Technical pathway
Use case example
McKinsey & Company
Click a row or column header for more info
The organizational requirements for generative AI range from low to high, depending on the use case.
High
Low
Tech talent
Use software-as-a-service (SaaS) tool
Build software layers on model API
Fine-tune open-source model in-house
Train a foundation model from scratch
Process adjustments
Proprietary data
Costs
Freeing up
customer support representatives’ time for higher- value activities
Accelerating the pace at which research scientists can identify relevant cell features for drug discovery
Helping relationship managers keep up with the pace of public information and data
Changing the work of software engineering
Tech talent
Use software-as-a-service (SaaS) tool
Build software layers on model API
Fine-tune open-source model in-house
Train a foundation model from scratch
Process adjustments
Proprietary data
Costs
Many SaaS tools offer fixed-fee subscriptions of $10 to $30 per user per month; some products have usage-based pricing
Up-front investment is needed to develop the user interface, integrate the solution, and build postprocessing layers
Running costs for API usage and software maintenance
Initial costs ~2x more than building on API due to increased human capital costs required for data cleaning and labeling and model fine-tuning
Initial costs ~10–20x more than building on API due to up-front human capital and tech infrastructure costs
Running costs for model maintenance and cloud computing similar to the above
Use software-as-a-service (SaaS) tool
Build software layers on model API
Fine-tune open-source model in-house
Train a foundation model from scratch
Costs
Proprietary data
Tech talent
Requires large data science and engineering team with PhD-level knowledge of subject matter, best-practice MLOps, and data and infrastructure management skills
Experienced data science and engineering team with MLOps knowledge and resources to check or create labeled data needed
Software development, product management, and database integration capabilities are needed, which require at least 1 data scientist, machine learning engineer, data engineer, designer, and front-end developer
Little technical talent is needed— potentially for selecting the right solution and light integration work
Use software-as-a-service (SaaS) tool
Build software layers on model API
Fine-tune open-source model in-house
Train a foundation model from scratch
Costs
Tech talent
Proprietary data
Foundation models can be trained on large publicly available data, although long-term differentiation comes from adding owned labeled or unlabeled data (which is easier to collect)
A proprietary, labeled data set is required to fine-tune the model, although in some cases it can be relatively small
Because the model is used as is, no proprietary data is needed
Because the model is used as is, no proprietary data is needed
Use software-as-a-service (SaaS) tool
Build software layers on model API
Fine-tune open-source model in-house
Train a foundation model from scratch
Costs
Proprietary
data
Tech talent
Process adjustments
Including the above, when training on external data, thorough legal review is needed to prevent IP issues
Processes for triaging and escalating issues to humans are needed, as well as periodic assessments of model safety
Processes may be needed to enable storage of prompts and results, and guardrails may be needed to limit usage for risk or cost reasons
Processes largely remain the same, but workers should systematically check model results for accuracy and appropriateness
Process adjustments
Process adjustments
Tech talent
Use software-as-a-service (SaaS) tool
Build software layers on model API
Fine-tune open-source model in-house
Train a foundation model from scratch
Process adjustments
Proprietary data
Costs
Freeing up
customer support representatives’ time for higher- value activities
Accelerating the pace at which research scientists can identify relevant cell features for drug discovery
Helping relationship managers keep up with the pace of public information and data
Changing the work of software engineering
Up-front investment is needed to develop the user interface, integrate the solution, and build postprocessing layers
Running costs for API usage and software maintenance
Software development, product management, and database integration capabilities are needed, which require at least 1 data scientist, machine learning engineer, data engineer, designer, and front-end developer
Because the model is used as is, no proprietary data is needed
Processes may be needed to enable storage of prompts and results, and guardrails may be needed to limit usage for risk or cost reasons
Helping relationship managers keep up with the pace of public information and data
Tech talent
Use software-as-a-service (SaaS) tool
Build software layers on model API
Fine-tune open-source model in-house
Train a foundation model from scratch
Process adjustments
Proprietary data
Costs
Freeing up
customer support representatives’ time for higher- value activities
Accelerating the pace at which research scientists can identify relevant cell features for drug discovery
Changing the work of software engineering
Helping relationship managers keep up with the pace of public information and data
Many SaaS tools offer fixed-fee subscriptions of $10 to $30 per user per month; some products have usage-based pricing
Little technical talent is needed—potentially for selecting the right solution and light integration work
Because the model is used as is, no proprietary data is needed
Processes largely remain the same, but workers should systematically check model results for accuracy and appropriateness
Changing the work of software engineering
Helping relationship managers keep up with the pace of public information and data
Changing the work of software engineering
Helping relationship managers keep up with the pace of public information and data
Changing the work of software engineering
Helping relationship managers keep up with the pace of public information and data
Changing the work of software engineering
Tech talent
Use software-as-a-service (SaaS) tool
Build software layers on model API
Fine-tune open-source model in-house
Train a foundation model from scratch
Process adjustments
Proprietary data
Costs
Helping relationship managers keep up with the pace of public information and data
Accelerating the pace at which research scientists can identify relevant cell features for drug discovery
Freeing up
customer support representatives’ time for higher- value activities
Changing the work of software engineering
Initial costs ~2x more than building on API due to increased human capital costs required for data cleaning and labeling and model fine-tuning
Higher running costs for model maintenance and cloud computing
Experienced data science and engineering team with MLOps knowledge and resources to check or create labeled data needed
A proprietary, labeled data set is required to fine-tune the model, although in some cases it can be relatively small
Processes for triaging and escalating issues to humans are needed, as well as periodic assessments of model safety
Tech talent
Use software-as-a-service (SaaS) tool
Build software layers on model API
Fine-tune open-source model in-house
Train a foundation model from scratch
Process adjustments
Proprietary data
Costs
Helping relationship managers keep up with the pace of public information and data
Freeing up
customer support representatives’ time for higher-value activities
Changing the work of software engineering
Accelerating the pace at which research scientists can identify relevant cell features for drug discovery
Initial costs ~10–20x more than building on API due to up-front human capital and tech infrastructure costs
Running costs for model maintenance and cloud computing similar to the above
Requires large data science and engineering team with PhD-level knowledge of subject matter, best-practice MLOps, and data and infrastructure management skills
Foundation models can be trained on large publicly available data, although long-term differentiation comes from adding owned labeled or unlabeled data (which is easier to collect)
Including the above, when training on external data, thorough legal review is needed to prevent IP issues
Freeing up
customer support representatives’ time for higher- value activities
Accelerating the pace at which research scientists can identify relevant cell features for drug discovery
Freeing up
customer support representatives’ time for higher- value activities
Accelerating the pace at which research scientists can identify relevant cell features for drug discovery
Freeing up
customer support representatives’ time for higher- value activities
Accelerating the pace at which research scientists can identify relevant cell features for drug discovery
Freeing up
customer support representatives’ time for higher- value activities
Accelerating the pace at which research scientists can identify relevant cell features for drug discovery