How do you practically get started with applying AI to the deal lifecycle?
AI development framework
By assigning a responsible person, or a team, to create a step-by-step AI development framework before your organisation starts its AI journey. We provide an example of the framework below if you are choosing to custom-build or fine-tune an off the shelf model:
2. Prioritise use cases
- Define use cases based on objectives, available resources and risk tolerance
- Consider resources such as technical capabilities and budget
- Consider tasks that are low risk, time consuming and repetitive
- Involve risk management experts
- Create a sandpit environment with clear guardrails, and allow select staff to experiment with AI and pass on lessons learnt
3. Identify sources of funding
- For the cost of design, deployment, training and maintenance (costs will differ depending on whether custom building or licensing an off the shelf model)
- Redeploy capital from other technology projects or raise new capital?
4. Identify risks and mitigating actions for use cases
- Create risk management framework (see section 4 of this resource) [link to it]
- Involve legal and risk management experts
- Weigh up benefits and risks of the tools
6. Define infrastructure to use
- Buy vs build vs hybrid (see Buy or build? section below)
- Experiment with the tools and system integration
- Perform supplier evaluation
- Agree on specific AI model to use
7. Develop AI governance and business policies
- Assign person responsible for implementation and evaluation of tools
- Ensure business’s policies cover responsible AI use
- Include data protection, privacy, confidentiality and intellectual property considerations
8. Manage the change from the outset
- Create change management plan for both people and processes
- Define clear governance, communication strategy and a detailed execution plan
- Outline expected timeline to implement
10. Train and test the AI models
- Develop error analysis process
- Assess outputs against precision and accuracy required
- Test and adapt the tools accordingly
11. Build team and skills
- Identify technical specialists required
- Define people strategy for implementation and for business as usual
- Determine new training needs and upskill up-skill staff or role redesign for job roles augmented by technology
- Design learning program to expand the AI skills of M&A teams, including the responsible use of AI
12. Implement the tools
- Consider implementing AI tools in stages – start with a simple model and then gradually improve or add on
- Incorporate user feedback into the implementation phase
- Create thorough documentation to support with future maintenance and onboarding new joiners
13. Monitor performance of the tools
- Continuously evaluate outputs against expected output and set objectives
- Including ongoing supplier oversight (if AI tools bought)
Firms that select to use off the shelf Gen AI tools, without further customisation may not need to go through steps 5 and 9 as the tools will already have been trained using pre-determined prompts on large quantities of curated and unbiased open-source data. The tools will; however, still need to be tested to ensure they are working as expected.
Use cases
Read how Moore Kingston Smith implemented AI in its audit practice and the steps it took to ensure its success.
As of August 2024, many organisations with M&A activities report that while they are exploring AI to shape their M&A deals, their use of generative AI solutions is in the proof-of-concept and training phase and not yet at full scale deployment. Lessons will be learnt as they shift into the next phase, and we will share these as they become available.
Data cleansing
When it comes to cleaning up data to use along the deal lifecycle, each firm needs to consider what risk they are willing to take – there is a school of thought that waiting for ‘perfect data’ is not always practicable, and it can often not be cost effective to clean up the data perfectly. Using statistical techniques, it would be possible to find the errors and decide whether to fix them or instead to work around them.
Buy or build?
When it comes to implementing AI tools, users generally have the following options:
- use free AI tools such as ChatGPT (although this is not recommended due to pooling of public data and data confidentiality being comprised);
- buy pre-built third-party solutions (i.e. licensing an existing AI tool and ringfencing training data to the organisation only); and
- building custom built- solutions to meet the unique needs of a particular user.
Each approach has its considerations, as outlined in the table below.
| Licensed model | Proprietary model | |
| Upfront investment | low | high |
| Time of deployment | low | high |
| Ownership of algorithm | none | yes |
| Control of data | unknown | known |
| Understanding of inner workings | low | high |
| Benefit from rapid development and improvement | high | low |
| Ongoing maintenance costs | low | high |
| Pressure to provide computer power | low | high |
The risks associated with implementing AI will differ slightly depending on which approach is taken.
Due to the significant cost and time to train a model, most firms will either buy third-party AI solutions and use them as intended, or they may choose a hybrid approach of licensing a third-party AI model and fine-tuning it by training it on additional data that is proprietary to the organisation.
Firms using the same off the shelf AI models will have access to the same widely available data and will likely have similar outputs from the tools. The key differentiator between firms will be those that train their models using unique data.
Ownership of the source code and data
Open-source AI refers to artificial intelligence software whose source code is freely available for modification by anyone.
Open-source AI encourages collaboration and transparent development.
They are typically free to use and require far more modest computing infrastructure, since the models have already been pre-trained. Meta’s Llama and Mistral AI’s Mixtral models are examples of open-sourced large language models that organisations use as deep learning models.
Proprietary AI systems refer to AI software that is owned by a specific organisation or individual and where the source code is kept secret and is controlled by the organisation that developed it. Proprietary AI is developed for commercial purposes and may be sold or licensed to other businesses.
Proprietary AI can be either bought off-the-shelf or custom-built, but it is always owned by a specific entity. Custom-built AI, however, is specifically designed for unique needs and may or may not be proprietary.
Mindset
Having the right mindset towards AI is a crucial factor to its implementation success. This includes having the following traits:
- Adaptability: be open to try something new
- Curiosity: ask lots of questions and have a genuine interest in learning more
- Growth mindset: be open to constantly learning and updating knowledge and skills
- Problem-solving: have a drive to find solutions
- Collaborative: work closely with different specialists (legal, IT, HR, audit, cyber, risk management etc.)
Disclaimer
This AI in Corporate Finance content is being provided for information purposes only. ICAEW will not be liable for any reliance you place on the information in this material. You should seek independent advice © ICAEW 2024