75% of DIY Agent Architectures Will Fail, And Forrester’s Reasoning Deserves More Attention

There’s a seductive logic to building your own AI agents. You know your business better than any vendor. Your workflows are unique. Off-the-shelf solutions won’t capture the nuances. And besides, your engineering team is world-class, they can handle it.

Forrester thinks you’re wrong. In their 2025 Predictions for Artificial Intelligence, the firm predicted that three out of four enterprises attempting to build advanced agentic architectures on their own would fail. Not struggle. Not underperform. Fail.

The prediction didn’t get the attention it deserved when it was published, partly because it competed for airtime with flashier headlines about AI capabilities and partly because the enterprise technology press had already developed prediction fatigue. But Forrester’s reasoning was specific enough to be actionable, and the months since publication have validated it with uncomfortable precision.

Why the failure rate is so high

Forrester’s analysts didn’t just throw out a scary number. They identified the specific technical barriers that make DIY agentic architectures categorically harder than other enterprise technology builds.

The core argument: agentic AI architectures are “convoluted, requiring diverse and multiple models, sophisticated retrieval-augmented generation stacks, advanced data architectures, and niche expertise.” That’s a dense sentence, so let me unpack what it means in practice.

A production agentic system isn’t one AI model doing one thing. It’s an orchestration layer managing multiple specialized models, each with different strengths, latencies, cost profiles, and failure modes. The agent needs to decide which model to invoke for which task, handle graceful fallback when a model produces garbage, manage context windows across multi-step workflows, and maintain state across interactions that may span hours or days.

Then layer in RAG, retrieval-augmented generation, which sounds simple in architecture diagrams but requires solving hard problems in document chunking, embedding quality, retrieval relevance, and real-time index updates. A RAG stack that works for a demo with 50 documents behaves very differently from one handling 5 million documents across 12 different formats with daily updates.

Then add the data architecture requirements: clean, well-structured, consistently formatted data that the agent can actually use. Most enterprises don’t have this. They have data spread across decades-old systems, each with its own schema, access patterns, and quirks that only the people who’ve been there for ten years understand.

Then staff it. Forrester’s point about “niche expertise” is the quiet killer. The intersection of ML engineering, distributed systems architecture, security engineering, and domain expertise required to build production agentic systems is narrow. These people exist, but there aren’t enough of them, and they’re expensive.

Rowan Curran, a Forrester analyst, was even more specific in his comments to TechTarget: the idea of multi-agent and multi-ecosystem architectures, where an AWS agent communicates with a Microsoft agent and then a Salesforce agent, “is not what we see as being a feasible thing in the short term” because of the lack of testing and validation capabilities for such architectures.

The parallel Forrester predictions that compound the problem

The 75% failure prediction didn’t exist in isolation. Forrester’s 2025 forecast included several adjacent predictions that collectively paint a picture of an industry heading for a correction.

Craig Le Clair, Vice President and Principal Analyst at Forrester, predicted that implementation challenges would stall 25% of agentic AI efforts even among organizations using vendor platforms rather than building their own. The drivers: vague business objectives, premature integration in decision-making, and the challenge of determining the right level of autonomy to balance risk and efficiency.

Le Clair also predicted that GenAI would orchestrate less than 1% of core business processes in 2025, with deterministic automation handling the heavy lifting while AI provided “bursts of insight and efficiency.” That framing, AI as a complement to deterministic systems, not a replacement for them, runs directly counter to the narrative driving most enterprise agent investments.

The firm also predicted that most enterprises fixated on AI ROI would scale back their efforts prematurely. As Jayesh Chaurasia and Sudha Maheshwari wrote in Forrester’s AI blog: “The expectation for immediate returns on AI investments will see many enterprises scaling back their efforts sooner than they should. This retreat risks stifling long-term growth and innovation as leaders realize that the ROI from AI will unfold over a more extended period than initially anticipated.”

So the enterprise AI picture, according to Forrester, is this: 75% of self-built agent systems will fail, 25% of vendor-built systems will stall, GenAI will handle less than 1% of core processes, and most organizations chasing fast ROI will quit too early. That’s not a technology adoption curve. That’s a reckoning.

Gartner’s data reinforces the pattern

Forrester’s prediction gained additional weight from Gartner’s parallel research. A Gartner survey found that only 15% of IT leaders were considering fully autonomous AI agents. Meanwhile, 74% viewed agents as a new attack vector, a concern that directly conflicts with the “deploy agents everywhere” strategy most enterprises are pursuing.

The 74% number is striking because it means enterprise security leaders have already concluded that agents create more risk than they mitigate, even as their counterparts in operations and product are racing to deploy them. That internal tension, security teams pulling one direction, business teams pulling the other, is a recipe for the kind of governance vacuum that turns experimental deployments into production liabilities.

When Forrester later assessed the accuracy of their 2025 predictions, the agentic AI failure call was tracking as expected. Their mid-year review noted that “the vast majority of the agentic use cases currently hitting production are deployed off of vendor platforms, and for good reason, building agentic architectures is much more challenging than what the very robust vendor marketing may have you believe.” They referenced the MIT finding that 95% of organizations investing in generative AI showed zero P&L impact, estimating that “agentic investments are in the same ballpark.”

What mature organizations are doing instead

Forrester’s prescription was clear: “Mature companies will recognize these limitations and opt to collaborate with AI service providers and systems integrators, leveraging their expertise to build cutting-edge agentic solutions.”

That sounds like a vendor pitch, but the logic underneath it is sound. The organizations succeeding with agentic AI share several characteristics that distinguish them from the 75% heading for failure.

They start with a specific, well-defined use case, not a horizontal “agent platform.” The successful deployments I’ve observed in my work and through my involvement with CoSAI and IETF AGNTCY are narrowly scoped. An agent that handles a specific category of customer inquiry. An agent that automates a defined compliance review workflow. An agent that monitors a specific set of infrastructure metrics and escalates anomalies. Not “an agent that does everything.”

They fix the workflow before automating it. This is the step that most organizations skip because it’s unglamorous. If your current process requires six handoffs between three teams and nobody can explain why step four exists, layering an AI agent on top of that process will automate confusion at machine speed. The organizations getting value from agents first map and optimize the target workflow, then automate the optimized version.

They treat governance as a prerequisite. Before any agent touches production, successful organizations have answered: Who is accountable when this agent acts wrong? What can this agent do, and what is it prohibited from doing? How do we monitor the agent’s decisions in real time? What’s the rollback plan when something goes wrong?

They buy specialized vertical solutions and build only the integration layer. This is Forrester’s core recommendation, and it maps to what I’ve seen work in practice. A purpose-built agent for IT service management, or for financial reconciliation, or for code review, built by a team that has spent years refining it, will outperform a from-scratch build by a general engineering team every time. The engineering team’s value isn’t in building the agent itself. It’s in integrating the agent securely into the enterprise’s specific environment.

The build-versus-buy decision has changed

The traditional build-versus-buy framework assumed that building gave you customization and control at the cost of speed, while buying gave you speed at the cost of flexibility. For most enterprise technologies, that tradeoff was genuine.

For agentic AI, the equation has shifted. Building from scratch doesn’t just sacrifice speed, it sacrifices reliability, security, and the accumulated learning from thousands of production deployments that the specialized vendors have already internalized. The complexity Forrester identified isn’t the kind of complexity that smart engineers overcome with enough time. It’s the kind that requires organizational infrastructure, dedicated ML ops teams, continuous evaluation pipelines, safety testing frameworks, that most enterprises haven’t built yet.

The 75% failure prediction isn’t a commentary on engineering talent. It’s a commentary on organizational readiness. And the organizations honest enough to recognize that gap will redirect their engineering resources from building agents to governing, integrating, and securing the agents they buy, which is where the real competitive differentiation lives anyway.

Forrester gave the enterprise AI market a number: 75%. The market would be wise to take it seriously before the cancellation notices start arriving.