Deloitte University Press has published a white paper “How much time and money AI can save the US government” detailing industry potential for government and AI.

Key problems identified

A recent governing survey of state and local officials found that excessive paperwork burdens was causing trouble for 53% personnel in getting their work done in a 35-to-40-hour week.

Documenting and recording information, was found to be the most time-consuming activity for both federal and state workers. This task captured 10% of both federal and state government work hours


Augmentation with AI frees up 25% of labor hours for more complex tasks;

Depending on levels of federal investment, estimates predict that 96.7 million -1.2 billion federal labor hours will be saved translating to 3.3 billion to 41.1 billion government dollars.

Depending on levels of state investment, estimates predict that 4.3-33.8 million state labor hours will be saved translating to 121 to 931 million government dollars.

Our solutions:

I) Cognitive Robotic Process Automation: Image recognition: Option identifier

Using Deep Learning, we built an application that can analyze PDF / JPEG scans of forms submitted by citizens and identify:

  • Options chosen by user: Yes or No for Ten sections. This application handled the problems of a) Multi-label classification b) Pose problem. The usage of the application reduced human labor and reduced human error.

Fig: Application in English and then in Spanish with Handmade marks to choose options.

Application identified Tick marks, Cross marks, Circle and Shade marks by the user.

  • Presence of Signature

Figure: Presence of Signature in application is identified

  • Presence of Handwriting anywhere in the document. If there were any handwritten messages then human processing was required. The application correctly identified such cases. Here a combination of OpenCV functionality and Deep learning is used.

Figure: Handwriting anywhere in the application is identified

The Deep Learning application that we built used

i) Inception model, ii)Convolutional Neural Networks.

II) Cognitive Robotic Process Automation: Image recognition: Survey summary

A second Deep Learning product recognizes the choices marked by user in answering a survey. A spreadsheet is generated of the results of all survey respondents. The department staff only needs to see the spreadsheet to analyze results of the complete survey.

FFigure: Survey document; results identified by application provided as a spreadsheet