I did create a constraint based authoring tool, named PATH, which can be used to author tutoring systems. I made some experiments with it and I got some conclusions regarding your question,which are sumarized in the next lines: The first variable is the size of the models. Normally, the bigger the models is, the higher the number of constraints that will have to be modelled, although it also depends on how the model can be reused for different tasks. Another important issue is the characteristics of the domain where the ITS is built. These will determine the difficulty of modelling the domain formally. Regarding the constraint modelling, it is directly influenced by the characteristics of the domain and the precision that is needed in the diagnosis. If the evaluation of the task needs to be very detailed, it will require the definition of more constraints, so the cost of generating the task will be greater as well. What is more, the definition of the interpretation model is subjective, that is, there are many ways to model steps and situations. Therefore, depending on the skills of designers, some designers will take longer than others when they define the models. 

Due to all these variables, it is not feasible to obtain objective metrics that realistically determine the general cost of creating a task model with PATH, because the creation of each IILS will have its own issues. 

This part of my research is not published yet, anyway, it may help you as a guideline for your question.

The challenge in such tutoring system is the level of complexity of the subject, the level of tutoring (levels of depth), and the number of situations/combinations/possibilities of tutoring support at the point of need. in addition of correctly identifying the problem, issue or error, the authoring of the remedy should be effective in helping the student. It can therefore be a complex problem, and depending on the problem simplicity-complexity, number of erroneous problems, etc, an approach using  error analysis, though doable, might only be effective for relatively simple problem scenarios.

Perhaps an alternative approach might be considered, and that is to use social learning model in which the participants play a role in creating what i call convergent knowledge as their contribution adapts yet converges to the right/best solution via student participation, with the teacher becoming a facilitator. Of course, the student interactions can be captured to document-categorise various different student responses to the problem, and that might then be used the basis of providing an automated tutor solution for that specific problem later on.

Bob,

there are several authoring tools worth considering:

• Flatworld http://www1.flatworldknowledge.com/cbe
• CCKF http://realizeitlearning.com
• Acrobatiq  http://acrobatiq.com/products/the-learning-dasboard/
• DecisionSim http://decisionsimulation.com
• Branchtrack https://www.branchtrack.com
• Simwriter http://simwriter.com
• SmartSparrow https://www.smartsparrow.com https://www.youtube.com/watch?v=A1oLmPAFcA4

Some of these may have links to associated studies on associated development time.

We are well advanced on a new web-based tool known at the moment as ‘Branch’. The idea of this tool is to simplify the process for the development of competency-based adaptive tutoring-systems including simulation components to embed in webpages or iBooks. The idea is to place the development process in the hands of instructors rather than developers. Simple building blocks may be developed, reused, repurposed and shared. We believe this will have a significant impact on development time. There are all sorts of factors that can impact on development time all related to the scope of your project.

Following.