DSS is a mature technology in organizations, providing support for structured and semi-structured decisions, mostly in production, operations, marketing and transportation [9]. Support for unstructured decisions is still primitive. In a pathological case, an executive at a Japanese electronics company used a version of “rock, paper or scissors” to make a $20 m decision [10]. The key to supporting these types of decisions still lies in understanding how they are made. The phases of decision making are widely understood as Intelligence, Design and Choice although these are not necessarily sequential [11]. The Intelligence phase concerns the identification of the problem, an area that has received much attention in the organizational literature as problem formulation. According to [12], managers become aware of problems through informal channels of communication. Then they attempt to pigeonhole the problem into “categories” such as “mission and goals,” “organizational structure,” “resources” etc. [12] [13]. Cognitive scientists believe that categorization is followed by a mental representation of the problem that includes details such as constraints, objects, procedures and operators which enables them to solve it [14]. Both the problem formulation and categorization stages involve retrieving information from stored schemata, an area that technology could support. This stage overlaps with the Design phase of the decision-making process. Choices are made to maximize decision-maker utility, as in the classical literature with experienced managers giving extra attention to see that decision constraints are fulfilled [14]. Understanding these stages will certainly contribute to the state-of-the-art of DSS.

In the meantime, specific contributions are required in three areas. The first is Decision Representation—to develop a graphical method of representing a decision, similar to the representation of a process (see also challenge#9). A universal methodology, like UML for representing alternatives, constraints, information, rationale, micro-decisions, actions and impacts of a decision is required. Along with this, a tool that provides an unconstrained interface to formulate the problem and additionally to find a solution is required. With the tool, a decision-maker should be able to “sketch” decision alternatives, sub-problems, assumptions, weights, etc. and freely hop among these (see [15] for a similar idea). Drawing a diagram with the Microsoft Visio tool, using different symbol sets, is a comparable design metaphor. Such a “doodling” tool is one of the great challenges in IS. The second challenge is modeling subparts of the decision with existing DSS models such as L.P., Probabilistic, or simple algebraic models. For example, can we model the impacts of a decision as a mathematical relationship between the decision objective and a possible action [16] ? Is such an approach feasible? What information will the subparts require? Is it available from the decision situation as depicted by the model? How can we aggregate results from these sub-decisions? If it is not possible to model subparts with DSS models, is it at least possible to have a schemata for each category of problem, in a manner somewhat analogous to word/email templates? This is the third area where research is required.

Natural language interfaces (NLI) will be the ultimate interface to information systems since no menus or user training is required. NLIs have been developed for databases [17], visualization software, medical and tourist information systems [18] [19]. But not all domains lend themselves to NLI applications, and graphical software being one example. So the first question to address is, for what types of software are NLIs best suited to? Perhaps it is not the software as much as certain actions that are suited to NLI. What are those? Assuming this is addressed, other issues arise during the process of interacting with the software [20]:

1) Understanding the User’s Request—there are two approaches to understanding the user’s request. The first is to break the request into parts of speech to understand the subject and object of the request. This presents the usual problems of processing NL which will be addressed in time. The second and more promising approach is to map the request into a pre-defined pattern. The request and its intent have to be understood in the context of the user’s previous actions as well his/her current software context in order to select a sequence of actions.

2) Mapping the Request into a Sequence of Actions—once the request is understood, the next step is to identify a sequence of actions that will fulfill the objective. Classical issues in means-ends analysis and hierarchical planning arise (see for example [21] ). A compromise approach where the complexity is limited should be taken to ensure that these problems do not arise. What then is the right level of complexity that can be handled by the system without running into planning problems? Additional questions that arise are, are there canonical forms for task-execution sequences? Can execution sequences be somehow auto-generated from the software itself to avoid coding them explicitly?

3) Determining the User’s Context—the most important problem here is identifying the user’s present context. Issues arising here are how can context be determined within an application in relation to the task that the user is carrying out? What knowledge should the system have about previous and current actions? About his/her computing environment? If the user is asking for a heading change in a file, how does the system recognize the file the user referring to?

4) Error handling and help—the system needs to interact with the user when their request is not understood, incorrect or poorly phrased. What are effective ways of dealing with help and error handling? Should the system present alternative formulations of the request? Suggest a list of functions that the user can use? Or fill in missing parameters from what it knows?

Task or functional support is concerned with technological support for performing daily tasks. A simple example is a researcher preparing a research paper that has a list of references at the end. He/she ought to be able to ask the system to use IEEE referencing style for formatting them, but this is not possible in current systems without using specialized software. Despite the plethora of information technologies available, office workers are poorly supported in their work [6]. They cannot automatically send emails or collect information from an email message (some of this capability is appearing in contemporary systems). Technology support for professionals is even poorer. It has not evolved past the use of specialized software such as CAD/CAM, LexisNexis database or a business intelligence tool. So an engineer carrying out a design cannot ask the system a question such as “What is the yield point for a 2" thick 10' long I beam?”

Functional support requires four types of capabilities on the part of office systems [6]: 1) Answer requests for information from documents stored in the system, 2) Store and retrieve assorted information such as a vendor offering a particular type of discount, 3) Answer questions about the employee’s domain/work (essentially knowledge management) and 4) Carry out an action or a sequence of actions. The system should support the office worker like an assistant [6].

The first requirement is to answer requests for information that is in the system. The field of information retrieval focuses primarily on retrieving documents using keywords/terms [22], but in this realm, retrieval of parts of documents is an important problem [6]. It is up to IS researchers to identify different scenarios in which this requirement arises and then pursue a solution using techniques from the field of information extraction (see for example [23] ). The second capability in the list requires suitable organization of information and perhaps the use of AI technologies [5]. The third requirement, question answering capability requires knowledge to be encoded using a suitable representation method (challenge#6) as well as natural language interfaces (challenge#2). Although there are some proposals [24] [25], research is still required to operationalize these schemes to develop technologies for a production implementation. Lastly, the requirement to provide clerical support for mundane tasks requires re-architecting office systems so that its functionality is available as services to end-users as well as other applications [6]. For example, it should be possible to send an email while editing a document or to schedule a meeting from an email client. Here also different usage scenarios need to be identified. This and NL capability (challenge#2) will allow requests such as “Email the Frankfurt report to all VPs of the company” to be fulfilled by the system [6].

The information and internet revolutions have spawned a number of different file formats for documents, images and data. For example, there are docx, html, xml, pdf, txt, and rtf formats for documents and bmp, jpg, png, and gif formats for images. At present, we will confine our discussion to physical file formats for documents only. A variety of applications including workflows, patient records, maintenance records, web searches and IoT (Internet of Things) related communications require standardized archival file formats that have to withstand the test of time. Is such a universal document format possible? XML has been introduced as such a standard but it has several shortcomings [26]. The first is ambiguity in the meaning of tags used since these are user-generated and subject to linguistic problems. Secondly, it lacks mechanisms for encoding relationships between elements i.e. other than as a hierarchy of elements [26]. Thirdly XML is verbose and requires a DTD file for interpreting it correctly. Fourthly, it is possible to encode only the linear or hierarchical structure of document elements, e.g. “a”, “b”, “c” (linear) or “a” (“b” (“c”)) (hierarchical) where a, b and c are elements of the document. So a table structure is awkward to encode [27]. A similar problem occurs with office forms [6]. Modeling of form components as normalized tables makes it difficult to model operations on them in the same way that hierarchical structure in hierarchical databases made queries difficult [6]. An intermediate representation therefore presents a layer of complexity. A possible solution is to go beyond basic data types (e.g. integer, character, text, etc.), standardize more complex data types such as data ranges, min-max values, list of tasks, meeting minutes, etc. and incorporate them into the interpreter/browser. Conceivably these could be implemented as XML extensions. It will enable any document type to be modeled independently of the technology. For example, document A could be encoded as a + b, while document B could be a + b + c, and document C could be b + d where “a” “b”, “c” and “d” are components of a document. As a concrete example, “a” could be user profile on a web site and “b” could be a list of patient medications, dosages, etc., so “a + b” would be a patient record. Only a comprehensive use-case analysis can reveal the complex types that are required. This addresses only part of the problem. The second issue is to have a more elegant solution to describe the appearance of the document, that would integrate well with the semantic model. Markup languages such as HTML may be thought of as a possible solution, but in these, data is stored without interpretation. For example, items such as graphs would be stored simply as images rather than with their semantic content, i.e. what is the graph about?

Executive information systems (EIS) have been introduced in the eighties. Their main functions are to provide high-level summarized information (from operational databases) and to provide support for analyzing this information. The information can be internal or external but generally speaking, external information is more valuable than summarized data from operational databases. The data and analysis capability are currently provisioned by Business Intelligence (BI) tools [28] [29]. These can provide summarized information in a variety of formats with an almost infinite variety of data couplings. BI tools appear to have replaced EIS completely, but some fundamental problems remain even with the addition of data/analysis capability.

The first is access to external information. Most EISs do not provide access to external information [30], presumably to avoid exposing sensitive company information. BI tools tend to be standalone tools with the ability to import data. Top management needs access to soft, non-financial data and strategic information that is currently a gap in EIS capabilities [29]. Regardless of the source and type of information, executives use it to make strategic decisions [31] and this is another bottleneck in the hyper-reactive 21^st century business environment. To keep up with this velocity, it is necessary to rely on AI techniques to process and analyze trends; for example, to assess how a company’s valuation is affected by a change in regulation. This has been missing in EIS research and is obviously the most important problem in this area. While there has been some research, there are no good models of how an organization or an organizational unit is affected by environmental forces (see also challenge#1). Secondly, there is no support for analyzing environmental information, such as creating scenarios or simulations [31]. The third issue is with respect to filtering the information. Critical success factors (CSFs) or strategic business objectives are often used as development methodologies for executive information [32] but in practice the volume of information is vast and dynamic, making CSFs an unsuitable paradigm. When executives are responding to a specific situation, they will need to obtain and analyze very specific information. For this reason, filtering information is an important research problem. Filtering requires the user’s context which is typically ignored in Information Retrieval systems [33]. For executives, the user’s context translates into markets, products, customers, etc. but these will not be simple encodings. So there is a need for user representations of context and at the same time, representations of document content are also needed to enable retrieval, spilling into other related challenges (challenges #3 and #6).

Professional knowledge is deep knowledge associated with a particular domain such as stock trading, marketing, banking, engineering, etc. [24]. It describes objects, events, actions, situations, concepts, objectives or policies but consists mostly of abstract concepts [24]. In professional knowledge, there are complex relationships between such concepts including mathematical, axiomatic, logical, temporal, structural, etc. Traditionally rules have been used to model this knowledge but these are not useful for representing declarative knowledge such as “a ‘trade’ (an ‘action’) being one of ‘options’, ‘stocks’, ‘bonds’ or ‘mutual funds’ ”. This is a typical class-subclass relationship (structural) that is better modeled with declarative schemes. Professional knowledge is filled with abstract concepts that are defined in terms of other abstract concepts, creating a knowledge engineering challenge [24]. In the previous example, “trade”, “options”, “stocks” etc. are all abstract concepts with structural relationships among them. There could be elaborations or restrictions on both concepts and relationships such as conditions under which the relationship is valid or some sort of agreement that enforces the relationship. For example, the allowable condition of a trade is that there should be sufficient funds in the account. These funds could consist of “cash” or “unsettled trades” or both. Thus concepts can have elaborations on relationships, but there are additional requirements for the representational mechanism. Firstly, the knowledge stored in the system should be viewable/editable by employees, so it should be at the “conceptual level” rather than at the “implementation level” of Brachman’s knowledge levels. Graphical representations are easier to comprehend so this is a requirement. Secondly, the representation scheme should support the formation of abstractions so that concepts may be defined in terms of other concepts. Class-subclass relationships and concept definitions are common abstraction mechanisms. Thirdly, the scheme should support modularity which takes the form of partitioning. It should not be a monolithic graph that will present difficulties in comprehension, update or assertions of facts. In other words, it should be easy to divide the representation into parts and perhaps store these parts or make assertions independently. Fourthly, the scheme should be extensible, so that new concepts/constructs can be easily added. Lastly, since the amount of information in organizations is large, the techniques should scale up to accommodate volume [5].

Knowledge management (KM) is the explicit management of organizational knowledge, including tools and processes to create, store, access and disseminate organizational knowledge [34]. KM ideas have been in existence for at least two decades now, yet very little progress has been achieved in terms of technology. The prevailing approach in IT is to use knowledge repositories [35] with little consideration being given to the problem of retrieval or the need for keeping thousands of knowledge items up to date [5]. The trend toward virtual organizations and exploding knowledge in high-tech industries exceeds the bounded rationality of organizations, if not their budgets. The leader in addressing KM problems was not academia but Siemens with its Sharenet [2]. The repository approach is not practical if the use case requires answers to fine-grained queries (see also Challenge #2 and #3) such as “What should be the blade angle for a jet engine, if it needs 16,000 lbs of thrust at an altitude of 25,000 feet, facing a head wind of 15 knots?” Such queries can be answered only through AI approaches [5]. But these are poorly developed at best [24]. Progress in representing complex knowledge is needed as outlined in Challenge #6. Since knowledge in a repository has to be updated, it should be both human and machine processable. For this reason, graphical representations should be used [24]. Further, the knowledge should be checked for consistency, so maintaining knowledge integrity is another challenge for researchers. A philosophical issue here is when knowledge is updated, should it replace existing knowledge or should it be another version of the same knowledge? Would having ten versions of the same knowledge item defeat the purpose of a KM system? The details and functionality of the KM engine should be worked out as well as the best way to enhance the user experience (UX). These issues are definitely within the scope of IS research.

The internet of things (IoT) is upon us with billions of smart devices that communicate with other smart devices. Designing applications that rely on IoT and communications with other systems to achieve distributed intelligence will be a major challenge for the future [36]. A car that has been sold in a dealership should be able to communicate to a Dealer Inventory database as well as to the Department of Motor Vehicles to update its registration [3]. How to achieve this without causing chaos or conflicts while preserving security and privacy is the “mother of all problems” and can serve as a gateway for many useful applications.

The basic IoT architecture is a wireless sensor network that collects data from sensors (e.g. a utility meter). This data could be sent to a local data store or a cloud-based server [37]. In some cases, actions can be taken based on the data, such as turning on a pump. For some applications, the data is cleaned and processed while in others such as meter readings, it is collected for other purposes. Data can also be aggregated and mined for patterns/conclusions [38]. There are networking/architectural challenges in having common transport and application protocols given the heterogeneity of sensor/actuator devices and networks [37]. Maintaining security and privacy in a ubiquitous data rich environment is also a major hurdle [39]. For IS researchers, the challenge lies in developing a framework that can facilitate data sharing (at the user level) since this will be a crucial component of distributed AI. The necessary protocols for handling this need to be developed. Issues that arise are: What data is public vs private? What checks have to be carried out before allowing information access from a device such as a health monitor? If a building detects a person with a weapon entering it, can it identify the person and check his/her firearm permit? Can it alert law enforcement if there is no permit? Philosophically, where should the intelligence that carries this out be located? Who should update it?

Software engineering is a massive but mature industry. Potential developers are confronted by a bewildering array of tools and methodologies for specifying requirements. Existing methodologies can be classified into two main branches based on whether the development approach is traditional or agile, using object-oriented development [40]. There are a number of methodologies in each of these for specifying the development. In the traditional approaches, there are tools such as DFDs, structure charts, data dictionaries [41], while for the latter approach there are tools such as UML diagrams, sequence and interface diagrams. Most research approaches have focused on specialized systems or on individual tools. See for example, [42] [43]. There is a compelling need for a uniform and integrated specification method (that is not unwieldy) to document or define: 1) objects and their relationships, 2) processes, 3) process calling sequences, 4) data relationships, 5) decisions 6) interfaces and interface relationships, 7) system features, 8) business rules and 9) organizational knowledge. Such a universal method ensures that the same methodology can be used for different types of systems and additionally, for creating a system’s entire documentation in an integrated fashion. UML has been proposed as one such standard, but it has several limitations [44]. Among these is the inability to capture: 1) system-actor interactions (for example, a system informing an actor of its unavailability), 2) organizational context (for example, interactions between actors), 3) use case relationships (for example, part/subpart), 4) state-dependent system behavior (for example, a use case depends on a system’s state). Modeling of information flows is also awkward in UML, but is a basic requirement of a specification methodology [44]. A universal methodology that overcomes these limitations and adapted to specifying/documenting different types of systems is required.