In a previous article,
we mentioned that clients and potential clients invariably ask how they can use Artificial Intelligence to add value to their business.
The next question inevitably asked is "how much will this cost?".
From experience, almost all clients initially have no idea whatsoever.
An AI Project which includes Machine Learning, Natural Language Processing and a Web Interface or API
is inherently difficult to price.
The reason for this is that Machine learning is applied data science and cannot be considered regular
programming. Initially, implementation is more difficult for organizations
that already have a non-machine learning infrastructure as compared to a
"tabula rasa" start-up.
Specifying requirements is the main problem, as they often change midway through a
project as a result of new features, development issues, integration
issues, user acceptance and more. Commonly, clients know that there is
undoubtedly value to be obtained from their AI project but cannot define
requirements and are not sure of the expected results. Given an unclear
project definition it is not possible to create a matching project solution
Expected results are also another major difficulty. Critical to any Machine
Learning project, clear expectations in regards to accuracy, precision,
recall and scores must be set at the very start. It is important to clarify what can
be done, what cannot be done, what the expected results are and
what the value provided will be for a client’s business.
As an example, we will price a project where requirements
involve building a specialized crawler that retrieves ‘about us’ texts
from corporate web sites and feeds them to an AI Module that classifies
them by conceptual similarity. Next in a similar manner, all current customer
‘about us’ pages are crawled and classified. The goal
is to compare the first with the second set with a view to identifying
those new companies from the first set that are good potential fits as new
clients (i.e. AI based lead generation).
Let’s begin with the following assumptions:
o the client does not have an ML server
o the texts to be processed do not contain sensitive information
o the texts can be securely transferred offsite on to the cloud
The process will be as follows:
o Compare a number of word embedding methods and classification
o Deploy the most accurate model as a web service
The Project Plan
Now let’s make a plan.
We consider three toolset options and proceed considering which
one best matches a client’s current technology infrastructure.
The project will live on Windows Azure.
We will use:
o The Azure ML Workbench
o An Azure Data Science Virtual Machine with GPU
o Word2vec word embeddings, Doc2vec
o TensorFlow, CNTK and Keras
The project will live on Amazon Web Services (AWS)
We will use:
o 2 AWS instances (one for web and one as the ML server)
o TensorFlow, Keras, Elasticsearch, Flask
o AngularJS GUI
o Jupyter notebook
The project will live on Fortuitapps servers
We will use:
o The Fortuit AI Engine
10 hours for instance provisioning and certificate installs
60 hours qualifying, evaluating, implementing, testing and selecting an
optimal ML Model
30 hours User Interface (GUI) Development
10 hours of support, training and follow up
Total 110 Hours
So, for a total of 110 hours at a rate of £90 per hour the total cost
would be £9900.
At Fortuitapps our aim is always to provide pricing at an appropriate level that will
neither create a net loss or
conversely producing an unfairly large profit margin. Our moto: “what we do works”.
About the Author