In financial advisory, artificial intelligence does not win a mandate, does not hold a negotiation and opens no address book. It writes the first draft of almost everything else. And the best description of what it produces does not come from a software vendor, it comes from the regulator.
Tandem works with finance teams on AI applied to their deliverables. This article describes what concretely changes in running a mandate, from pitch to signing, and above all what does not. It addresses advisory boutiques, not investors.
The exact word for what AI produces on a deal
In its February 2026 study on AI use by financial market participants, the French financial markets authority surveyed law firms specialised in financial law and Paris advisory and audit firms. Its conclusion fits in one sentence. AI is designed there as a working assistant, producing first drafts comparable to a junior associate's work, always reviewed and validated by a responsible professional.
That sentence deserves to be pinned in a meeting room. It states both the size of the gain and its exact limit. A junior associate produces a great deal, very fast, and nobody sends that output to a client unreviewed. Nothing more, nothing less.
The gains reported in the same study are sober and credible. One firm having integrated these tools extensively reports 2 to 3 hours saved per week per user. It also reports up to 30 % less time on contract review, thanks to automatic comparison against its internal policies. Of the four advisory and audit firms that responded, three already show an advanced level of adoption.
The six deliverables of a mandate, and the one AI never touches
A mandate runs through known steps, and almost each one ends in a document. That is what makes the profession so exposed to the subject. The pitch, the angle note, the teaser, the memorandum, the due diligence tracker, the clause review, all are written deliverables.
The management meeting is the exception, and an instructive one. It produces nothing written at the time, the minutes come afterwards. No tool sits in your place across from an owner hesitating to sell. The time reclaimed on the six other lines therefore only has value if it goes back into that one.
Who produces, who reviews, who signs
The real organisational question is not whether a boutique uses AI. It is whether its review chain has been rewritten accordingly. Three roles, and only one changes hands.
In practice the review changes nature. It bears less on form, which arrives clean already, and much more on figures and assertions. That is a different job, more demanding in attention and less in time. It deserves to be named in job descriptions rather than endured.
The teaser and the memorandum, where formatting used to eat the week
On a sell-side mandate, what takes time is not the substance, it is the formatting. An anonymous teaser and a full memorandum represent dozens of hours of layout, chart tweaks and version round-trips.
What changed in recent months is not that AI can generate slides, it has been able to for a long time. It is that it finally respects a brand and outputs genuinely editable PowerPoint. The lever is to write your visual identity down once, colours, typefaces, spacing, logos, then reuse it on every document. I documented the whole method in my newsletter edition on slide generation, and the site keeps a comparison of the tools in our article on AI presentation tools.
For a boutique, the benefit goes beyond comfort. A branded teaser takes half a day instead of a week. You reach the market earlier, and you iterate on the angle rather than the layout. The document stays anonymous until the confidentiality agreement is signed, and that constraint is written into the template once and for all.
What AI controls better than it produces
Here is the use almost nobody sets up, and the one that returns the most on a mandate. On several deliverables, AI is worth far more in review than in drafting. It then produces no extra page, it prevents the error that shows up in negotiation.
- On listed comparables. Your database already provides the companies and their aggregates. The real issue is comparability. Rule out a company in an exceptional situation, harmonise earnings adjustments, and never mix trailing and forward figures in the same statistic.
- On a leveraged buyout model. Control bears on links between tabs, the balance sheet closing each year, circular references and cash that must never go negative. These are mechanical checks, exactly what a machine does better than a human at eleven at night.
- On advisers' reports. The real work is not summarising each report, it is cross-checking them and surfacing contradictions, citing report and page. A divergence between the financial and the legal report is negotiation information.
The same principle applies to the spreadsheet, which remains an analyst's real workstation. We detailed the three tooling levels, from the in-file add-in to the coding agent for heavy consolidations, in our article on AI applied to Excel and data.
Capture the house standard rather than re-explaining it
A boutique has a standard. A memorandum structure, an order of sections, phrasings that have proven themselves, things that never go in a teaser. That standard lives in the heads of two or three people, and it passes on by successive correction of juniors.
That is exactly what can now be documented. The format is called a skill, the house procedure made available to the model. Written once, it applies to every file. The subject is covered in depth in our article on the end of the prompt as the unit of work.
The useful novelty is being able to show rather than describe. You record your screen while doing the task, you narrate out loud what you are doing and why, and the demonstration becomes a reusable procedure. That is what we used to do manually at Tandem, observing teams to reconstruct the real process rather than the one in the internal handbook. I detailed the manipulation in a July 2026 post.
Where the market stands, and who has already deployed
The advisory market works against a backdrop of harder exits. In 2025, French private equity players exited 14.0 billion euros at historical cost, up 9 %, but across a falling number of transactions (France Invest and Grant Thornton, March 2026). Fewer deals, more work per deal, is precisely the context in which execution becomes a differentiator.
Deployment in investment banking is no longer experimental. The vendor Rogo states in April 2026 that its agent Felix is used by more than 35,000 finance professionals across more than 250 institutions, among them Rothschild & Co, Jefferies, Lazard, Moelis and Nomura (Rogo press release, April 2026). Gabriel Stengel, co-founder and chief executive, sums up the movement, the leading institutions are moving from automating isolated tasks to firms tooled end to end.
Two readings coexist, and both are true. The first is that tooling is becoming commonplace, so it stops being an advantage. The second is that a boutique without it now feels the comparison, when a client asks in a meeting how the file will actually be executed.
What cannot be automated in a mandate
Four things resist, and none is solved by a better model. Naming them saves six months spent trying to work around them.
- Access to acquirers. A list of potential acquirers is built in a few hours. Knowing which one will answer, and through whom to reach them, comes from nowhere but a maintained address book.
- Valuation judgement. A tool produces a range and sensitivities. Deciding the price to take to market, and standing behind it in front of a seller, remains an act of advice.
- Negotiation. Clause review flags departures from the standard and proposes alternative drafting. It does not hold a balance of power in a room.
- Signing the work. It commits the firm. No tool takes that responsibility, and no licence agreement transfers it.
A technical limit comes on top of those four. An agent works on documents gathered locally, not on an online data room consulted read-only without downloads. That constraint shapes how a review is organised, and it is settled with the seller when the process opens.
How a boutique organises itself so it holds
The most advanced firms surveyed by the regulator have all formalised a governance, and it is nothing sophisticated. A committee, a usage charter, audits, and continuous training. That is the reproducible minimum in a twenty-person firm.
| Once for the firm | On every mandate | |
|---|---|---|
| Tool | Professional plan, hosting, no-reuse commitment | Nothing, the tool is not re-debated |
| Standard | Teaser, memorandum and minutes structure | The key messages of the file |
| Perimeter | General rule on client documents | Which documents enter, and when they leave |
| Review | Who reviews what, and at what seniority | The time reserved before delivery |
The applicable regulatory frame is not reinvented either. It is the same as for the wider financial sector, and we detailed it, from the EU AI Act to operational resilience obligations, in our article on AI in finance. The investor side, the fund that owns and resells, is covered in our article on AI in private equity.
Which deliverable should an M&A boutique start with?
With the management meeting minutes, and that often surprises. They are frequent, they are tedious, they commit nothing towards the client, and they give the team a first win within days. Nobody will ever defend the right to retype their notes at night.
Then take control of the model and the comparables, before document generation. It is counter-intuitive, and it is the best return of the lot. You produce nothing more, you stop handing a client a file that does not balance.
The teaser and the memorandum come third, once your brand is written down and your standard captured. At that point the gain is counted in days per mandate rather than hours. The full scoping approach sits on our AI audit page, and the sector offer on our AI agency for finance page.
One last rule, valid on every mandate. If your tool cannot say that it is unsure, it has no business near a figure you are about to deliver. Require doubt to be flagged before you require layout quality.
