From an M&A due diligence perspective, there are still some relevant limitations to the use of AI algorithms, in particular:
- Accessibility of information: data in virtual data rooms (VDRs) is usually well protected and the sell side is often not willing to have their confidential information used for AI training. Also, the VDR provider is generally mandated by the seller. Hence, developing and training algorithms for M&A due diligence may be very limited "on the fly" in running transactions. Instead, this process may be based on a vast amount of public and anonymized data.
- Reliability and data protection: LLMs still tend to bluntly "invent" facts (often referred to as "hallucinations") or to leak parts of the input data they have been trained on. While this is not a considerable issue if the LLM is being used to plan a family vacation or to compose Christmas cards, it is a showstopper for using sensitive data in a VDR and relying on AI findings.
The opening: analysis of publicly available information and setting up a VDR
For the buy side, the first step in a due diligence process is often to delve into the public sources available on the target. The advantage of such outside-in due diligence is the (often vast) availability of information. Further, the buy side can perform a first analysis of the target without zeitnot (time pressure). At this stage, generative AI can examine public sources such as press releases, ad-hoc announcements, financial reports, prospectuses, and media coverage for ongoing tax and legal disputes. The issues identified in such an AI-generated report may serve as customized input for the information request list and the management interviews in the due diligence process.
On the sell side, setting up a VDR is often time-consuming and burdensome, tying up substantial operational resources. AI algorithms can streamline the process by automatically organizing the uploaded documents. Further, the software may support the team by checking the documents for sensitive information and directly proposing redactions. This capability provides considerable time and efficiency benefits to the sell side, particularly for transactions involving many documents with employee details, competitively sensitive information, and intellectual property.
The middlegame: VDR due diligence
Once a VDR has been opened to the potential buyer(s), AI can play to its primary strength by detecting gaps and providing summaries of the documents uploaded. Both make extracting relevant information easier for the due diligence teams, particularly on larger transactions and across different languages.
Example: The annual report mentions that the target has sold a real estate property in FY 2022. AI can pick this up autonomously and check whether all the documentation that is common in connection with such a sale is available. In seconds, the software could flag a missing notarial deed or a tax declaration where the purchase price does not match the amounts in the financial statements.
The limiting factor is, once again, the availability of relevant data to train the AI. While there are sufficient precedents for targets involved in real estate transactions, other interdependencies are more challenging to spot for AI algorithms. Examples are the various permissions required for staff leasing and medical businesses.
One area with significant potential is the analysis of legal documents. Well-drafted agreements have a precise and logical structure similar to software code, using largely standardized language. Current AI-powered due diligence software can already look for critical clauses such as "change-of-control "and "non-compete" provisions in the target's contracts.
Hence, AI software could identify potential issues and risks such as:
- The target has paid a dividend without deducting withholding tax. No treaty clearance and/or related tax forms are available in the VDR.
- The target has sold a subsidiary in FY19 to an unrelated third party. The underlying share purchase agreement (SPA) has several unusual, buyer-friendly indemnity clauses.
- Three of the target group's credit facilities have "change-of-control" clauses, requiring the respective entities to seek a waiver for the transaction.
- The board resolutions in FY18 are not compliant with applicable corporate rules as they lack the necessary signatures.
- The target did not impair trade receivables against a distressed third party.
- There are significant off-balance sheet items due to ongoing product liability litigation in the US.
After identifying a potential risk, the next step is determining its probability and assessing its impact. For this, humans rely not only on legal and financial research but also on "soft" information such as their experience from similar cases, discussions with colleagues and authorities, academic training, as well as familiarity with human reactions. AI algorithms need access to this "soft" information in a digestible form, allowing them to learn the patterns and to reproduce the risk assessments.
Example: For not (or not timely) submitting a withholding tax form, the law provides fines of up to EUR 7,000, depending upon the taxpayer's level of culpability in the individual case. However, the tax authorities' unpublished practice is consistently applying the maximum fine in all cases. Unless AI is trained for this specific case or has access to a large number of fine notifications allowing it to identify the (questionable) pattern, risk quantification for a non-submitted withholding tax form may not be accurate.
Overall, compliance and (potential) findings are regularly discussed in management interviews. Such conversations are an essential source for experienced M&A practitioners to get an understanding of the target's risks and procedures. While it is still difficult for AI algorithms to participate in such management interviews, the written Q&A process may partially bridge this gap.
In the last step, generative AI can produce first drafts of due diligence reports based on the confirmed findings as well as on additional inputs from the due diligence team.
