12 answers
CareerVillage Office Hours
Student
Palo Alto, California
589
Questions
Asked
779 views
Can you share an example of how your team overcame an AI-related challenge through human ingenuity?
Can you share an example of how your team overcame an AI-related challenge through human ingenuity?
Login to comment
12 answers
PwC Impact
884
Answers
Orlando, Florida
Updated
PwC’s Answer
One experience that stands out was when my team built a Custom GPT for a PDR assistant to generate templates. While it was a great starting point, we noticed the outputs were not always accurate or aligned with what was needed. That pause pushed us to think more critically about how the tool was being used. We analyzed the errors, mapped out what needed to change, and added a mandatory layer of quality checks. We are still refining it, and the process has reinforced an important lesson that AI can support the work, but it takes human judgment, curiosity, and hands-on review to make it truly effective.
Our team was asked to deliver a three‑year, year‑over‑year view of the financial and resource investment required to stand up a major technology platform. The challenge was that the data lived in multiple places, the audience was senior partners, and the turnaround time had to be fast and highly accurate.
Our manager structured the team intentionally. Four of us were assigned, two focused on building the presentation shell using AI tools, and two focused on pulling together and validating the data. Each role had a backup so the work could continue if someone was unavailable. At peak, we were working twelve‑hour days, six days a week.
AI helped us move quickly on structure and draft content, but it could not efficiently combine complex financial data from multiple systems or produce reliable charts. That work required human judgment, so the data consolidation and chart creation were done manually in Excel. Once the charts were complete, AI was used again to help generate slide text based on those visuals.
Our manager provided the storyboard, meaning the narrative, the required views, and the flow of the presentation were clearly defined upfront. This human guidance was critical. While AI generated conclusions, those conclusions were not always accurate and needed to be reviewed, challenged, and refined by the team to ensure they truly reflected the data and the business reality.
Over the four‑week effort, the team had to adapt continuously. One person was unexpectedly out for two weeks, another could not work weekends, and another was unavailable during the final week. Because of the redundancy built into the team and strong human collaboration, the work still came together.
In the end, six people contributed to delivering a partner‑ready presentation that summarized a forty‑million‑dollar program. AI accelerated parts of the process, but human ingenuity, planning, judgment, and teamwork were what ultimately made the outcome successful.
Our team was asked to deliver a three‑year, year‑over‑year view of the financial and resource investment required to stand up a major technology platform. The challenge was that the data lived in multiple places, the audience was senior partners, and the turnaround time had to be fast and highly accurate.
Our manager structured the team intentionally. Four of us were assigned, two focused on building the presentation shell using AI tools, and two focused on pulling together and validating the data. Each role had a backup so the work could continue if someone was unavailable. At peak, we were working twelve‑hour days, six days a week.
AI helped us move quickly on structure and draft content, but it could not efficiently combine complex financial data from multiple systems or produce reliable charts. That work required human judgment, so the data consolidation and chart creation were done manually in Excel. Once the charts were complete, AI was used again to help generate slide text based on those visuals.
Our manager provided the storyboard, meaning the narrative, the required views, and the flow of the presentation were clearly defined upfront. This human guidance was critical. While AI generated conclusions, those conclusions were not always accurate and needed to be reviewed, challenged, and refined by the team to ensure they truly reflected the data and the business reality.
Over the four‑week effort, the team had to adapt continuously. One person was unexpectedly out for two weeks, another could not work weekends, and another was unavailable during the final week. Because of the redundancy built into the team and strong human collaboration, the work still came together.
In the end, six people contributed to delivering a partner‑ready presentation that summarized a forty‑million‑dollar program. AI accelerated parts of the process, but human ingenuity, planning, judgment, and teamwork were what ultimately made the outcome successful.
PwC Impact
884
Answers
Orlando, Florida
Updated
PwC’s Answer
AI should be leveraged but not trusted entirely. An example that my team overcame an AI-related challenge was when the AI was not giving us valuable metrics when we asked the model to evaluate the effectiveness of applications based on a rubric that we fed it. We realized that there were certain inputs and factors that could only be evaluated by human ingenuity.
Challenge AI via prompts and example to provide you with better analysis.
For example, evaluating AI-generated insights to ensure they align with team's overall goals. My team worked on creating a Custom GPT that determines Project Q/NQs for R&D Credit studies. It wasn't always correct, so human review was needed as well as training the model to 'think the way a tax professional would think'
Challenge AI via prompts and example to provide you with better analysis.
For example, evaluating AI-generated insights to ensure they align with team's overall goals. My team worked on creating a Custom GPT that determines Project Q/NQs for R&D Credit studies. It wasn't always correct, so human review was needed as well as training the model to 'think the way a tax professional would think'
Priyanka K
Software Engineer
29
Answers
Little Elm, Texas
Updated
Priyanka’s Answer
Certainly! Let’s explore an example of how human ingenuity can solve challenges posed by AI systems—this can apply in various real-world scenarios. Here’s a hypothetical but realistic example based on common experiences with AI deployment:
Scenario: AI Misinterprets Customer Sentiment in a Chatbot
Challenge:
An AI-powered chatbot designed to handle customer service inquiries was frequently misinterpreting customer sentiment. For instance, when a customer typed, "I'm so thrilled to finally get this resolved," the chatbot would incorrectly classify it as a negative sentiment because the sentence contained the word "resolved," which was often paired with complaints in the training data. As a result, the chatbot would escalate the issue unnecessarily to a human agent.
Human Ingenuity to the Rescue:
The team combined human observation, creativity, and problem-solving to address the problem:
Step 1: Identifying the Root Cause
A cross-functional team of engineers, linguists, and customer service experts manually reviewed conversations flagged as problematic by the chatbot. They noticed that the sentiment analysis AI was overly reliant on keywords without fully understanding context or tone.
Insight: Humans realized that the issue wasn’t just about training data but about the need for the AI to interpret combinations of words and punctuation with nuance (e.g., "thrilled" + "finally" is positive).
Step 2: Designing a Hybrid Approach
Instead of relying solely on AI, the team implemented a hybrid model:
They added a secondary layer of human oversight for flagged conversations.
A team of linguistics experts augmented the training data with edge cases, including diverse sentence structures, emojis, and punctuation styles.
The chatbot was also programmed to ask clarifying questions in ambiguous situations, like: “It sounds like you’re happy with the resolution—did I get that right?”
Step 3: Leveraging Creativity in Training
The team used creativity to simulate real-world conversations:
They designed mock chats with varied emotional tones.
They incorporated slang, cultural phrases, and regional nuances, which AI often struggles with.
They ran workshops where customer service representatives role-played interactions to create new training data.
Step 4: Continuous Feedback Loop
To ensure the problem didn’t recur, they implemented a feedback mechanism:
Human agents flagged future cases where the AI misinterpreted sentiment, feeding this data back into the system for retraining.
The Outcome:
The chatbot’s sentiment analysis accuracy increased by over 90%.
Escalations to human agents dropped by 40%, saving time and resources.
Customer satisfaction scores improved as the chatbot responses became more empathetic and contextually accurate.
Key Takeaway:
This example highlights that while AI is powerful, it’s not infallible. Human ingenuity—through observation, collaboration, and creative problem-solving—can bridge gaps in AI systems. It’s a reminder that AI works best as a tool to augment human intelligence, not replace it. Together, humans and AI can achieve smarter, more nuanced solutions.
Scenario: AI Misinterprets Customer Sentiment in a Chatbot
Challenge:
An AI-powered chatbot designed to handle customer service inquiries was frequently misinterpreting customer sentiment. For instance, when a customer typed, "I'm so thrilled to finally get this resolved," the chatbot would incorrectly classify it as a negative sentiment because the sentence contained the word "resolved," which was often paired with complaints in the training data. As a result, the chatbot would escalate the issue unnecessarily to a human agent.
Human Ingenuity to the Rescue:
The team combined human observation, creativity, and problem-solving to address the problem:
Step 1: Identifying the Root Cause
A cross-functional team of engineers, linguists, and customer service experts manually reviewed conversations flagged as problematic by the chatbot. They noticed that the sentiment analysis AI was overly reliant on keywords without fully understanding context or tone.
Insight: Humans realized that the issue wasn’t just about training data but about the need for the AI to interpret combinations of words and punctuation with nuance (e.g., "thrilled" + "finally" is positive).
Step 2: Designing a Hybrid Approach
Instead of relying solely on AI, the team implemented a hybrid model:
They added a secondary layer of human oversight for flagged conversations.
A team of linguistics experts augmented the training data with edge cases, including diverse sentence structures, emojis, and punctuation styles.
The chatbot was also programmed to ask clarifying questions in ambiguous situations, like: “It sounds like you’re happy with the resolution—did I get that right?”
Step 3: Leveraging Creativity in Training
The team used creativity to simulate real-world conversations:
They designed mock chats with varied emotional tones.
They incorporated slang, cultural phrases, and regional nuances, which AI often struggles with.
They ran workshops where customer service representatives role-played interactions to create new training data.
Step 4: Continuous Feedback Loop
To ensure the problem didn’t recur, they implemented a feedback mechanism:
Human agents flagged future cases where the AI misinterpreted sentiment, feeding this data back into the system for retraining.
The Outcome:
The chatbot’s sentiment analysis accuracy increased by over 90%.
Escalations to human agents dropped by 40%, saving time and resources.
Customer satisfaction scores improved as the chatbot responses became more empathetic and contextually accurate.
Key Takeaway:
This example highlights that while AI is powerful, it’s not infallible. Human ingenuity—through observation, collaboration, and creative problem-solving—can bridge gaps in AI systems. It’s a reminder that AI works best as a tool to augment human intelligence, not replace it. Together, humans and AI can achieve smarter, more nuanced solutions.
PwC Impact
884
Answers
Orlando, Florida
Updated
PwC’s Answer
On one project, our team faced a machine learning model that produced accurate but opaque recommendations. Instead of accepting outputs at face value, we convened a cross functional group to deconstruct the model, stress test results, and create an interpretability framework. This mix of technical rigor and human creativity built trust in the solution and ultimately improved adoption.
On our engagement the we automated certain testing. Our DA team was getting stuck in one logic in the cost roll testing. Our team liaised with the DA team and were able make the workflow which provided the right results.
One example that stands out was when our team was piloting an AI tool to support analysis and content development. While the tool was efficient, we quickly realized that some of the outputs were confidently written but lacked important context and occasionally misinterpreted nuanced client requirements.
Instead of relying on the tool at face value, we applied human judgment and domain expertise. Team members cross-validated the outputs against source materials, asked deeper clarifying questions, and reframed prompts to better reflect the business objective. We also established a simple review framework checking for accuracy, bias, tone, and alignment to client strategy before anything was shared externally.
By combining AI’s speed with our critical thinking, collaboration, and contextual understanding, we turned a potential risk into a productivity gain. The experience reinforced that AI can accelerate work, but human discernment, accountability, and teamwork are what ensure quality and trust.
On our engagement the we automated certain testing. Our DA team was getting stuck in one logic in the cost roll testing. Our team liaised with the DA team and were able make the workflow which provided the right results.
One example that stands out was when our team was piloting an AI tool to support analysis and content development. While the tool was efficient, we quickly realized that some of the outputs were confidently written but lacked important context and occasionally misinterpreted nuanced client requirements.
Instead of relying on the tool at face value, we applied human judgment and domain expertise. Team members cross-validated the outputs against source materials, asked deeper clarifying questions, and reframed prompts to better reflect the business objective. We also established a simple review framework checking for accuracy, bias, tone, and alignment to client strategy before anything was shared externally.
By combining AI’s speed with our critical thinking, collaboration, and contextual understanding, we turned a potential risk into a productivity gain. The experience reinforced that AI can accelerate work, but human discernment, accountability, and teamwork are what ensure quality and trust.
PwC Impact
884
Answers
Orlando, Florida
Updated
PwC’s Answer
I had the opportunity to work on a problem analyzing the PwC Professional snapshots across the firm. In this study we began with a simple classification based on word embeddings the challenge was how confident could we be in the mapped embedding classifications so we overlayed a simpler more statistical benchmark as a guardrail to improve our confidence in AI's classification. This helped us to support AI in the tech stack and gave us assurance in our recommendations.
I recently built an assignment optimization model in Excel using ChatGPT and Copilot to allocate learners to courses based on their interests. The model performed well on randomized test data, but when we applied it to real registration data, the results were poor. At first, we assumed this was simply due to multiple constraints like room and class capacities.
Through team discussion, however, we identified two issues: first, the AI-generated VBA code contained a logic error; second, the model was constrained by fixed room sizes that didn’t reflect demand. By revising the VBA code and creatively adjusting room capacity — splitting larger spaces unevenly to expand popular classes — we resolved both issues. The model then successfully assigned every learner to their preferred courses.
AI helped us accelerate model development, but it was human problem-solving, critical thinking, and creativity that made the solution truly work.
I recently built an assignment optimization model in Excel using ChatGPT and Copilot to allocate learners to courses based on their interests. The model performed well on randomized test data, but when we applied it to real registration data, the results were poor. At first, we assumed this was simply due to multiple constraints like room and class capacities.
Through team discussion, however, we identified two issues: first, the AI-generated VBA code contained a logic error; second, the model was constrained by fixed room sizes that didn’t reflect demand. By revising the VBA code and creatively adjusting room capacity — splitting larger spaces unevenly to expand popular classes — we resolved both issues. The model then successfully assigned every learner to their preferred courses.
AI helped us accelerate model development, but it was human problem-solving, critical thinking, and creativity that made the solution truly work.
PwC Impact
884
Answers
Orlando, Florida
Updated
PwC’s Answer
It takes me a few minutes to think about my email and sometimes, when you are blinded by emotions, it takes more time to filter through that and send an email that is fair to the receiver. AI can easily help you change your mood through your writing.
Many a times, I brainstormed on my own on how can I work around faster leveraging the ready reports from tools readily available rather than manually struggling with gathering data. Trust me, it boosts the efficiency , happiness and confidence of our working environment to the next level. I would encourage seniors on the project to share such ideas rather than leaving it to the new joiners to figure out ways on their own when its too early for them to learn the nuances of the complex project leading to undue super stressful environment which is needless in our firm's overall collaborative environment.
My team was working on a project that used an AI model to classify and summarize unstructured employee feedback at scale. The AI was good at clustering text into themes, but it often:
Misinterpreted subtle emotions or sarcasm,
Overlooked cultural context in phrasing, and
Produced generic summaries that didn’t capture the nuance leadership needed.
If we had delivered the output “as is,” the results would have been misleading and risked eroding trust in the analysis.
The Human Ingenuity
We took a human-in-the-loop approach here to bring human judgment unlocked meaning and trust.
Many a times, I brainstormed on my own on how can I work around faster leveraging the ready reports from tools readily available rather than manually struggling with gathering data. Trust me, it boosts the efficiency , happiness and confidence of our working environment to the next level. I would encourage seniors on the project to share such ideas rather than leaving it to the new joiners to figure out ways on their own when its too early for them to learn the nuances of the complex project leading to undue super stressful environment which is needless in our firm's overall collaborative environment.
My team was working on a project that used an AI model to classify and summarize unstructured employee feedback at scale. The AI was good at clustering text into themes, but it often:
Misinterpreted subtle emotions or sarcasm,
Overlooked cultural context in phrasing, and
Produced generic summaries that didn’t capture the nuance leadership needed.
If we had delivered the output “as is,” the results would have been misleading and risked eroding trust in the analysis.
The Human Ingenuity
We took a human-in-the-loop approach here to bring human judgment unlocked meaning and trust.
PwC Impact
884
Answers
Orlando, Florida
Updated
PwC’s Answer
When faced with inconsistent AI results, we used human judgment and collaboration (reframing prompts, adding context, etc). The mix of critical thinking and teamwork produced a stronger output than AI alone.
When our team first started piloting AI tools on R&D tax credit studies, we quickly realized that while the technology was powerful, it wasn’t perfect. The AI could organize interview notes and draft activity summaries in minutes, but it often misclassified work—treating routine quality testing as qualifying research or overlooking iterative design efforts that actually counted. Rather than dismiss the tool, we relied on our own ingenuity to bridge the gap. We built a simple validation framework, applied our technical judgment to stress-test the AI’s reasoning, and refined our prompts until the outputs aligned with our standards. Over time, this hybrid approach turned AI into a reliable partner: it gave us speed on first drafts while our expertise ensured accuracy and compliance. The result was faster, higher-quality work and a process we were able to share with other teams to help them adopt AI with confidence.
When our team was working on AI use case to generate recommendations from unstructured documents. The model was producing confident but inconsistent outputs. Instead of tuning the model immediately, we stepped back and reframed the problem - clarifying what decision the recommendations were meant to support and what level of certainty was actually required. Through critical thinking, we identified gaps in the input assumptions and added human review checkpoints for high-impact cases.
When we were working on a divestiture, the AI model gave us TSA cost allocations that looked precise but didn’t reflect the real service complexities. We stepped in to test scenarios, challenge assumptions, and layer in business context the model couldn’t capture. That mix of AI speed and human judgment gave us fairer, more defensible costs for both sides.
Working sessions, office hours and open communication lines to help with general setup and adoption of AI systems. One of the biggest barriers to entry was the actual setup for some of our teams and just getting initially started using the tools. Having office hours, team chats and user guides on setup and best practices has helped to get teams ramped up and using the tools to our benefit.
When our team first started piloting AI tools on R&D tax credit studies, we quickly realized that while the technology was powerful, it wasn’t perfect. The AI could organize interview notes and draft activity summaries in minutes, but it often misclassified work—treating routine quality testing as qualifying research or overlooking iterative design efforts that actually counted. Rather than dismiss the tool, we relied on our own ingenuity to bridge the gap. We built a simple validation framework, applied our technical judgment to stress-test the AI’s reasoning, and refined our prompts until the outputs aligned with our standards. Over time, this hybrid approach turned AI into a reliable partner: it gave us speed on first drafts while our expertise ensured accuracy and compliance. The result was faster, higher-quality work and a process we were able to share with other teams to help them adopt AI with confidence.
When our team was working on AI use case to generate recommendations from unstructured documents. The model was producing confident but inconsistent outputs. Instead of tuning the model immediately, we stepped back and reframed the problem - clarifying what decision the recommendations were meant to support and what level of certainty was actually required. Through critical thinking, we identified gaps in the input assumptions and added human review checkpoints for high-impact cases.
When we were working on a divestiture, the AI model gave us TSA cost allocations that looked precise but didn’t reflect the real service complexities. We stepped in to test scenarios, challenge assumptions, and layer in business context the model couldn’t capture. That mix of AI speed and human judgment gave us fairer, more defensible costs for both sides.
Working sessions, office hours and open communication lines to help with general setup and adoption of AI systems. One of the biggest barriers to entry was the actual setup for some of our teams and just getting initially started using the tools. Having office hours, team chats and user guides on setup and best practices has helped to get teams ramped up and using the tools to our benefit.
PwC Impact
884
Answers
Orlando, Florida
Updated
PwC’s Answer
When faced with inconsistent AI results, we used human judgment and collaboration (reframing prompts, adding context, etc). The mix of critical thinking and teamwork produced a stronger output than AI alone.
When our team first started piloting AI tools on R&D tax credit studies, we quickly realized that while the technology was powerful, it wasn’t perfect. The AI could organize interview notes and draft activity summaries in minutes, but it often misclassified work—treating routine quality testing as qualifying research or overlooking iterative design efforts that actually counted. Rather than dismiss the tool, we relied on our own ingenuity to bridge the gap. We built a simple validation framework, applied our technical judgment to stress-test the AI’s reasoning, and refined our prompts until the outputs aligned with our standards. Over time, this hybrid approach turned AI into a reliable partner: it gave us speed on first drafts while our expertise ensured accuracy and compliance. The result was faster, higher-quality work and a process we were able to share with other teams to help them adopt AI with confidence.
When our team was working on AI use case to generate recommendations from unstructured documents. The model was producing confident but inconsistent outputs. Instead of tuning the model immediately, we stepped back and reframed the problem - clarifying what decision the recommendations were meant to support and what level of certainty was actually required. Through critical thinking, we identified gaps in the input assumptions and added human review checkpoints for high-impact cases.
When we were working on a divestiture, the AI model gave us TSA cost allocations that looked precise but didn’t reflect the real service complexities. We stepped in to test scenarios, challenge assumptions, and layer in business context the model couldn’t capture. That mix of AI speed and human judgment gave us fairer, more defensible costs for both sides.
Working sessions, office hours and open communication lines to help with general setup and adoption of AI systems. One of the biggest barriers to entry was the actual setup for some of our teams and just getting initially started using the tools. Having office hours, team chats and user guides on setup and best practices has helped to get teams ramped up and using the tools to our benefit.
When our team first started piloting AI tools on R&D tax credit studies, we quickly realized that while the technology was powerful, it wasn’t perfect. The AI could organize interview notes and draft activity summaries in minutes, but it often misclassified work—treating routine quality testing as qualifying research or overlooking iterative design efforts that actually counted. Rather than dismiss the tool, we relied on our own ingenuity to bridge the gap. We built a simple validation framework, applied our technical judgment to stress-test the AI’s reasoning, and refined our prompts until the outputs aligned with our standards. Over time, this hybrid approach turned AI into a reliable partner: it gave us speed on first drafts while our expertise ensured accuracy and compliance. The result was faster, higher-quality work and a process we were able to share with other teams to help them adopt AI with confidence.
When our team was working on AI use case to generate recommendations from unstructured documents. The model was producing confident but inconsistent outputs. Instead of tuning the model immediately, we stepped back and reframed the problem - clarifying what decision the recommendations were meant to support and what level of certainty was actually required. Through critical thinking, we identified gaps in the input assumptions and added human review checkpoints for high-impact cases.
When we were working on a divestiture, the AI model gave us TSA cost allocations that looked precise but didn’t reflect the real service complexities. We stepped in to test scenarios, challenge assumptions, and layer in business context the model couldn’t capture. That mix of AI speed and human judgment gave us fairer, more defensible costs for both sides.
Working sessions, office hours and open communication lines to help with general setup and adoption of AI systems. One of the biggest barriers to entry was the actual setup for some of our teams and just getting initially started using the tools. Having office hours, team chats and user guides on setup and best practices has helped to get teams ramped up and using the tools to our benefit.
Michael Rill
Program Management
5
Answers
Redmond, Washington
Updated
Michael’s Answer
Right now, there's a big change in how teams create AI products. In the past, making software started with a detailed plan that engineers would follow. But with AI, that approach doesn't quite fit. Instead of just writing plans, teams now focus a lot on creating evaluations.
An evaluation is like a test to see if the AI is working as people expect. Instead of just making sure a button works or a file saves, we ask: did the AI give a helpful answer? Did it solve the problem in a natural and trustworthy way?
Here are some examples:
For a writing assistant: does the AI make a sentence clearer without changing its meaning?
For a math helper: does it not only find the right answer but also show steps a student can follow?
For a business tool: does it summarize a meeting well enough that someone who missed it feels informed?
These questions can't be graded automatically. They rely on judgment, subtlety, and context. Right now, evaluations need experts—people who know what "good" looks like and can explain tricky situations. AI models get better because humans guide them with their expertise.
This is where human creativity shines. Our team found that the best results happen when experts design evaluations that match what customers really want, not just what's technically correct. It's a partnership: the model brings speed and scale, while humans bring the wisdom to judge quality.
So, AI isn't replacing human work here; it's enhancing it. The evaluations we create today are possible because humans and AI work together, each doing what they do best.
An evaluation is like a test to see if the AI is working as people expect. Instead of just making sure a button works or a file saves, we ask: did the AI give a helpful answer? Did it solve the problem in a natural and trustworthy way?
Here are some examples:
For a writing assistant: does the AI make a sentence clearer without changing its meaning?
For a math helper: does it not only find the right answer but also show steps a student can follow?
For a business tool: does it summarize a meeting well enough that someone who missed it feels informed?
These questions can't be graded automatically. They rely on judgment, subtlety, and context. Right now, evaluations need experts—people who know what "good" looks like and can explain tricky situations. AI models get better because humans guide them with their expertise.
This is where human creativity shines. Our team found that the best results happen when experts design evaluations that match what customers really want, not just what's technically correct. It's a partnership: the model brings speed and scale, while humans bring the wisdom to judge quality.
So, AI isn't replacing human work here; it's enhancing it. The evaluations we create today are possible because humans and AI work together, each doing what they do best.
PwC Impact
884
Answers
Orlando, Florida
Updated
PwC’s Answer
I recently faced a challenge updating macro codes for a new trial balance format. I used ChatGPT to speed up the coding, but the AI’s output didn’t fully align with our business rules. By testing, spotting gaps, and refining the code iteratively, I combined AI’s efficiency with my domain knowledge to deliver an accurate solution. It showed me how human oversight and ingenuity turn AI from a tool into a true partner.
In one project, we explored using AI tools to accelerate research and analysis for complex regulatory and technical issues. While AI was helpful in summarizing large volumes of information, it struggled with context-specific interpretations and edge cases that required professional judgment.
The team addressed this by redesigning the workflow: we used AI as a first-pass assistant, then layered in human review, cross-functional discussion, and scenario testing. Human ingenuity came into play in identifying where AI outputs could be misleading, asking better follow-up questions, and integrating practical business and regulatory considerations that AI could not infer on its own.
This hybrid approach improved efficiency while maintaining accuracy and accountability, reinforcing that AI is most powerful when paired with human expertise rather than used as a replacement for it.
In one project, we explored using AI tools to accelerate research and analysis for complex regulatory and technical issues. While AI was helpful in summarizing large volumes of information, it struggled with context-specific interpretations and edge cases that required professional judgment.
The team addressed this by redesigning the workflow: we used AI as a first-pass assistant, then layered in human review, cross-functional discussion, and scenario testing. Human ingenuity came into play in identifying where AI outputs could be misleading, asking better follow-up questions, and integrating practical business and regulatory considerations that AI could not infer on its own.
This hybrid approach improved efficiency while maintaining accuracy and accountability, reinforcing that AI is most powerful when paired with human expertise rather than used as a replacement for it.
PwC Impact
884
Answers
Orlando, Florida
Updated
PwC’s Answer
Our team was faced with the challenge of ensuring AI usage is fair, bias free and secure. Through collation of human ingenuity, cross functional communication and collaboration, our team was able to devise an operating model that enabled fair and secure use of AI within the client organization.
Team collaboration and innovation. You cannot have this when working as an individual. You become stronger when everyone works together - Learning and Teaching at the same time.
The human element of AI is so important. AI can make much of what we do more efficient but it without human expertise, critical thinking, and professional skepticism - we might miss correcting AI in what may seem small but amounts to a larger error or misdirection.
Using AI to critique our business plans and technology designs to make them more holistic.
We were using GPTs to generate user-stories and not getting good results. The issue was, we expected GPT to be more intelligent than it was/is.
To solve for the issue, we treated GPT as a super-intelligent 10 year old who is great at understanding Qs and prompts, but needs to be provided highly curated steps to get the expected results.
We defined each user-role, scope area, context to get user-stories in manageable chunks vs. expecting GPT to spit out all the user stories at one time.
We are currently placing a strong emphasis on Education and Awareness by hosting open sessions and interactive discussions. These initiatives aim to ensure that our team is well-prepared for future developments and understands how others are leveraging emerging technologies.
We have internal protocols for updating our genAI custom models so they do not stagnate on older input datasets.
We ran a 'spark spotting' session with our team and then dedicated 2 hours to use AI to create quick hitting working prototypes to address our business opportunities.
Team collaboration and innovation. You cannot have this when working as an individual. You become stronger when everyone works together - Learning and Teaching at the same time.
The human element of AI is so important. AI can make much of what we do more efficient but it without human expertise, critical thinking, and professional skepticism - we might miss correcting AI in what may seem small but amounts to a larger error or misdirection.
Using AI to critique our business plans and technology designs to make them more holistic.
We were using GPTs to generate user-stories and not getting good results. The issue was, we expected GPT to be more intelligent than it was/is.
To solve for the issue, we treated GPT as a super-intelligent 10 year old who is great at understanding Qs and prompts, but needs to be provided highly curated steps to get the expected results.
We defined each user-role, scope area, context to get user-stories in manageable chunks vs. expecting GPT to spit out all the user stories at one time.
We are currently placing a strong emphasis on Education and Awareness by hosting open sessions and interactive discussions. These initiatives aim to ensure that our team is well-prepared for future developments and understands how others are leveraging emerging technologies.
We have internal protocols for updating our genAI custom models so they do not stagnate on older input datasets.
We ran a 'spark spotting' session with our team and then dedicated 2 hours to use AI to create quick hitting working prototypes to address our business opportunities.
PwC Impact
884
Answers
Orlando, Florida
Updated
PwC’s Answer
- Within the Advisory Factory we continue to push ourselves to drive better human alignment with agent outputs. As part of this, our Factory team created a patent pending technology that analyzes and clusters human feedback to improve agent outputs over time to ensure those outputs better align with human expectations.
- We did it in our projects we were actually using Google AI to scan thousands of documents and put them into an application and what we found was we had to train the model constantly to gain the level of accuracy finally we got close but we still had to put a human in the loop to make sure that the AL algorithm didn't get it right we still have someone manually corrected
- We did it in our projects we were actually using Google AI to scan thousands of documents and put them into an application and what we found was we had to train the model constantly to gain the level of accuracy finally we got close but we still had to put a human in the loop to make sure that the AL algorithm didn't get it right we still have someone manually corrected