Category: Tech News

The interplay between ownership, outsourcing, and remote work

As we enter 2025, artificial intelligence (AI) is taking center stage at companies across industries. Faced with the twin challenges of acting decisively in the short run (or at least appearing to do so to reassure various stakeholders) and securing a prosperous future for the company in the long run, executives may be compelled to launch strategic AI initiatives. The aims of these initiatives can range from upgrading the company’s technical infrastructure and harvesting large amounts of high-quality training data, to improving the productivity of employees and embedding AI across the company’s products and services to offer greater value to customers.

Organizing in the right way is crucial to the successful implementation of such AI initiatives and can depend on a company’s particular context, e.g., budgetary constraints, skills of existing employees, and path dependency due to previous activities. This article takes a closer look at the interplay between three key dimensions of organizing for AI in today’s complex world: ownership, outsourcing, and proximity. We will see how different combinations of these dimensions could manifest themselves in the AI initiatives of various companies, compare pros and cons, and close with a discussion of past, present, and future trends.

Note: All figures and tables in the following sections have been created by the author of this article.

Guiding Framework

Figure 1 below visualizes the interplay between the three dimensions of ownership, outsourcing, and proximity, and this will serve as the guiding framework for the rest of the article.

The ownership dimension reflects whether the team implementing a given initiative will also own the initiative going forward, or instead act as consultants to another team that will take over long-term ownership. The outsourcing dimension captures whether the team for the initiative is primarily staffed with the company’s own employees or external consultants. Lastly, the proximity dimension considers the extent to which team members are co-located or based remotely; this dimension has gained in relevance following the wide experimentation with remote work by many companies during the global COVID-19 pandemic and throughout the escalation of geopolitical tensions around the world since then.

Although Figure 1 depicts the dimensions as clear-cut dichotomies for the sake of simplicity (e.g., internal versus external staffing), they of course have shades of gray in practice (e.g., hybrid approaches to staffing, industry partnerships). In their simplified form, the boxes in Figure 1 suggest eight possible ways of organizing for AI initiatives in general; we can think of these as high-level organizational archetypes. For example, to build a flagship AI product, a company could opt for an internally staffed, co-located team that takes full long-term ownership of the product. Alternatively, the company might choose to set up an outsourced, globally dispersed team, to benefit from a broader pool of AI talent.

Organizational Archetypes for AI

Table 1 below provides an overview of the eight high-level organizational archetypes, including real-life examples from companies around the world. Each archetype has some fundamental pros and cons that are largely driven by the interplay between the constituent dimensions.

Archetypes with high ownership tend to offer greater long-term accountability, control, and influence over the outcomes of the AI initiative when the level of outsourcing is minimal, since in-house team members typically have more “skin in the game” than external consultants. But staffing AI experts internally can be expensive, and CFOs may be especially wary of this given the uncertain return on investment (ROI) of many early AI initiatives. It may also be harder to flexibly allocate and scale the scarce supply of in-house experts across different initiatives.

Meanwhile, archetypes that combine a high level of outsourcing and low proximity can allow AI initiatives to be implemented more cost-effectively, flexibly, and with greater infusion of specialized external expertise (e.g., a US-based company building an AI product with the help of externally sourced AI experts residing in India), but they come with cons such as external dependencies that can result in vendor lock-in and lower retention of in-house expertise, security risks leading to reduced protection of intellectual property, and difficulties in collaborating effectively with geographically dispersed external partners, potentially across time zones that are inconveniently far apart.

Current and Future Trends

As the real-life examples listed in Table 1 show, companies are already trying out different organizational archetypes. Given the trade-offs inherent to each archetype, and the nascent state of AI adoption across industries overall, the jury is still out on which archetypes (if any) lead to more successful AI initiatives in terms of ROI, positive market signaling, and the development of a sustained competitive advantage.

However, some archetypes do seem to be more common today — or at least have more vocal evangelists — than others. The combination of high ownership, low outsourcing, and high proximity (e.g., core AI products developed by co-located in-house teams) has been the preferred archetype of successful tech companies like Google, Facebook, and Netflix, and influential product coaches such as Marty Cagan have done much to drive its adoption globally. Smaller AI-first companies and startups may also opt for this organizational archetype to maximize control and alignment across their core AI products and services. But all these companies, whether large or small, tend to show strong conviction about the value that AI can create for their businesses, and are thus more willing to commit to an archetype that can require more funding and team discipline to execute properly than others.

For companies that are earlier in their AI journeys, archetypes involving lower ownership of outcomes, and greater freedom of outsourcing and remote staffing tend to be more attractive today; this may in part be due to a combination of positive signaling and cautious resource allocation that such archetypes afford. Although early-stage companies may not have identified a killer play for AI yet, they nonetheless want to signal to stakeholders (customers, shareholders, Wall Street analysts, and employees) that they are alert to the strategic significance of AI for their businesses, and ready to strike should a suitable opportunity present itself. At the same time, given the lack of a killer play and the inherent difficulty of estimating the ROI of early AI initiatives, these companies may be less willing to place large sticky bets involving the ramp-up of in-house AI staff.

Looking to the future, a range of economic, geopolitical, and technological factors will likely shape the options that companies may consider when organizing for AI. On the economic front, the cost-benefit analysis of relying on external staffing and taking ownership of AI initiatives may change. With rising wages in countries such as India, and the price premium attached to high-end AI services and expertise, the cost of outsourcing may become too high to justify any benefits. Moreover, for companies like Microsoft that prioritize the ramp-up of internal AI R&D teams in countries like India, it may be possible to reap the advantages of internal staffing (alignment, cohesion, etc.) while benefiting from access of affordable talent. Additionally, for companies that cede ownership of complex, strategic AI initiatives to external partners, switching from one partner to another may become prohibitively expensive, leading to long-term lock-in (e.g., using the AI platform of an external consultancy to develop custom workflows and large-scale models that are difficult to migrate to more competitive providers later).

The geopolitical outlook, with escalating tensions and polarization in parts of Eastern Europe, Asia, and the Middle East, does not look reassuring. Outsourcing AI initiatives to experts in these regions can pose a major risk to business continuity. The risk of cyber attacks and intellectual property theft inherent to such conflict regions will also concern companies seeking to build a lasting competitive advantage through AI-related proprietary research and patents. Furthermore, the threat posed by polarized national politics in mature and stagnating Western economies, coupled with the painful lessons learned from disruptions to global supply chains during the COVID-19 pandemic, might lead states to offer greater incentives to reshore staffing for strategic AI initiatives.

Lastly, technologies that enable companies to organize for AI, and technologies that AI initiatives promise to create, will both likely inform the choice of organizational archetypes in the future. On the one hand, enabling technologies related to online video-conferencing, messaging, and other forms of digital collaboration have greatly improved the remote working experience of tech workers. On the other hand, in contrast to other digital initiatives, AI initiatives must navigate complex ethical and regulatory landscapes, addressing issues around algorithmic and data-related bias, model transparency, and accountability. Weighing the pros and cons, a number of companies in the broader AI ecosystem, such as Zapier and Datadog, have adopted a remote-first working model. The maturity of enabling technologies (increasingly embedded with AI), coupled with the growing recognition of societal, environmental, and economic benefits of fostering some level of remote work (e.g., stimulating economic growth outside big cities, reducing pollution and commute costs, and offering access to a broader talent pool), may lead to further adoption and normalization of remote work, and spur the development of best practices that minimize the risks while amplifying the advantages of low proximity organizational archetypes.

Organizing for AI was originally published in Towards Data Science on Medium, where people are continuing the conversation by highlighting and responding to this story.

You can’t afford to remain an AI-ignoramus, even if your product isn’t using an LLM

If you’re a Software Architect, or a Tech Lead, or really anyone senior in tech whose role includes making technical and strategic decisions, and you’re not a Data Scientist or Machine Learning expert, then the likelihood is that Generative AI and Large Language Models (LLMs) were new to you back in 2023.

AI was certainly new to me.

We all faced a fork in the road —should we invest the time and effort to learn GenAI or continue on our jolly way?

At first, since the products I’m working on are not currently using AI, I dipped my toes in the water with some high-level introductory training on AI, and then went back to my day job of leading the architecture for programmatic features and products. I reassured myself that we can’t all be experts in everything — the same way I’m not a computer vision expert, I don’t need to be an AI expert — and instead I should remain focused on high level architecture along with my core areas of expertise- cloud and security.

Did you face a similar decision recently?

I’m guessing you did.

I’m here to tell you that if you chose the path of LLM-semi-ignorance, you’re making a huge mistake.

Luckily for me, I was part of a team that won a global hackathon for our idea around using AI for improving organizational inclusivity, and that kicked off a POC using GenAI. That then led to co-authoring a patent centered around LLM fine-tuning. I was bothered by my ignorance when AI terms flew around me, and started investing more in my own AI ramp-up, including learning from the experts within my organization, and online courses which went beyond the introductory level and into the architecture. Finally it all clicked into place. I’m still not a data scientist, but I can understand and put together a GenAI based architecture. This enabled me to author more patents around GenAI, lead an experimental POC using an LLM and AI agents, and participate in a GenAI hackathon.

What I learned from this experience is that GenAI is introducing entirely new paradigms, which are diametrically opposed to everything I knew until now. Almost every single fact from my computer science degrees, academic research, and work experience is turned on its head when I’m designing a GenAI system.

GenAI means solving problems using non-deterministic solutions. We got used to programmatic and deterministic algorithms, allowing us to predict and validate inputs against outputs. That’s gone. Expect different results each time, start thinking about a percentage of success vs absolute correctness.

GenAI means results are not linear to development effort investment. Some problems are easy to solve with a simple prompt, others require prior data exploration and complex chains of multiple agentic AI workflows, and others require resource heavy fine tuning. This is very different than assessing a requirement, translating it to logical components and being able to provide an initial decomposition and effort assessment. When we use an LLM, in most cases, we have to get our hands dirty and try it out before we can define the software design.

As a software architect, I’ve begun assessing tradeoffs around using GenAI vs sticking with class programmatic approaches, and then digging deeper to analyze tradeoffs within GenAI — Should we using fine tuning vs. RAG, where is an AI agent needed? Where is further abstraction and multi-agents needed? Where should we combine programmatic pre-processing with an LLM?

For every single one of these architectural decision points- GenAI understanding is a must.

I became a Software Architect after over a decade of experience as a Software Engineer, developing code in multiple languages and on multiple tech stacks, from embedded to mobile to SaaS. I understand the nuts and bolts of programmatic code, and even though I’m not writing code anymore myself, I rely on my software development background both for making high level decisions and for delving into the details when necessary. If as tech leaders we don’t ensure that we gain equivalent knowledge and hands-on experience in the field of GenAI, we won’t be able to lead the architecture of modern systems.

In other words — I realized that I cannot be a good Software Architect, without knowing GenAI. The same way I can’t be a good Software Architect if I don’t understand topics such as algorithms, complexity, scaling; architectures such as client-server, SaaS, relational and non-relational data bases; and other computer science foundations.

GenAI has become foundational to computer engineering. GenAI is no longer a niche sub-domain that can be abstracted away and left to Subject Matter Experts. GenAI means new paradigms and new ways of thinking about software architecture and design. And I don’t think any Software Architect or Tech Leader can reliably make decisions without having this knowledge.

It could be that the products and projects you lead will remain AI free. GenAI is not a silver bullet, and we need to ensure we don’t replace straightforward automation with AI when it’s not needed and even detrimental. All the same, we need to be able to at least assess this decision knowledgeably, every time we face it.

I’m going to end with some positive news for Software Architects — yes we all have to ramp-up and learn AI — but once we do, we’re needed!

As GenAI based tools become ever more complex, data science and AI expertise is not going to be enough — we need to architect and design these systems taking into account all those other factors we’ve been focused on until now — scale, performance, maintainability, good design and composability — there’s a lot that we can contribute.

But first we need to ensure we learn the new paradigms as GenAI transforms computer engineering — and make sure we’re equipped to continue to be technical decision makers in this new world.

Why Every Software Architect Needs to Learn GenAI was originally published in Towards Data Science on Medium, where people are continuing the conversation by highlighting and responding to this story.