Brendan Humphreys is the CTO of Canva. He returns for JDN on the integration of AI into the software development cycle within his teams.
JDN. How do your teams use AI in their software production cycle? Have you set up internal tools, like a Copilot, to support your engineers on a daily basis?
Brendan Humphreys. We have a very liberal policy in the adoption of AI at Canva. We want each employee to experience these tools to understand how they can apply them to work. We have given a mandate wide enough to experiment, by creating time and space for this, and we support this approach with very permissive licenses. We therefore have licenses for many tools on the market. In the field of technology, we use most of the tools you know: Sourcegraph, AMP code, Cursor, Claude, Chatgpt for code, etc. We make the licenses available to our engineers in a fairly permissive manner. If someone wants a license for a tool, they can simply get it.
With our suppliers, we emphasize a lot to obtain consumption -based pricing rather than paying for unused licenses. We have 2,300 engineers; Providing a license from each tool to each engineer becomes very expensive. We are therefore trying to push suppliers to a pricing for use. Sometimes our engineers offer us new tools and we also have a community of engineers and a very new IT department. We are therefore aware of the changes in tools and we have good direct relations with many of the major suppliers, which allows us to know what is being prepared.
When you buy a license, what criteria do you look at?
We are primarily a security analysis. We have a very sophisticated analysis and test of supplier program. We therefore want to know what data movements take place around these models, which data is used for training, and we want to understand the threat models that these LLM introduce. Generally, the path we follow is as follows: we carry out a quick evaluation of the threat. Once the security team is satisfied, we go to a small pilot project. This pilot program gives us a quick return to find out if this tool is adapted to our needs. Once it is a success, we then consider a more permissive license deployment to the entire technological department.
Do your teams work mainly with IDE enriched by AI, or do they also experience code agents?
We are only at the beginning of the use of complete automation with agents in the Software Engineering. By having many solutions of commercial agents for software development, the challenge is that our code base is simply too large. We have tens of millions of lines of code, and most of the tools would not even manage to clone the Github deposit.
“One of the main advantages that we observe is that these tools allow engineers to stay much longer in a state of flow”
We have therefore done internal development and we have the beginnings of an internal “agentic” support. This is certainly something that we experience. We discuss with the main suppliers and we hope they will take care of these more important use cases, because it is obviously very lucrative for them. We believe, for example, that OpenAi Codex is an excellent example which we expect to be soon capable of managing the size of our deposit.
Is the return on investment of these tools already tangible?
Many fashionable metrics are actually quite superficial. For example, when suppliers proudly tell us that “30% of their code is generated by AI”, I remain skeptical. It is not at all the kind of data that we follow with us. What interests me more is the perception that engineers have their own productivity. On this point, we note that around 80% of engineers consider themselves more productive using these tools. We also observe an increase in work volume (from the output, editor’s note) among engineers who use these tools.
One of the main advantages that we observe is that these tools allow engineers to stay much longer in a state of flow. You have this peer program by your side, which acts as a constant companion that you can consult, but to whom you can also start delegating tasks when you are in this state of concentration. Previously, without the tools of AI, you could have interrupted a task to, for example, notice something to correct in the code base, then you would have gone to do something else, and before you realize it, you found yourself juggling with three tasks at the same time.
Beyond their profits, have you also identified limits or constraints in the use of these AI tools by your teams?
We know very well that the tools hallucinate. They are sometimes mistaken with disconcerting insurance. They must therefore be supervised by highly qualified human engineers. It is an undeniable limitation. We also note that even the most advanced models, with large context windows, are not proportionally efficient on the large code bases. Even with a very large context window, there are decreasing yields compared to its size.
Have you seen differences in the way junior and senior developers appropriate and use AI on a daily basis?
Yes, we observe differences. Junior engineers use these tools to prototyper and identify technical solutions to their problems. Our experienced engineers, from the intermediate level to senior, use these tools as a kind of super-power. They integrate them into their IDE and use them as a program program to check their work and generate code, such as a really intelligent self -compulsory.
Do you plan to use fully automated agents, capable not only to generate code but also to provide more complex tasks, such as migration or software maintenance?
At this stage, we do not see room for complete automation. These tools must be supervised. I think that the automation that enthusiasts comes in when you have a well -defined problem: you can then use an agency code solution to produce a solution proposal. But this proposal then remains in an environment of development or prototyping where a human can both examine the code and test the functionality, and it is this person who ultimately validates the change. I don’t think these tools are mature enough to give them total confidence in autonomous mode.
Some companies are already planning to substitute part of the developers, especially juniors, by AI. Do you think this scenario could occur in the long term, including at Canva?
In the medium term, we certainly do not think that this is a possibility, for the reasons I mentioned. I think that in the end, humans must remain those who guide, those who exercise aesthetic judgment and define quality, those who assess and assume the responsibility of the end result.
“In the end, the coding represents only a tiny part of the job of Software Engineer”
For this, it is essential that junior engineers continue to integrate the industry, bringing their enthusiasm, their passion and their innovative ideas. In the end, the coding represents only a tiny part of the profession of Software Engineer. We are enthusiastic about these tools that assist us in this aspect, but the job of software engineer accomplished has so many other dimensions that AI cannot replace it. Even with AI, we therefore always see a promising future for junior engineers in the field of Software Engineering.
