1、1AI Agents:What They Are,How They Work,and Where Organizations Might Best Use Them 2024|infodeep-Deep AnalysisAI Agents:What They Are,How They Work,and Where Organizations Might Best Use Them By:Matt MullenAnalyst BriefDeep Analysis The EssentialsIn 2024,the introduction of“agents”derived from artif

2、icial intelligence(AI)dominated the relatively immature marketplace in generative AI for enterprises.Whereas in 2023,discussion of enterprise AI centered upon its use as an assistant within business applications,2024 saw automation,process orchestration,and integration built onto that foundation,cre

3、ating the notion of agents.Using increasingly sophisticated,specialist models and enterprise data along with software engines that can make determinations on purpose,plan,and execution,agents promise a high degree of specificity in carrying out sets of tasks.Where previous iterations of technology h

4、ad to rely on highly detailed predetermined plans,agents can create their own plans based on sets of actions that they have been cleared to operate.In this report,we examine the characteristics of agents and how they extend the capabilities of assistants to achieve more specific,complex outcomes.We

5、also look at the operational and technical architectures that support agents operation and classify the supported actions based on complexity.Finally,we discuss the specific use cases that an enterprise might consider appropriate for deploying agents and how the preparation,implementation,and operat

6、ion of such use might look.Agents can enable enterprises to shift into a level of complex process orchestration that previously would have been resource prohibitive(in both cost and skill).To take advantage of these capabilities,a priority planning task is creating and supporting the actions needed

7、to make plans executable by prospective agents.The specificity that makes agents useful means that every action they are approved to perform needs to be supported to some extent by additional enterprise software(for example,business applications)and customer-specific configuration.Traversing the age

8、nt landscape as outlined within this report is a vital first step for organizations that recognize their own operational challenges among those proposed by the industry.Successful execution will bring a bounty of potential benefits,provided that execution is based on sound planning and successful ha

9、rnessing of human skills and experience.2AI Agents:What They Are,How They Work,and Where Organizations Might Best Use Them 2024|infodeep-Deep AnalysisFinally,looking at the short to medium term the next 12-24 months we acknowledge that initially,agents will be deployed primarily by organizations wit

10、h existing,application-specific workflows and codebases that can be utilized right away within single-vendor environments.From this starting point,the provisioning of libraries of actions from which agents can be assembled and configured will become part of vendor-provided capabilities,with gradual

11、integration support for third-party applications emerging.This might seem an unambitious trajectory,but it will still be challenging for even highly engaged organizations to achieve.Background:From Assistants to AgentsIn October 2023,Deep Analysis released the report“Workplace AI Market Analysis:Gen

12、erative AI and the Desktop(R)Evolution”,1 which detailed how relatively new technological advances had enabled generative AI(GenAI)assistants to become a brand-new element in many desktop software applications.A year later,at the time of writing this report,many of those software applications are al

13、ready augmenting those existing still fresh assistants with agents,often with very similar descriptions and overlapping messaging as to their capabilities.That report attempted to explain the moving parts around assistants including the most cited use cases and suggested approaches for implementatio

14、n and this report attempts to do the same for this next market iteration.AI agents represent a significant advance from the scope largely attributed to assistants,and this report seeks to explain that advance,how agents extend the vision for assistance much further than the“assistant”iteration,and h

15、ow in doing so,significant existing elements of an organizations information architecture start drawing much closer together.As we will discuss,agents enable Information,Transaction,and Extended use cases,which as they increase in complexity also deliver an equivalent amount of specificity.The langu

16、age of agents In part because agents came into being as products so hot on the heels of AI assistants,it has been difficult for many to comfortably articulate ways to distinguish between the two.Some software vendors have iterated their product naming and descriptions to fold agent capabilities into

17、 existing branding,while others have thrown out their branding,introduced something new,and renamed previously launched products accordingly.Confusion was perhaps not the goal,yet it has certainly been an unintended consequence.Figure 1How Agents,Reasoning Engines,and Actions Are Connected A Simplif

18、ied FlowPlanActionsActionsActionsResultsAgentReasoning EngineRequester3AI Agents:What They Are,How They Work,and Where Organizations Might Best Use Them 2024|infodeep-Deep AnalysisFor clarity,this report will primarily refer to agents,reasoning engines,and actions(see Figure 1).Agents are software e

19、ntities that interact with human or machine requesters to collect,reiterate,and ultimately return information.Reasoning engines are used by agents to create execution plans to satisfy the instructions provided through that interaction.Actions collect information,work with a transaction,or perform an

20、 extended function as laid out in the plan.Once the actions have been executed and results collected,the reasoning engine can supply the agent with information that allows it to communicate a response that a human and if applicable,a machine can understand.Specificity and complexity If we use that d

21、escription of an agent-managed interaction and compare it to those that might occur using an assistant,we see a lot of similarities.This is primarily because agents are based on many of the same technologies that underpin assistants,but with the aim to deliver more complex outcomes.In that sense,age

22、nts are broadly cumulative of assistant functionality;functions like content assembly and summarization become part of the armory that agents have at their disposal.The first assistants of this generation of AI exclusively used large language models(LLM)as a discrete source of control.Using somethin

23、g like OpenAIs ChatGPT meant making a request(a prompt)that would be interpreted and responded to solely by the underlying LLM.The LLM had to provide its own reasoning engine capabilities to interpret the request,use its own information storage to compile the response,and then deliver that back to t

24、he requesting entity(human or machine).This can produce responses often very complex ones but they are at a base level,general,and non-specific to the context of the requesting entity.This non-specific context was recognized early and an associated information architecture called retrieval-augmented

25、 generation(RAG)was proposed to allow the LLM to access internal knowledge sources in constructing the response.2 This meant that organization-specific information could populate the assistants response,increasing the results specificity but also increasing the complexity of the supporting applicati

26、ons required.Of note,RAG is also occasionally referred to as“grounding,”a broader term used to denote an AI-generated responses connection to“ground truth”(a factual basis for a response).In that sense,RAG is always used for grounding,but not all grounding is RAG.Agents are designed to bring the ult

27、imate in specificity by allowing the employment of Figure 2LLMs to RAG to Agents:Potential Increase in Response Specificity*LLM large language model*RAG retrieval-augmented generationResponse SpecificityLLM*(non-specific)LLM+RAG*(adds organization-specific information)Agents with LLM+RAG (adds actio

28、ns that enable more targeted responses to specific complex requests)4AI Agents:What They Are,How They Work,and Where Organizations Might Best Use Them 2024|infodeep-Deep AnalysisLLMs,RAG,and preconfigured actions to help resolve requests(see Figure 2).RAG is typically organization-specific,but it is

29、 a large repository often a vector database with data from potentially thousands of ingested unstructured documents.This means that the response may still lack the precision required for first attempt(zero-shot)resolution.Actions enable agents to respond specifically,with calls to end points for inf

30、ormation or transactions,or to enact automations to meet requests.These actions can be effectively stacked within a plan created by a reasoning engine,enabling responses to specific,complex requests in what appears to the user as a single step.Information Architecture and LogicThis section expands o

31、n the logical construction of agents and how they are assembled and maintained.Descriptions are necessarily generic and try to draw the broad commonality of contemporary approaches across a range of software vendors.Agents manage the interaction between the human or system that enacts the transactio

32、n,the human who calls up the UI and types or speaks their request,or the external system that intermediates the same(through an integration).The agent is configured through an administration tool to set the parameters defining its scope of operations(see Figure 3).These include things like the sourc

33、e of any internal knowledge maintained in a RAG system or similar,and which LLM(s)are to be supported in operation.Assistants utilize those same capabilities for their operation,but for assistants the reasoning engine sits within the LLM.What is different for agent architecture is that the reasoning

34、 engine is abstracted outside the LLM and can be set up with specific sets of defined permitted actions.Within this architecture,the reasoning engine has a selected range of options including those within the LLM that it can build into its plan.This means that LLM-derived functionality like content

35、assembly and summarization familiar from assistants forms part of the armory here and can be combined with approved RAG repositories to create actions within such a plan.Figure 3How an Agent System Connects Together:Simplified Logic DiagramAgents manage the interaction between the human or system th

36、at enacts the transaction,the human who calls up the UI and types or speaks their request,or the external system that intermediates the same(through an integration).PlanAdministration ToolRAGLLMAgentReasoning EngineActions5AI Agents:What They Are,How They Work,and Where Organizations Might Best Use

37、Them 2024|infodeep-Deep AnalysisWhat are actions?A plan is not necessarily a single-step operation.Indeed,depending upon how the request is interpreted by the reasoning engine,it can include different actions that have to be assembled in sequence with dependencies and conditional logic(see Figure 4)

38、.These actions will likely combine those specified by software vendors and those specified by customers within an implementation project.For example,an agent employed primarily within a business application is likely to be preconfigured at least to some extent to be able to call upon a range of func

39、tions specific to that environment.These functions could perform basic tasks such as creating a case,creating a contact,or opening an existing case/contact,but perhaps within larger-scale operations such as updating multiple cases or accounts in a single request.These actions call upon functions tha

40、t already exist but might be more convenient to operate through the medium of the agent.The action itself has characteristics that help the reasoning engine determine the right call(s)as it assembles its plan for the response.System administrators set elements like input and output descriptions for

41、the action,user authentication permitted to call the action,what the context might be within a broader set of operations,and the likely response format.These operating parameters are often referred to as part of the“guardrails”that maintain the agents safe operation.What separates agents from previo

42、us generations of“chatbots”is the use of the reasoning engine to determine the plan dynamically.Chatbots tended to require tightly scripted decision trees to operate,with simple keyword triggers to move between nodes to reach predetermined end points.This means that their safe operation is a factor

43、of their Figure 4How Actions Might Be Collated Within the Administration of an Agent System*Triggers and responses help determine whether this is the right action to attribute to part or all of a request.limitations.Reasoning engines have a much looser set of predeterminations,which gives them more

44、flexibility but with the potential for a greater range of problematic outcomes if not well-marshaled.The use of actions to move along approved paths both designed by application providers and specified by customers is a fundamental part of how agents improve specificity.Action typesActions do not ha

45、ve to be specific to the business application in which they operate.So far,weve only discussed very basic localized outcomes for actions,but in essence they can potentially address any function or system they have permission to communicate with.The following are some example groupings for common act

46、ions.Trigger Response WorkflowTrigger Response API CallTrigger Response URL/EndpointTrigger Response External AutomationDescriptionRoleContext ResultOutput of Action*Described Action6AI Agents:What They Are,How They Work,and Where Organizations Might Best Use Them 2024|infodeep-Deep AnalysisBusiness

47、 application workflowsThe business application environment is likely to involve enacting workflows:both those that might be supplied out of the box by software vendors and those that have been customer-developed within implementations.This enables the reasoning engine to call a workflow,wait for an

48、expected response successful or otherwise and pass that information to the agent.Given that the agent/reasoning engine dialogue can generally manage state to some degree(remembering context both for a current session and potentially from previous sessions),this means that a form of exception handlin

49、g or process clarification could be possible.For example,where a workflow returns a“not found”reference,the agent can be prompted to ask the requester for clarification before replaying the request with an updated reference.Remembering that actions can be stacked up within an individual plan,this is

50、 one instance where process orchestration using agents is possible:workflows executed in order with dependencies between them the output of one being the input operator for the next handled by the reasoning engine.API/URL endpoint callsBeyond the business application itself,actions can be created th

51、at make calls to specific locations to collect information or perform specific actions(see Figure 5).For example,if defined sources of information are commonly requested within your organization or by your customers,then plugging that information into an action is a straightforward way to ensure qui

52、ck handling of that sort of request.An example is URLs:location,opening hours,contact details,facilities information,etc.Figure 5Simplified Example of a Request and Plan Set with Actions Going to Various SystemsThis method also allows for querying basic states such as order status or due dates by UR

53、L or API calls.In isolation,these can seem facile,however they might form a significant part of the overhead for some customer-or employee-facing teams,and providing a reliable and specific way to deflect the request away from them can free up resources.Also,remembering that actions can be stacked w

54、ithin a plan,the result of a state check might be critical to the next action within a broader plan,especially if its conditional(“if order not dispatched,then trigger order escalation”).Request:Make me an order from the best-selling items that are in stock,with descriptions.Reasoning:I need to chec

55、k sales,then stock,then make a summary of it all.Guardrails:Are these approved actions?Plan*:1.Make list of best-selling items order management2.Check list against available stock inventory management3.Summarize items into a nice description LLM summarizer4.Return to agent,with checkpoint for reques

56、ter to approve(possible return loops here to make changes to suggested order)5.If approved,send to cart to create order commerce*Note:plan does not include payment and delivery actions.7AI Agents:What They Are,How They Work,and Where Organizations Might Best Use Them 2024|infodeep-Deep AnalysisExter

57、nal automations/business processesMuch as enacting workflows within a business application reveals a potentially powerful form of process orchestration,enacting workflows outside the business application extends that potential.For the sake of convenience,we have focused on agents that might sit with

58、in a business application or common application suite.Organizations that are solely operating within such an environment will use that application set as their“run time environment.”Most organizations,however,are more heterogeneous and have multiple core applications where work takes place;perhaps a

59、 specific business process engine along with that provides a centralized run time to coordinate all the requisite parts.Other run times like within BPM,RPA,or low/no-code development platforms are additional external processes that might be triggered.Agents ideally will be able to kick off,terminate

60、,or query instances of workflows or automations within these environments through API calls as deemed necessary or appropriate.Some of these are likely to be relatively trivial:checking the status of a workflow instance by case ID,customer ID,or data range,for example.The use of agents becomes more

61、complex where transaction states might want to be changed:for example,where an order might need to be changed or canceled entirely.Ensuring authentication is one important element,but providing access to an agent so it can act based on an interpretation of intent might trouble some application owner

62、s.Here,it is likely about building confirmations and recording mechanisms within those actions.What started as AI becomes far more about process mapping and modeling at this stage.Use Cases for AgentsMuch has been made of the potential business impact of agents since they began emerging among AI pro

63、duct announcements during 2024.Sweeping predictions of transformative technology can be compelling,but its use has to be wedded to the day-to-day realities of an organizations ways of working.In practice,this means finding ways that agents can help introduce or improve tasks and processes that contr

64、ibute to their operations.This is especially important as agents are bringing with them a change in how this software is paid for,ensuring that the transactional impact of agents will be even more keenly observed,measured,and attributed.We will come back to this in the“Planning for Agent Projects an

65、d Deployment”section.As can be imagined from the technological capabilities described,the potential use cases for deploying agents are large in number and broad in scope,ranging from simple information to complex multi-stage transactions.To clarify this for better understanding,this section of the r

66、eport does two things:Creates a set of broad classifications for agent use cases,to demonstrate Ensuring authentication is one important element,but providing access to an agent so it can act based on an interpretation of intent might trouble some application owners.8AI Agents:What They Are,How They

67、 Work,and Where Organizations Might Best Use Them 2024|infodeep-Deep Analysisthe complexity of their planning,implementation,and operation.Outlines some of the sets of use cases that software vendors are proposing for agents,using the classifications to indicate the likely scale of work involved in

68、introducing them to everyday use.Note that this is a general,indicative guide and is not intended to support any particular vendors approach to their agent products.The guide is as broad as possible in encapsulating the current market and its suggested use cases for agent adoption.It is important to

69、 remember that agents are cumulative of the general functionality of assistants,which may be mentioned in discussions of specific use cases below.Assistants capabilities in use cases are discussed in our report“Workplace AI Market Analysis:Generative AI and the Desktop (R)Evolution.”3 Use case categ

70、orizationTo produce a straightforward,easily articulated classification of agent use cases,we use three groupings:Information,Transaction,and Extended.These are not especially scientific but provide a quick way to determine like complexity and time to value,and therefore the basics of an indicative

71、cost and value calculation(see Figure 6).Naturally,an organization can only estimate those calculations,as the operational impact on existing ways of working is likely to be specific to each.Figure 6Relative Time to Complexity and Time to Value of the Three Use Case ClassificationsIn this section,we

72、 suggest an illustrative but generic agent use case for each of the three classifications.We chose to look at task-level use cases rather than process-level,because this should help attribute the use cases more clearly against the action(s)needed to fulfill them.It is likely that complete use cases(

73、complete interaction maps,etc.)will be strung together,much as a reasoning engine devises a plan containing multiple actions.Each interaction is likely to contain straightforward and complex individual elements,which will translate into similarly varied sets of actions.We use general commerce for th

74、ese use cases because most people will be familiar with at least the requesting side of the transaction,if not the actioning side.It is important to remember that agents are cumulative of the general functionality of assistants.Information(read)Transaction(create)Extended(update)ComplexityTime to va

75、lue9AI Agents:What They Are,How They Work,and Where Organizations Might Best Use Them 2024|infodeep-Deep AnalysisInformationExample:Where is my order?The most straightforward classification,Information encapsulates broad swathes of use cases that are designed to deflect simple questions away from hu

76、man operators.Previously,they were primary candidates for the prior generation of chatbots and before that,search technology.That those previous generations of technology did not deflect information requests permanently perhaps suggests that the requests were not as simple as they appeared.In our“Wh

77、ere is my order?”example,detecting the initial intent and creating a plan of action should not be too difficult,especially as its likely such a dialogue flow already has been mapped for existing chatbots:confirming the request,asking for order identification,confirming basic order details,returning

78、current order status.These are likely existing,well-used informational end points on an API that,once returned,can remain in state for the conversation until it is concluded.The actions that likely need to be configured or developed are those to query whichever line-of-business application contains

79、the order information(e.g.,commerce or other order management platform).LLM functions like summarization can be used to format the output into something more conversational to be returned to the agent for response.Information use cases are likely to contribute many of the interactions collected with

80、in a research exercise examining existing channels(call center,chatbot,search logs).Understanding the current cost of fulfilling these requests versus the potential cost of deploying agent actions is recommended(see the section“Planning for Agent Projects and Deployment”).TransactionExample:Update m

81、y order.Transaction use cases are a step more complicated than Information ones,as they require external systems to be not just read,but also updated and confirmed.A Transaction use case might be strung together with an Information use case within a single interaction,as a dependent outcome(order de

82、layed,change contents of order,or cancel).This takes the interaction out of the scope of existing chatbots and into higher-value call center deflection,significantly increasing the complexity of the actions needed for a resolution plan.Update actions here must be able to interact with line-of-busine

83、ss systems so that they can manipulate all the elements of in this case the order(although it could equally be a case,invoice,or other business object).This means that the commerce system has to be extensible to work with the agent and that the agent needs permissions to take executive control of th

84、e order,albeit with confirmation steps from the requester.That way the agent can generate conformational elements like updated emails or invoices for the order as would happen natively within the commerce system.Understanding the current cost of fulfilling these requests versus the potential cost of

85、 deploying agent actions is recommended.10AI Agents:What They Are,How They Work,and Where Organizations Might Best Use Them 2024|infodeep-Deep AnalysisOne of the common frustrations with chatbots over the years has been that at critical points within an interaction,dialogues tended to require escala

86、tion to a call center or other channel.For Transaction use cases to be successful and for agents to be valuable,they need to be able to execute completely.If any of the required steps for an update cannot be completed by the agent through actions to the commerce platform,then the value of the entire

87、 use case is likely significantly reduced.ExtendedExample:Create an order.Extended use cases take the steps in Information and Transaction cases and build an additional magnitude of complexity within them.If we see Information as a“read”case and Transaction as an“update”case,Extended is a“create”cas

88、e where were adding at least the following into our required set of actions for our commerce example:product search,selection and quantity determination,discounting and couponing,and other order-specific elements.A potential simplified case would be commonly repeated orders requiring replication of

89、an existing order.Its essential to be very clear that a use case like this should solve an issue such as by deflecting inquiries from human operators in volume and not create and support actions that directly replicate almost an entire commerce system.It is easy to imagine cases for agents in this a

90、rea,but much harder to imagine the business case for developing them unless the agent is designed to be operated natively within the commerce system itself as a discrete entity.The closer requirements become to full Extended use cases,the more complex they are likely to be:bespoke and tied to specif

91、ic business applications(and therefore highly specific,thinking back to the chart in Figure 2).Use case examples(with categorization)We are in the very early stages of understanding the likely operational sweet spot for agents within organizations.This is partly because the cycle for assistants star

92、ted just 12 months ago and on limited real implementation experience,without any long-term outputs so its difficult to identify the shortfalls in functionality that agents can best fulfill.This means software vendors that are investing heavily in making agents the centerpiece of their current go-to-

93、market are having to cite use cases that they know exist and are marketable,rather than those which are applicable and deployable.As we noted when discussing the Extended use case classification,in the most complex case,an agent moves from a point of process orchestration to actually performing the

94、role of executing the user experience for a single application.The latter is,of course,valid but raises the potential of an upcoming generation of application-siloed agents needing mediation by executive agents,which for now is perhaps far too meta to be examined further(even if it can be imagined).

95、This is distinct from managing instances of agents within the same environment,utilizing the same set of actions.11AI Agents:What They Are,How They Work,and Where Organizations Might Best Use Them 2024|infodeep-Deep AnalysisIn this case,administration and management interfaces would be expected to s

96、cale as agent demand pushes the capabilities.In looking across a wide diversity of business application vendors to see the use cases they are explicitly lining up for their current interpretation of agents,we see a range from the exceptionally broad(“knowledge management,”“employee hiring,”“customer

97、 success representative”)to the recognizably specific(“order modification,”“warranty enquiries”).As can be imagined,only a subsection of these can be neatly categorized using our previously outlined use case classification.For example:Information:order status and tracking,product availability and sp

98、ecifications,product set-up and troubleshooting,membership inquiries,maintenance troubleshooting,facilities information,appointment checking,contracts research,building and maintenance operating procedures Transaction:order modifications,billing and payment support,returns,exchanges and refunds,shif

99、t scheduling,appointment adjustment Extended:claims and warranty,appointment scheduling,shipment management,inventory management,claims reportingPlanning for Agent Projects and DeploymentThe emergence of agents as a potential avenue for organizations to apply AI to process orchestration has been so

100、recent that no specific precedents exist for planning and deploying the technology.Our previous report,“Workplace AI Market Analysis:Generative AI and the Desktop (R)Evolution,”4 contains a section titled“Managing the(R)Evolution”that outlines best practices gleaned from both AI and process analysis

101、 projects that formed the basis of an approach to assistant projects.The sections“Testing and Evaluation”and“Delivery and Maintenance”offer a basic primer for agent project considerations.Additional elements can be considered addendums to the section“Planning and Participation,”specifically to the s

102、ub-sections“Identification and Analysis”and“Workforce Participation.”This section describes those elements and discusses cost planning specifically for agents.Planning and participationThe use of agents presents challenges beyond those for assistants.Assistants primarily deal with incremental,single

103、-step,task-based activities,and the concern is whether those functional increments are beneficial.With agents,the question becomes whether to transpose entire processes and build them on a task-by-task basis as actions,so that they can be dynamically reconstituted as versions of those processes base

104、d on a reasoning engines interpretations.This effectively allows the creation of a process instance that operates the process that has just been created(albeit from approved building blocks)and terminates it,potentially without any human intervention.12AI Agents:What They Are,How They Work,and Where

105、 Organizations Might Best Use Them 2024|infodeep-Deep AnalysisIdentification and analysis“Managing the(R)Evolution,”focuses heavily on the use of task mining to determine how well the skills an assistant could bring to the organization matched up with workers current ways of operating.With agents,we

106、re rushing from a broadly task-based to a process-based analysis,so process mining will provide a heap of additional indicative information to help guide the identification and capture of building blocks to turn into approved actions.Generically,process mining uses transaction data from business app

107、lications to build analysis of how process instances actually operate,revealing things like degrees of iteration/correction/intervention within process instances,time spent on specific tasks within those instances,and other insights.It identifies the critical paths and critical tasks within processe

108、s based on real transactional data and combines that with the processes operation frequency.As weve moved from the desktop to the application and from assistants to agents,process mining will add a significant amount of useful data for planning what actions to build.Workforce participationAs discuss

109、ed in“Managing the(R)Evolution,”systems like task and process mining are vitally important in understanding the contributions of applications to a current and future state of operation,but they do not deliver complete insight into how the organization functions practically.The workforce should provi

110、de a vital expert voice when selecting use cases for consideration,mapping the tasks and processes that an agent will need to be cognizant of to operate,and ensuring that the full range of exceptions is accounted for(including all the common mitigations).Cost planningProcess mining will add another

111、important source of planning data:the cost implication of using agents.This is touched on in“Managing the(R)Evolution,”where we imagined metered use of generative AI was coming.As part of the shift to agents,it has certainly arrived;specifically,consumptive pricing along with some focus on the outco

112、me of a transaction.Now,it is important to know not just how many instances of a process might take place,but also the frequency of the specific events that the agent will count as a successful outcome,which might not be the same.Where those events might not be specifically recorded within a traditi

113、onal line-of-business application,other considerations apply.For example,where a human clicking“end conversation”within a chat is considered a billable outcome,understanding how frequently that might occur could involve looking at current data retained by existing chatbot applications,rather than th

114、e transaction logs of a CRM or commerce system.As weve moved from the desktop to the application and from assistants to agents,process mining will add a significant amount of useful data for planning what actions to build.13AI Agents:What They Are,How They Work,and Where Organizations Might Best Use

115、 Them 2024|infodeep-Deep AnalysisOutlook(12-24 Months):“What Is Our Agent Strategy?”Providing an outlook for agents over only 12-24 months might seem unambitious until you recall that ChatGPT was first available for use in November 2022 24 months ago at the time of this reports writing.The short(12

116、months)and medium(24 months)terms are designed to offer practical support for those being asked“what is our agent strategy?”as many will imminently,if they have not already.In the short term,most agents deployment will focus on utilizing preconfigured sets of actions designed to primarily meet broad

117、,horizontal use cases(general line-of-business,rather than industry specific).This repurposing of existing workflows and application-specific code will predominate within business-application-facing agents specifically for customers heavily committed to single-vendor walled gardens(whether the agent

118、 itself and its use cases are internal or external facing).Again,this sounds unadventurous given the big promises software vendors are making publicly,but it will be vital in stress-testing those vendors agent tooling within trusted,existing processes where they have control over the surrounding tec

119、hnology stack.In short,the focus will be vendors outfitting the biggest and most loyal customers,to provide those first big case studies that will prove agents successful.In the medium term,the focus will broaden to vendor agent platforms providing stable and robust action*libraries.Initially,these

120、will ensure that assembling agents can easily access native functions and workflows,add support for common third-party applications,and then enable easy assembly and configuration of customer-defined actions.The focus will be on decreasing the time-to-value as much as possible so the consumption or

121、outcome events can start ringing the cash register as rapidly as possible.This onboarding will almost certainly involve employing an agent to help define and build such customer-defined actions;agents will perform largely the same task as wizards once did.Just on the horizon out of scope in the shor

122、t term but coming into view for the medium term are large action models.These are based on sets of interactions proposed to string together actions into recognizable task streams in the same way LLMs do with language.Working alongside LLMs in the next iteration of agent architectures,these new model

123、s could bring agents closer to the advertised autonomy that the current generation is not equipped to deliver.*Authors noteI wish we could call them skills;wouldnt that be better?Outfitting agents with skillsets would be a better way to explain all of this.Its a shame that term got associated with a

124、 logically similar way of adding capabilities to the generation of physical assistants(Alexa,Google Home,etc.)that now are at best gathering dust in homes,if not turning up in landfill.In the short term,most agents deployment will focus on utilizing preconfigured sets of actions designed to primaril

125、y meet broad,horizontal use cases(general line-of-business,rather than industry specific).14AI Agents:What They Are,How They Work,and Where Organizations Might Best Use Them 2024|infodeep-Deep AnalysisCall to Action VendorsFor software vendors looking to propose agents to both their current install

126、base and prospective customers,it is vitally important to provide simple on-ramps and realistic onward paths towards their use.This means offering clear advice on how organizations can get ready for agents,plan their use cases and initial projects,and develop internal roadmaps to determine their med

127、ium-term direction.This advice will be in addition to that already in place around the development of assistants(such as for data management).As attractive as it might be to pitch ideas around autonomous agents performing extended use cases,this is unlikely to be the reality for even the best-prepar

128、ed organization,even beyond the medium-term,and offers a glimpse into a vision that is closer to science fiction than a projectable future state.Moving beyond what is practically possible does no favors to vendors or customers;what is needed is the realization of projects that deliver the praxis nec

129、essary to bring pragmatic customers on board and grow general agent adoption.BuyersFor software buyers,matching agent capabilities to beneficial outcomes is of paramount importance.Ensuring that the use cases for agents are backed with research that determines what contribution and impact the advert

130、ised process orchestration would realize in their context is the primary pre-project goal.In short,“are agents useful to us?”The second important question to answer is“are we ready for agents?”Further,“do we have the operational maturity in our tasks and processes and the robust workflows required t

131、o make their orchestration even possible?”And,“have we identified specific use cases where agents could take on sets of tasks that modeling has shown will have measurable impact by deflecting tasks from humans?Answering these sorts of questions along with those around internal data management and da

132、ta quality that may already have been addressed for assistant use cases but also matter significantly here will help answer the question that buyers may be asked:“what is our agent strategy?”Endnotes1 https:/www.deep- https:/arxiv.org/abs/2005.114013 https:/www.deep- https:/www.deep- Agents:What The

133、y Are,How They Work,and Where Organizations Might Best Use Them 2024|infodeep-Deep AnalysisDeep AnalysisAbout Deep Analysis Deep Analysis is an advisory firm that helps technology vendors,buyers,and investors understand and address the challenges of innovative and disruptive technologies in the ente

134、rprise software marketplace.The firms work is built on decades of experience advising and consulting to global technology firms large and small,from SAP,Oracle,and HP to countless start-ups.Led by Alan Pelz-Sharpe,Deep Analysis works with technology vendors,buyers of enterprise technology,and invest

135、ors in the ECM and enterprise automation market to improve their understanding of the information management technology landscape and provide actionable guidance.Deep Analysis timely book,“Practical Artificial Intelligence:An Enterprise Playbook,”outlines strategies for organizations to avoid pitfalls and successfully deploy AI.Unstructured Data ManagementECM,KM,EX,etc.Market&Product StrategyProcess&Task AutomationIntelligent Document Processing(IDP)Thought LeadershipCompetitive IntelligenceDue DiligenceContact us:infodeep-+1 978 877 7915We Research InnovationOur Services