Creating Large Action Models for Mergers and Acquisitions
A reference model can serve as the process and action blueprint for a Large Action Model (LAM): it defines what needs to be done, in what sequence, by whom, using which information, and with what outcome. The LAM then learns to select and execute those actions through APIs, software tools, or human handoffs.
1. Convert processes into action units
Break each reference-model task into structured steps:
- Trigger: What starts the task?
- Objective: What must be achieved?
- Inputs: Required documents, data, or decisions.
- Actions: Concrete operations the agent can perform.
- Roles: Responsible person, reviewer, approver, or external party.
- Outputs: Produced documents, records, decisions, or status changes.
- Rules: Preconditions, constraints, and escalation criteria.
- Exceptions: What to do when information is missing or contradictory.
The M&A Reference Model is particularly suitable because it organizes work into tasks such as target search, target evaluation, due diligence, and IT integration, while the full version includes subtasks, assigned roles, and data objects. [manda-automation](https://www.manda-automation.com/free-reference-model)
2. Represent the model as machine-readable data
A useful action schema might look like this:
```json
{
"task": "Evaluate target fit",
"trigger": "Target added to shortlist",
"inputs": [
"target_company_profile",
"m&a_strategy",
"evaluation_criteria"
],
"actions": [
"retrieve_target_data",
"score_strategic_fit",
"score_financial_fit",
"identify_missing_information",
"prepare_recommendation"
],
"tools": [
"crm.search_company",
"financial_model.calculate_score",
"document.create_report"
],
"outputs": [
"fit_scorecard",
"risk_register",
"recommendation"
],
"approval": "corporate_development_lead"
}
```
This transforms descriptive process knowledge into a combination of state, actions, tools, and expected outcomes.
3. Build an action ontology
Create a controlled vocabulary for:
- Tasks and subtasks.
- Roles and permissions.
- Business objects, such as targets, contracts, financial statements, and integration plans.
- Events and status changes.
- Tools and APIs.
- Decision types.
- Risks and exceptions.
For example, “request missing financial statements” should be represented as a distinct action rather than left as vague prose. This lets the model learn that the action requires a target identifier, a list of missing documents, an authorized sender, and a follow-up deadline.
4. Create training examples
Generate demonstrations from the reference model. Each example should connect:
\[
\text{Context} + \text{User intent} \rightarrow \text{Next action or action sequence}
\]
Example:
```text
Context:
The target is on the shortlist. Revenue data is available, but customer concentration data is missing.
Intent:
Prepare the target for financial due diligence.
Expected actions:
1. Check the due-diligence data checklist.
2. Detect missing customer-concentration data.
3. Create an information request.
4. Assign it to the target contact.
5. Set a due date.
6. Update the diligence status.
```
The resulting dataset can support supervised fine-tuning, retrieval-augmented generation, or an agent planner. LAM development commonly requires action-sequence data, tool integration, grounding in the operating environment, and evaluation. [arxiv](https://arxiv.org/abs/2412.10047)
5. Connect actions to tools
Each executable action needs a tool contract:
```json
{
"name": "create_information_request",
"description": "Create a request for missing diligence information",
"parameters": {
"target_id": "string",
"documents": "array",
"recipient": "string",
"due_date": "date"
},
"permissions": [
"corporate_development"
],
"requires_approval": false
}
```
The LAM should not merely generate text such as “ask for the data”; it should produce a validated function call. Tool-learning systems generally require the model to understand available tools, plan their use, observe results, and revise the plan when necessary. (https://www.iese.fraunhofer.de/blog/large-action-models-multi-agents/)
6. Use the model as a process graph
The reference model can become a directed graph:
- Nodes: Tasks, decisions, documents, and states.
- Edges: Preconditions, sequence relationships, dependencies, and escalation paths.
- State: Current transaction phase and completion status.
- Policies: Rules determining which actions are allowed.
For example:
```text
Target shortlisted
↓
Evaluate strategic and financial fit
↓
Approve diligence
↓
Request diligence data
↓
Review findings
↙ ↘
Proceed Escalate issue
```
The LAM uses the graph to determine the next valid action instead of improvising an arbitrary workflow.
7. Train and evaluate in stages
A practical development path is:
1. Retrieval prototype: Give the agent the relevant process and task definitions at runtime.
2. Tool-use prototype: Add APIs for document retrieval, workflow updates, email, scoring, and reporting.
3. Supervised action model: Fine-tune on approved examples of contexts, tool calls, and outcomes.
4. Simulation: Test the agent against realistic M&A scenarios and incomplete data.
5. Human-in-the-loop deployment: Require approval for high-impact actions.
6. Continuous learning: Store successful, corrected, rejected, and escalated actions for evaluation and retraining.
Measure not only language quality, but also:
- Correct next-action selection.
- Validity of tool parameters.
- Compliance with role permissions.
- Completion of required subtasks.
- Accuracy of outputs.
- Appropriate escalation.
- Recovery from tool failures.
- Absence of unauthorized actions.
Because the publicly available page is explicitly a free, redacted version, a production implementation should verify licensing before extracting or commercializing the complete JSON dataset. (https://www.manda-automation.com/free-reference-model)
Parts of this blog entry might be AI generated
Advertising section
Build your AI skills by licensing our offerings
A Modern Post-Merger Integration Playbook: From M&A Models to AI Solutions
By Dr. Karl Michael Popp
Master integration due diligence to transform your M&A success. Learn more at manda-automation.com