A Teraflow analysis of recent Harvard Business Review Analytic Services research.
Key Takeaways
- 94% of enterprise leaders say well-connected data, processes and applications are critical to AI success. Only 27% say their organisation has them.
- 65% report their structured data is AI-ready. Only 39% say the same about their unstructured data, which holds the majority of institutional knowledge.
- 39% of organisations run AI as standalone tools sitting beside the workflow. Only 12% have AI embedded inside the workflow itself.
- 17% of enterprises have deployed agentic AI. 32% have stopped trying.
- The constraint on enterprise AI value is operating model, not technology.
Source: Harvard Business Review Analytic Services, “Bridging the Readiness Gap to the Agentic Enterprise”, December 2025, n=325 enterprise decision makers.
What is the enterprise AI readiness gap?
The enterprise AI readiness gap is the difference between what an organisation says it needs to make AI work and what it has actually built. The Harvard Business Review Analytic Services survey of 325 enterprise decision makers, conducted in December 2025, measured the gap at 67 percentage points. 94% of leaders say connected data, processes and applications are highly important to AI success. Only 27% say their organisation currently has them.
That gap is the single most predictive signal in the report. It explains why AI pilots stall, why productivity claims fail to land in the P&L, and why most agentic AI deployments are parked at the proof-of-concept stage.
It’s not a model quality problem. It’s not a compute problem. It is not, in the simple sense, even a talent problem.
In Teraflow’s enterprise practice across banking, insurance, telecommunications and retail, the same pattern shows up regardless of sector. Pilots are funded faster than foundations. Models are deployed faster than data is governed. The gap is structural, and it is an operating model problem before it is anything else.
Why does the AI readiness gap exist?
Three reasons, all of them organisational.
First, AI got a budget before data got a strategy. Pilots and point solutions were cheap, fast and visible. Foundational work on master data, semantic layers, content governance and integration is none of those things. When organisations optimise for speed of demo, they under-invest in the substrate that makes the demo scale.
Second, business and IT have been describing the same entities differently for years. “Customer” in the CRM is not the same as “customer” in the billing system. “Policy” in underwriting is not the same as “policy” in claims. AI does not tolerate that fragmentation. It surfaces inconsistency by producing answers nobody trusts.
Third, governance has been treated as a brake instead of a substrate. Most organisations bolt governance on after deployment, when the cost of refactoring is highest and the political appetite is lowest.
Amy Machado, senior research manager in the Content and Knowledge Discovery Strategies program at IDC, frames the consequence directly in the HBR report:
“If organisations don’t trust their data and content, they’re not going to trust the AI built on top of it. And if they don’t trust the system, they’re not going to deploy it broadly or rely on it for more complex decisions.”
That causal chain is the entire enterprise AI scaling problem in two sentences.
Why does unstructured data matter more than most enterprises admit?
The HBR survey makes the cleanest case for this in any recent piece of research. 65% of organisations say their structured data is somewhat or fully prepared for AI. Only 39% say the same about their unstructured data.
Unstructured data is the text, images, video, emails, PDFs, contracts, claims notes, meeting transcripts and call recordings that make up the majority of what an enterprise actually knows about itself. In the bank profiled in the HBR report, the operations director estimated that 60% of organisational data is unstructured. In insurance, financial services and the public sector, the proportion is often higher.
This matters because the highest-value AI use cases (underwriting decisions, claims triage, customer service, compliance review, knowledge work) depend on unstructured content. An organisation that is AI-ready on its database tables and unprepared on its document repositories is AI-ready on the wrong 40%.
Karma Hicks, director of operations and process improvement at American National Bank of Texas, describes the diagnostic the bank now uses:
“As we map our processes and look for efficiency opportunities, we’re focusing on the data itself, where people are sourcing it from and whether it’s flowing into our data warehouse. Is it structured or unstructured? Is it sitting on someone’s desktop?”
That last question is the one most enterprises avoid. The desktop, the inbox, the SharePoint folder, the shared drive nobody owns. Until those repositories are surveyed, classified and either integrated or retired, the AI initiative is operating on a small fraction of the knowledge it needs.
Machado adds the harder layer underneath:
“If you don’t have good data, you’re not going to have good [AI] outcomes. At its core, AI requires data that you can access. That means not only modernising your content but your content applications, too. So many companies still have legacy applications with data they can’t access because it’s hard to integrate and orchestrate. Modernising has really become a strategic imperative.”
The modernisation imperative extends past the data into the platforms that hold it.
Embedded versus standalone AI: which actually works?
Embedded AI works at enterprise scale. Standalone AI rarely does. The HBR data shows the split clearly.
| Approach | Share of organisations | What it looks like | Why it matters |
| Standalone AI | 39% | Separate app sitting beside the workflow. User context-switches to get help. | Context is lost on every switch. Decisions are not written back to systems of record. Adoption depends on individual discipline. |
| Embedded AI | 12% | AI built inside the workflow. Acts on context the system already has. | Compounds across the organisation. Produces audit trails. Adoption follows process design, not individual habit. |
| Mixed | 27% | Some workflows embedded, others standalone. | Transitional state. Value depends on which workflows were embedded first. |
| No AI in workflows yet | 19% | AI tooling not deployed in operational workflows. | Foundational work still ahead. |
Amy Machado of IDC puts the design principle in language that should be on every CIO’s wall:
“When you don’t build AI into the workflow, it feels like one more thing you have to learn. It’s not just an integration of systems. It’s an integration of processes and the experience itself. If it feels like just another tool, it’s not going to stick.”
Stephanie Woerner, principal research scientist at the MIT Sloan School of Management and director of the MIT Center for Information Systems Research, makes the corollary point about enterprise scale:
“Companies have had a hard time figuring out the value of standalone AI productivity tools. They see it, but it’s much more on the individual level. Where you start to make real strides is when you get to enterprise-wide solutions. But getting value from those is going to be a lot harder if you don’t have your data and processes in place.”
In Teraflow’s delivery practice this insight reshapes engagement design. The first question for any enterprise AI workstream is not which model, or even which use case, but which workflow. Until the workflow is mapped end to end and the points where context is lost and decisions are made are explicit, no embedding work can produce value.
What does agentic AI actually require?
Agentic AI requires everything traditional AI requires, plus more.
Agentic systems plan, decide and act autonomously within defined goals and guardrails. The infrastructure demands are correspondingly higher. Connected data across systems. Orchestrated workflows that can carry an end-to-end decision. Governance embedded in the substrate rather than reviewed afterwards. A feedback loop that lets the system learn from its own actions.
The HBR survey shows the current state. 17% of enterprises have implemented agentic AI. 47% are exploring or piloting. 32% are not currently moving forward.
That last number is the more interesting one. Almost a third of enterprises with AI experience have looked at agentic AI and decided to stop. They have learned from prior AI deployments that the foundations are not in place, and that another pilot built on the same fragmented base will produce the same disappointing result.
Stephanie Woerner of MIT Sloan summarises the bar:
“As companies move toward advanced and agentic forms of AI, the bar is being raised. Many organisations know what they need to do, but the real challenge is getting the systems, governance, and data capabilities in place to scale.”
The implication for executives is straightforward. Agentic AI is not the next AI initiative to fund. It is the AI initiative the foundations have to be ready for. Until the operating model can support it, additional investment in agentic pilots will produce the same outcome as the standalone AI investment that came before it.
How should enterprises measure AI value?
The HBR survey found that 51% of organisations measure AI success through productivity, 41% through work speed, 38% through cost savings, 33% through employee experience and 30% through return on investment.
Productivity sits 21 percentage points ahead of ROI. That gap is not a measurement framework. It is a defence mechanism.
Productivity is easy to claim. ROI is hard to defend. Hours saved is not money saved unless reallocated time produces value the business already tracks. As AI moves into higher-judgment work, the quality of the output matters more than the speed of the output.
Nick Tabbal, cofounder and principal consultant at Agentic Consulting, puts the measurement problem cleanly:
“ROI is very hard for someone to come in from the outside and calculate. You can start by measuring productivity gains, hours saved or output increased, and then translate that into dollars, but you also have to look at the quality of the work being produced.”
He adds the qualifier that matters for agentic systems specifically:
“It’s not just how fast people are moving. It’s whether the work coming out of these systems is actually more valuable and whether it’s being done with the right governance and controls in place.”
The leading practitioners in the HBR report measure differently. Karma Hicks at American National Bank of Texas describes the discipline:
“Just because you purchase an application doesn’t mean it’s going to get you the return. It’s how you use it and how you require it to be used.”
The bank now runs proof-of-concept evaluations tied to real workflows before purchase. ROI is assessed up front. Existing platforms are evaluated for under-utilised capability before new ones are added.
At the University of Maryland Global Campus, Sagar Sagiraju, vice president of platform engineering, uses adoption and engagement as proxies for value:
“From an ROI perspective, we look at adoption and engagement and whether that’s reflected in outcomes.”
The discipline is portable. Define what good looks like in metrics the business already cares about. Measure adoption, quality and decision integrity, not just speed and volume. If the AI investment cannot be tied back to an outcome the CFO already tracks, it is an expense line, not a strategy.
How does Teraflow help enterprises bridge the readiness gap?
Our enterprise AI enablement work structures itself around three frameworks that map directly to the gaps the HBR report identifies.
Phase Zero is the design thinking layer. Before any technology decision, Phase Zero maps the workflows that matter, identifies where decisions are made, where context is lost, and where institutional knowledge lives. It produces a target operating model that business and IT can agree on. This is the work the 27% have done and the 73% have not.
DAPA, the Digital AI Platform Accelerator, is the pre-engineered technical architecture that operationalises the Phase Zero output. It addresses fragmented foundations directly: data integration, content modernisation, semantic alignment, governance embedded in the substrate.
FloJo is the enterprise-grade agile delivery framework that takes the architecture into production. It is designed for the workflow-fit problem the HBR report flags as the largest predictor of AI value: AI built into the flow of work, not bolted alongside it.
The proof points are operational. Cell C delivered 10x growth in digital revenue. Comair and Kulula achieved 98% aircraft fill rates while serving 4 million passengers. Vodafone Portugal runs more than 2,900 pipelines processing 3.6 petabytes of data. The platform supports more than 150 ML models in production serving 37 million digital users.
The pattern across these engagements is consistent. Foundations first. Workflow fit second. Measurement frameworks tied to outcomes the business already values. The AI initiative succeeds because the operating model was designed to support it, not because the model was better than the alternatives.
the foundations are the strategy
The HBR Analytic Services research closes with a question from Amy Machado of IDC that should sit at the front of every enterprise AI strategy review:
“AI is going to keep getting more capable. The question for organisations is whether their data, content, and processes are ready to keep up.”
The answer for most enterprises today is no. 73% do not have well-connected foundations. 61% are not ready on unstructured data. 39% are running AI as standalone tools next to the workflow. 32% have stopped trying with agentic AI.
The path forward is not more pilots. It is the unglamorous foundational work that the 27% have already done. A shared data strategy signed by business and IT. Unstructured content prepared at scale. Workflows mapped and redesigned for embedded intelligence. Governance built into the substrate. Measurement frameworks tied to business outcomes.
Strengthening content, data and process foundations is not a precursor to AI success. It is AI success.
Frequently Asked Questions
What is AI readiness?
AI readiness is the state of an organisation’s data, content, processes and operating model that enables AI systems to deliver consistent, trusted, scaled value. It is not a single condition but a set of foundations: well-connected data across systems, prepared unstructured content, embedded workflows, governance integrated into how work happens, and shared definitions between business and IT.
What is the AI readiness gap?
The AI readiness gap is the difference between what enterprises say they need for AI to succeed and what they have actually built. HBR Analytic Services research published in December 2025 measured the gap at 67 percentage points. 94% of enterprise leaders say connected foundations are critical to AI success. 27% say their organisation has them.
Why do enterprise AI projects fail?
Enterprise AI projects fail more often because of operating model issues than technology issues. The HBR survey identifies data silos (54%), data security and privacy issues (48%), data format problems (46%), insufficient governance (46%) and unclear data strategy (45%) as the top barriers. Only 10% of respondents cited data scarcity itself. The problem is rarely too little data. It is fragmented, ungoverned and inaccessible data.
What is unstructured data and why does it matter for AI?
Unstructured data is information that does not fit into the rows and columns of a database. Examples include emails, PDFs, contracts, meeting transcripts, call recordings, images and video. It typically represents the majority of an enterprise’s data and contains most of its institutional knowledge. Only 39% of enterprises have prepared their unstructured data for AI, compared to 65% for structured data.
What is agentic AI?
Agentic AI refers to systems that can plan, decide and act autonomously within defined goals and guardrails. Unlike traditional AI that responds to single prompts, agentic AI can reason across documents and data sources, take multi-step actions, and complete workflows end to end. 17% of enterprises have deployed agentic AI according to recent HBR research, while 32% have decided not to proceed due to foundational gaps.
What is the difference between embedded and standalone AI?
Standalone AI is a separate application the user opens to get AI assistance, typically requiring context switching and manual integration of outputs. Embedded AI is built into the workflow itself, acting on the context the system already has without requiring the user to leave their primary application. 39% of organisations rely mostly on standalone AI. 12% have AI embedded in workflows. Embedded AI is the stronger predictor of sustained enterprise value.
How do I measure AI ROI?
Productivity metrics alone are insufficient. The leading practitioners in HBR’s research combine productivity (hours saved, output increased) with quality of outputs, adoption rates, decision integrity, and a tie-back to business outcomes the organisation already tracks. The discipline is to define what good looks like before deployment, run proof-of-concept evaluations against real workflows, and assess whether the AI investment is being used in ways that produce consequential outcomes.
What is an AI operating model?
An AI operating model is the combination of data strategy, governance, ownership, workflow design and measurement frameworks that lets AI systems scale beyond pilots. It includes shared definitions of key entities between business and IT, embedded governance, content preparation processes, integration architecture, and clear accountability for outcomes. The HBR research suggests this is the single largest constraint on enterprise AI value today.
How do I prepare unstructured data for AI?
Preparing unstructured data for AI is a multistage process. It requires extracting and classifying content, transforming it into a format downstream systems can ingest, applying context and taxonomy that matches the language of the business, and embedding governance into how the content is created and stored. Standardised documents like invoices are relatively easy to validate. Highly unstructured, text-heavy content (contracts, meeting notes, claims correspondence) requires more sophisticated capture, semantic alignment and orchestration.
Source: All statistics in this article are drawn from Harvard Business Review Analytic Services, “Bridging the Readiness Gap to the Agentic Enterprise”, sponsored by Hyland, published December 2025. The survey covered 325 members of the Harvard Business Review audience involved in AI decisions at organisations that have at least considered using AI.
About Teraflow: Teraflow.ai is an enterprise AI enablement consultancy operating across banking, insurance, telecommunications and retail. Teraflow’s frameworks (Phase Zero, DAPA, FloJo) help enterprises bridge the AI readiness gap from operating model through to production delivery.
