This is not another article about AI, machine learning or some fuzzy buzzword concept. This article is about concrete software development that yields immediate results for your organization.

The key components of our value creation methodology

WHEN was the last time you repeated the same steps in Excel but with different data to produce the same report so your boss can decide what to do with the information? Last week? Every week? I’m sorry for you.

WHAT if, instead, you could meet with an engineer, explain your processes and let him automate them for you? This is the most important variable in this value creation methodology: your expert knowledge.

The cool thing about automation, is that a process needs to be automated only… once. Subsequent executions of that process are free of human effort. Did you notice that Excel starts lagging around one million rows? Try a billion. Excel is not the correct tool to analyse big data. There are cloud computing tools for that. Add some math concepts to the mix like statistics (yes stats still work in 2020), and you can come up with nicely automated models that push your processes and knowledge even furthur by making millions of calculations per second on billions of data records. Something you can only dream of doing yourself on your laptop.

Still reading? Good. Now comes the main course.

How we freed our adtech operations staff from routine tasks so they can focus their skills on real issues and therefore create enhanced value for the clients

Our automation methodology consists of 4 activities.

Activity one: The interview – documenting expert knowledge

This is the most crucial activity in your automation journey: capturing the expert knowledge from your staff and translating it into complete and unambiguous documentation. Meet with the experts, ask questions and document the following: what they do, why they do it, how they make decisions, the exceptions, the processes. No small detail should be overlooked. Start broad, just like you were building a table of contents, and then dig down into each aspect.

A thorough interview should be exhausting for both parties. Otherwise the output of that interview would be incomplete. Why? We do our everyday tasks based on our own experiences and habits, most of the time executing them without thinking. To be able to capture expert knowledge, down to the very details, we have to actually think, make sense and establish logical links between the small actions we take in a typical day.

After the initial interview, the interviewer should organise the documentation into a logical sequence of actions using flow-charts, pseudo-code or any other suitable tool. Most probably some iterations of interview follow-ups and documentation refinements will be required. When are these iterations completed? You can answer that by a question: Could a new employee take the documentation and learn the job without asking too many questions? If the answer is yes, then the interview activity is complete.

The output of this first activity is documented expert knowledge.

Activity two: Development – Building a working automation

This activity is about building the automation. Your engineering team will take care of this. 

There is not much to say here except that it is very important to keep a channel of communication open with your expert. Keep the hype high. You may refer to your expert for clarifications. If you have too many questions, go back to Activity one: Interview and complete the documentation properly. Without complete documentation, your implementation will be erroneous and your automation project will fail.

The output of this activity is a working automation.

Activity three: Buy in – Validating expert knowledge and automation

This activity is about gaining the expert’s trust on what was built. How? Simple: the automation software sends recommendations to the expert. These suggestions are to be reviewed and applied by the expert. Because these recommendations are based on expert knowledge, most of them should be considered valid. Expect adjustments to be made to the working automation and to the documented expert knowledge in this activity. As the expert gains trust in the automation, recommendations will be applied with less and less questioning and verification.

The output of this activity is validated documented expert knowledge and a working automation.

You are now ready for the final activity.

Activity four: Audit – Monitoring and adjusting the automation

Recommandations are now automatically applied by the working automation. The expert must have access to a log that lists all the changes done by the automation. The purpose of this log is for the expert to monitor (audit) the behavior of the automation. The expert determines the appropriate sampling of the audit.

Any deviation from what is desired must be reported by the expert to the engineering or product team. The documented expert knowledge and working automation are adjusted accordingly. If required the working automation may go back to Activity three: Buy in.

This activity has no output. Audit should be performed regularly as long as the automation is active. 

A concrete example from district m

district m is an ad tech company based in Montreal, Canada. district m offers ad tech platforms to both media buyers and sellers. FLO, district m’s demand-side platform, is offered to both self-served and managed clients.

In the context of managed clients, our ad ops team executes many repetitive tasks every day in order to optimize programmatic campaigns, so that our clients get the best performance on their ad budget. Keeping pacing steady (the rate at which budget is spent over time) is one of these tasks.

Because ads displayed on any device must go through an auction process, the higher bid price will determine who wins the impression. Therefore, bid price is a variable that has a strong impact on pacing. The higher the bid, the more you spend and vice-versa.

The optimal bid price equals the current market price. The market price is always changing due to countless factors. In absolute terms, the ad ops would have to adjust the bid price continuously in order to bid an optimized price according to the current pacing. Ad ops typically have dozens of campaigns to manage. A human cannot analyse and update dozens of campaigns continuously. But a machine can.

We built an automation based on expert knowledge, big data and math to continuously adjust the bid price on all campaigns in order to keep pacing steady.

Two immediate benefits manifested themselves when we turned on the machine on a campaign: perfect pacing and a decrease in bid price up to 60%. This means that our clients are able to purchase up to 60% more ads with the same ad budget. What about performance? Viewability, click-rate and conversion-rate, etc. remained the same meaning that a campaign generates 60% more clicks, conversions, etc. when this automation is enabled. Too good to be true? We had the same feeling and rechecked everything to make sure there was no mistake.

What about auditing the automation? The automation communicates with our ad ops team via the Slack messaging app when changes happen. Our FLO app displays the change history at the campaign or line item level.