal version of this. The system reads the financials, rebuilds the earnings, tests each add-back, stresses the bank math, and keeps every figure connected to the page it came from. Human corrections are recorded with the reason they were made, so the document ends up carrying both halves: what the machine read, and what a person decided.
You can walk a sample deal without signing up for anything at jcobee.com/samples.