IBM published their latest research on why agentic AI projects fail and what separates the companies getting returns from everyone else. Here's what their findings mean for a mid-market business — and the four lessons that matter most.
IBM surveyed thousands of executives and distilled their findings into four recommendations for getting real returns from AI. Their advice is sound — but it's written for enterprises with platform teams, orchestration layers, and governance departments.
If you run a business with 10 to 200 people, the principles still apply. The implementation looks completely different. Here's what each of IBM's four findings actually means at your scale, and what you can do about it without an enterprise budget.
Companies that define ROI goals and KPIs before deploying AI agents see dramatically better outcomes. IBM themselves target US$4.5 billion in productivity gains. Top performers saw 18% ROI, well above average. The key: measuring business outcomes — productivity, revenue, cost savings, satisfaction — not technology metrics.
You don't need a KPI dashboard. You need to answer one question before you start: "How many hours is this workflow currently burning, and what would it look like if that number was halved?" If you can't measure the before, you'll never prove the after. The businesses getting results aren't measuring AI performance — they're measuring whether people got their Tuesday afternoons back.
Practical implication: Before any AI project, map the time. How long does the crew briefing actually take? How many hours go into the compliance check? How long before the proposal reaches the client? These are the numbers that prove return — and they're the numbers most consultants never ask for because it's easier to show an impressive demo than to measure operational reality.
68% of AI-first organisations achieving the highest ROI have mature governance frameworks, compared to just 32% of others. Over half of CEOs are delaying major AI investments until governance is clear. Without oversight, AI introduces risk faster than it delivers returns. IBM recommends AgentOps platforms, compliance monitoring, and policy enforcement layers.
You don't need an AgentOps platform. You need one thing: the document your team already approves should be the same artifact that drives the AI. When the team updates the document, the AI's behaviour updates with it. That's governance without a governance department. The approved document IS the guardrail — not a separate policy sitting in a folder nobody opens.
Practical implication: IBM is right that governance separates winners from losers. But at mid-market scale, governance isn't a platform — it's architecture. The question isn't "do we have an oversight committee?" It's "if someone changes the rules tomorrow, does the system update automatically or does it keep running on yesterday's instructions?" Single source of truth eliminates version drift. That's your governance framework.
Without effective orchestration, organisations get fragmented efforts and diluted returns. A central orchestration layer should coordinate agents, assistants, and automation so they don't operate in silos. IBM warns that focusing on task automation instead of end-to-end workflow transformation creates "agent sprawl" — many disconnected tools that don't contribute to the same outcome.
Agent sprawl is already happening in mid-market businesses — it just looks different. It's three people using ChatGPT for different things with no shared approach. It's a marketing tool that doesn't talk to the CRM. It's AI-generated content that nobody checks. Orchestration at your scale means building systems around whole workflows, not sprinkling AI onto individual tasks. One crew briefing system that handles the whole process is worth more than five separate AI tools.
Practical implication: IBM recommends an orchestration platform. At your scale, the orchestration layer is the workflow itself — designed end-to-end, with clear inputs, outputs, and handoffs. The architecture matters more than the tools. A well-designed workflow on open standards outperforms a collection of best-in-class tools that don't talk to each other. And it survives when the next model ships.
31% of employees admit to actively sabotaging their company's AI strategy. 81% of executives recognise they need the right people with the right incentives. IBM ran an internal AI challenge with 167,000 participants to drive adoption. Their recommendation: turn employees into ambassadors through empowerment, incentives, and hands-on involvement.
You can't run a 167,000-person challenge. But you have something IBM doesn't: you can sit with the three people who actually do the work and build the system around their Tuesday. The most powerful adoption strategy at mid-market scale is to start with the workflow that's causing the most pain, build the solution with the people who live it, and let the result speak for itself. When the crew briefing that took 3 hours now takes 15 minutes, nobody needs convincing.
Practical implication: The sabotage problem is real, but it's usually a fear problem — people are afraid of being replaced. The antidote isn't a change management program. It's showing people that the AI handles the part of their job they hate so they can spend time on the part they're actually good at. Build with the team, not for the team. The first working system converts more sceptics than any all-hands presentation.
IBM's four recommendations — define ROI, governance, orchestration, employee adoption — are not independent. They're a system. And they all point to the same underlying insight: the businesses that get returns from AI aren't the ones with the best technology. They're the ones that changed how work gets done.
The report confirms what operational experience teaches: technology is the easy part. The hard part is knowing which workflows to transform, which to eliminate, and which to leave alone. That's not an AI question — it's a business judgment question. And it's the question that most AI consultants skip because they're in the business of deploying AI, not deciding whether AI is the right answer.
The first question shouldn't be "what can we automate?" It should be "what should we stop doing entirely?" IBM's own data shows why: 75% of AI initiatives didn't deliver. Many of those automated workflows that shouldn't have existed in the first place.
IBM's data points to a layer that their enterprise recommendations don't name directly: operational intent. The missing ingredient in 74% of failed initiatives wasn't more AI capability — it was that nobody translated the business's goals into machine-actionable constraints. Which trade-offs does this process make every day? What matters more — speed or accuracy? Cost or quality? Human operators learn these rules through years of experience. AI agents need them spelled out before they can make a single useful decision.
The systems I build use frontier generative AI — the same models behind Claude and ChatGPT — not the enterprise machine learning platforms IBM recommends. No training data. No model development lifecycle. Working operational systems, usually inside a few weeks, where the document your team approves is the artifact that drives the AI.
If 75% of AI initiatives fail to deliver expected ROI, the problem isn't the AI. The problem is that most organisations automated the wrong things. Before you invest in any AI solution, the highest-return move is to identify the 30–40% of workflow steps that shouldn't exist at all. Eliminating dead work costs nothing, risks nothing, and makes everything that follows simpler and cheaper.
A structured session where we identify the 2–3 workflows burning the most time, map what's worth building, and flag what should be stopped altogether. You walk out with a concrete plan.
Book the Certainty Diagnostic →