Any enterprise business that cares about reliable AI automation has expectations about what each such automation application will do, what process it will follow, what its behavior will be, etc. This is especially true because many of the high-value processes being automated may have regulatory requirements, security constraints, and compliance considerations. Platforms that “engineer” AI solutions without formal mechanisms create brittle, unreliable, and high-maintenance. In Thunk.AI, reliability begins with the application model.
Every AI automation in Thunk.AI is an instance of a well-defined application model called a thunk.
The business user (typically the owner of a business process) describes what they want the application to do in a granular way. Much of the logic may be expressed in natural language, but it is definitely not one mega “prompt” that can be interpreted in many different ways each time it is run. There are important artifacts in the thunk application model:
Thunk Application Model
Workflow Plan
AI Tools
Modular reuse of tools
AI Instructions
MCP connections and business applications
Content Folders
Users and Access Controls
All these artifacts together are used to express the intent of the automated process. The application model is a contract between the platform and the user designing the thunk. When it comes to AI reliability, the information expressed in the application model captures the definitive intent of the AI automation.
In keeping with the principle of minimizing granularity, an overall business process is typically represented as a deterministic flow of smaller steps. Each workflow step becomes a unit of AI logic. Within a step, individual tools invoked by the AI agent might also contain encapsulated AI logic. The logic in a step may also search content folders, which may also encapsulate AI logic that governs how information is indexed, searched, and retrieved.
AI Instructions are the representation of the user’s intent for one granular unit of AI logic (most commonly, one step of a workflow process). Just because intent is specified does not mean that the AI instructions have no ambiguity, inconsistency, or incompleteness. As with any natural language instruction, those issues certainly arise. However, there are features in the model that allow the user to provide greater clarity at finer granular levels as needed. And of course, there is an AI agent to help with the authoring and refinement of the application itself.
There are three aspects to AI instructions:
Guidance
What it should do, includes examples which are particularly useful as “show-by-example” guidance. In other AI systems, this might be described as “prompts” or “instructions”.
Context
What information it has access to and informs its decisions. This might include information gathered in earlier steps and content from business applications.
Constraints
What the agent is allowed to do, and what it cannot do. In other AI systems, this might be described as “guardrails”.
A key element of the application model is schematized workflow state (a set of strongly structured properties that each workflow automation is expected to “fill in” as it proceeds). Specific properties from the workflow state are identified as Inputs (minimizing context) and Outputs (minimizing agency) for each AI instruction. The scoped context as well as the strong constraints imposed by the schema play a significant role in enhancing AI reliability.
Another key aspect of the application model is the use of AI tools. These might be defined by the user designing the automation, or they may enable access to business applications via the MCP protocol. The AI Instructions provide the ability to specify exactly why tools are available for any unit of AI logic, and to further constrain the parameter values that the AI agent is allowed to provide the tool if it decides to invoke it. For tools that provide context, these mechanisms allow the specification of the minimal relevant context. For tools that enable actions, these mechanisms allow the specification of minimal agency.
The guidance in the AI instruction can include deterministic hints like a planned sequence of AI tools calls, or a sequential sequence of prompts. This derives from the principle of persisting and reusing prior decisions to improve consistency.