Situation: AI chatbot with the requirement to minimise response time (ie Speed) and maximise customer satisfaction (ie Quality).

Here the trade-off is between 'Speed of responses' and 'Quality of responses'. Faster responses would also increase customer satisfaction as more queries can be handled in less time. On the other hand, detailed personalised responses may take more time but customer satisfaction could be high due to better resolutions generated from the bot.

 

Strategy for balancing objectives:

1) Instruct bot to send acknowledgement to customer query within 20 seconds

eg: "Dear 'customer', i understand you have raised a query related to order #abc123. I am looking into this now"

 

This would keep the customer 'warm' and he/she would be assured that the bot has started working on the query.

 

2) Set rule for bot to fetch all details related to the order and respond to specific query within next 30 seconds

eg: "I have checked your order and I can see that your order contains two items - item A would be delivered to your door step by xyz date, which is before time delivery as you can see. item B is getting slightly delayed due to cross border customs verification process. Apologies for this! Allow me sometime to escalate this issue and get you a resolution"

 

This would satisfy the customer to some extent, since part of the information/status is being shared within 1minute, and the bot is working on resolving the rest of the issue.

 

3) Set rule for bot to escalate to human agent and inform customer regarding the same - set time limit as 60 seconds

eg: "Thankyou for your patience and I apologise again for the delay in delivery of item B. Issue related to item B has now been escalated to a human agent, and he/she would get back to you in the next few hours. Hope i was able to help you with your query resolution."

 

This stage ensures that the escalation is also done by the bot and a human agent would work on the resolution, since this is now a complex issue to be handled appropriately.

 

This strategy addresses both 'speed' and 'quality' as the bot is operating in stages, and buying itself some time to resolve slightly complex queries. In doing so, the bot is able to construct the 'personalised' resolution that the customer is looking for, and also proactively actioning the escalation to human agent for complex issues.

In artificial intelligence domain, outcome of any question (process) depends on the input variables (existing references available datasets) and the algorithms (program) which is developed on available data sets.

Incase of any goals clashes, firstly we need to retrospect the input data set for its accuracy, authenticity and purpose.

We need to search within the developed algorithm (program) for any errors (e.g. syntax, semantics etc.), or any undesirable clause which can modify or influence the outcome.

It may also happen that based on the data, AI agent is exploring new possibilities which has not been heard off and is completely new. 

Based on thorough human assessment, we can decide the actual outcome.  If it still clashes, we can modify the algorithm (through advance discriminator) to achieve the target goal.

 

 

Per my understanding AI would face this kind of trade off in most of the service industries where the AI agent need to handle 2 or more objectives such as time bound, accurate, precise, customer delight, etc. Taking example of Customer Service in E-commerce..

 

Scenario : Customer queries are been handled by an AI Agent. Here the AI needs to manage response time with accurate and precise replies to queries regarding orders, deliveries, return, refund, etc.

 

How you would guide the AI to handle the trade-off –

1. Identify the Issue/query type : AI should first recognize the type of query if it’s an order status, return, refund or just an information.

2. Assess the urgency : Basis the query and customer responses assess the urgency attached to the query

3. Assess the complexity: such as order status can be simpler to handle whereas lost or refund may need additional information or time to reply.

 

What logic, rules, or signals would help it make the right call?

 

1. Rules: Set the priority rules such as simpler queries like order status to prioritize for quicker reply over the queries like lost items or refund may need additional information or time to reply

2.Logic: For the prioritized queries response may be shorter with predetermined pattern of text such as delivery time, tracking or details of delivery agent etc.  For queries requiring longer time to add follow up question to gain more understanding and give more detailed and precise reply including adding steps needed by customers to take

3. Signals: Customer Response (Important to assess what the customer says during the conversation. Basis those sentiments replies may need changes) and Response Time (Overall response time to be checked to ensure it gives quicker responses per the urgency)

• Raise the conflicts.
• Discuss it and take a piece of advice from the team users.
• Apply the strict rules.

When AI employees work in customer service, they often need to find a way to be quick and kind at the same time. While strictly adhering to process rules can expedite responses, it may overlook individual customer needs, potentially leading to customer dissatisfaction. 

Consider customer delight full first and this are the one of our six values of organization, on the other hand, might mean personalized exchanges that are different from what is usually done, which could slow things down. To help the AI handle we need to follow below approach:
Organized success factors that measure how well you're doing in terms of response time and how satisfied your customers are like customer sentiment analysis score. In accordance the AI can judge acts by how they affect both. Next is setting up escalation procedures means teaching the AI to recognize when a person needs to step in, like when a customer's body language shows they are dissatisfied. This approach makes sure that hard problems get the attention they deserve, even if they are difficult to understand or deal with. If we take one example a customer gives an NPS score of 3 and says, "The billing system is confusing, and last month I was charged too much." Score (NPS): Identify as a Detractor (scores from 0 to 6). Feedback in full: Words such as "confusing," "overcharged," and "last month" indicate a specific financial issue. Prioritize immediate communication: It's crucial to address feedback promptly. Communicate with the customer in a manner that suits their preferences. For example, "We're so sorry to hear about the problem that you had with our billing system and the extra charge you got last month." We understand how upset you are. If possible, please provide more details so we can quickly investigate and resolve the issue. Focus on solving the problem instead of advocating right away. The main goal is to resolve the problem and rebuild trust, not to get a better NPS score right away. Possible to escalate to a human agent: Because the problem is a billing error, the AI may escalate the case to a billing expert who can fix the overcharge and explain how the billing system works. 

Sticking to the process:

One situation where we need AI to take a trade off is speed vs accuracy. AI must quickly decide whether the content can be moderated, meaning deciding where human validation is required. 

 

Logics/Factors to be considered:

1. Determine the severity of the content / Modification requested

2. Confidence level of the content based on parameters that are available

3. User reputation scoring. Determine the score of the user who proposed the change/content creator

4. Time constraints for human validation

On the basis of above a quick scoring mechanism needs to be built which will be the decision making logic for the content & the content provider. 

 

Rules for AI to follow:

On the basis of the above logics, thresholds have to be created for where AI can take a call directly and let the change/content go live.

 

For cases where the threshold is passed, then it needs to be passed on for human validation. 

 

 

 

 

 

 

 

 

Where the AI agent must balance 2 or more competing objectives, accurate and helpful whilst ensuring the customer feels supported and valued requires 3 elements to be in place: Logic; Rules and Signals.

 

One such scenario would be when employing a Customer Support AI Chat Bot.

To guide the AI agent, we can use a combination of logic, rules, and signals. Here's a possible approach:

 

Logic:

1. Weighted scoring: Assign weights to different aspects of the response, such as accuracy, completeness, and relevance. The weights can be adjusted based on the customer's preferences and priorities.

2. Threshold-based evaluation: Set thresholds for response time and customer satisfaction. If the response time exceeds the threshold, the agent can adjust its response to prioritize speed over accuracy or completeness.

3. Contextual understanding: Use natural language processing (NLP) and machine learning algorithms to understand the customer's context, tone, and language. This can help the agent tailor its response to the customer's needs and preferences.

 

Rules:

1.*Response time tiers: Establish response time tiers, such as:

    - Tier 1: Quick responses (e.g., < 1 minute) for simple queries.

    - Tier 2: Standard responses (e.g., 1-5 minutes) for more complex queries.

    - Tier 3: In-depth responses (e.g., > 5 minutes) for critical or complex issues.

2. Customer segmentation: Segment customers based on their preferences, behavior, and value to the organization. For example:

    - VIP customers may require more personalized and rapid responses.

    - High-value customers may require more comprehensive and accurate responses.

 

Signals:

1. Customer feedback: Collect feedback from customers on their satisfaction with the responses. This can help the agent adjust its weights, thresholds, and rules to better meet customer needs.

2. Agent performance metrics: Track metrics such as response time, accuracy, and completeness. This can help identify areas for improvement and optimize the agent's performance.

3. Contextual signals: Use contextual signals, such as the customer's location, device, or previous interactions, to inform the agent's responses and improve customer satisfaction.

 

Weightage:

·        Accuracy = 0.4

·        Completeness = 0.3

·        Relevance = 0.2

·        Response Time = 0.1

My thoughts on this question - 

This is a tricky question to answer because once you give a set of instructions to AI, it will give a response after optimizing for the clashing goals. So purely from an AI perspective there is no clash. The clash here refers to more between the AI output and human expectation. The best way to handle it is keep revising the instructions, the logic, the rules, the knowledge base and continue to re-train the agent. And wherever we feel that the AI output is still not in sync with our understanding, AI should treat it as an exception and let a human handle it.

 

There are some must read answers - Swapnil Madhav Chaukar, Pratish Deshpande, Swarandeep Kaur Juneja, Mohammed Jaffer, Amit Kumar.

The best answer to this question is written by Mohammed Jaffer. Well done!