For decades, enterprise technology strategies have hinged on a deceptively simple question: should we build it, buy it, or partner for it?
The answer has never been simple, and artificial intelligence (AI) has put an entirely new spin on it. Particularly the emergence of agentic AI solutions, a new class of AI systems, capable of reasoning across workflows, initiating tasks and chaining actions without constant human oversight.
PYMNTS Intelligence in the September 2025 PYMNTS Data Book series finds that when it comes to buying, building or partnering for agentic AI solutions, CIOs and CFOs are weighing not just cost, control and speed, but a new dimension: autonomy. How much agency should a machine have inside the enterprise, and how much of that agency should the enterprise itself own?
After all, the report found half of “high impact” gen AI companies, meaning those with significant gen AI applications carrying significant business value and risk, are using or pursuing agents.
The New Frontier of Autonomy
The arrival of agentic AI is increasingly exposing fault-lines in corporate risk tolerance, technical debt and vendor strategy, pushing companies toward a nuanced “build and buy” hybrid that mirrors both the promise and peril of this new technology.
PYMNTS data found that 71% of active adopters of agentic AI are building in-house, while 43% buy off the shelf. Most tellingly, none of the report respondents granted agents “substantial or full autonomy,” although 43% allowed “moderate” system access with human approval.
Advertisement: Scroll to Continue
Every major technology cycle starts with overreach. Early enterprise AI promised to replace whole departments; robotic process automation (RPA) vowed to end manual work. The pattern repeats: vision races ahead of reliability. With agentic AI, however, the lesson has finally been learned. Enterprises are prioritizing focused, measurable use-cases rather than moonshots.
The more an AI can chain actions together, the more it can inadvertently create cascading effects. That’s why most enterprises are deliberately deploying agentic systems in “low-to-moderate risk” functions first: planning, reporting, data synthesis or internal knowledge management.
This focus is not just about risk. It’s about integration. Agentic systems are only as effective as the ecosystems they inhabit. To reason across workflows, they must access clean data, reliable APIs and consistent governance structures. Few enterprises can offer that level of coherence across every function, so they’re starting where the connective tissue is already strong.
This strategy echoes the early cloud-migration playbook: start with low-risk workloads, prove ROI, then expand. But unlike cloud adoption, where infrastructure could be commodified, agentic AI forces organizations to confront the very architecture of their workflows. It’s less about moving applications and more about rethinking the relationships between humans, machines and decisions.
Building in this context doesn’t always mean coding from scratch. Rather, it involves assembling and orchestrating modular components — foundation models, domain-specific data-sets, workflow APIs — into bespoke agentic architectures. The emphasis is on shaping behavior, constraints and integration points that align with the company’s risk appetite and regulatory context.
Still, few enterprises can or should attempt full vertical integration. The complexity of maintaining secure, reliable and up-to-date AI stacks is immense. That’s why even the most engineering-centric firms continue to “buy” foundational tools: orchestration platforms, vector-databases, agent frameworks and monitoring dashboards.
Vendors have responded with a flood of “agent-as-a-service” offerings — tools that let companies deploy semi-autonomous systems without reinventing the infrastructure wheel. These platforms often provide guardrails, governance templates and human-in-the-loop features allowing organizations to experiment safely.
The financial argument for hybrid build-and-buy also rests on optionality. Building core capabilities preserves flexibility for future shifts in model economics or vendor landscapes. Buying complementary components reduces opportunity-cost and accelerates experimentation.
Source: https://www.pymnts.com/