.webp){kind=icon}
0
.webp){kind=icon}
28 January 2026, 10:00:00 am AEDT
.webp){kind=icon}
16 min
Hard Lessons From 100+ Deployments Across IBM’s Ecosystem
What I Wish I Had Understood Earlier – by Steny Sebastian, Principal Data and AI Platforms
Over the past year, I have been deeply immersed in enterprise transformations. Not slideware. Not proofs of concept. Real systems operating in production across HR, finance, insurance, sales, and core operations.
Many of these lessons emerged through close collaboration with IBM Product Managers, IBM APAC Technical and Sales Leader, IBM Client Engineering, and Octane AI engineers. Working together on real delivery challenges, sharing perspectives, and building agentic AI systems in practice made this journey both demanding and genuinely exciting, and shaped the thinking reflected here.
What surprised me most was not how advanced the technology had become.
It was how frequently capable, well-funded, and highly motivated organisations still struggled to move AI beyond experimentation.
Across more than 100+ enterprise businesses and AI initiatives, the pattern became impossible to ignore. The organisations that faltered were no less ambitious or less intelligent.
They made the same three mistakes.
And I have made versions of all three myself.
If you are serious about deploying agentic AI that delivers measurable outcomes, these are the lessons enterprise reality teaches you quickly, and often the hard way.
The Reality Check Most AI Programs Avoid
Most enterprises launch AI initiatives with genuine optimism. Boards approve funding. Executives champion innovation. Teams move quickly.
Then momentum quietly stalls.
Around 84 per cent of organisations are experimenting with GenAI. Only 26 per cent move into limited deployment. Just 10 per cent operate AI at scale.
This gap is not caused by a lack of models, tooling, or budget.
It is caused by an execution failure.
Due to unclear ownership.
By architectures that collapse under real-world complexity.
Agentic AI only succeeds when it is designed to run work, not just assist it.
The Three Mistakes That Kill Enterprise AI and the Rebuttals That Actually Work
Mistake #1: We Tried to Fix Everything at Once
❌ “Let’s automate the entire HR function.”
✔️ “Let’s fix resume screening first.”
This mistake is driven by optimism.
When leaders finally see what AI can do, narrowing the scope feels like underthinking the opportunity. Why fix one workflow when you could redesign the whole function?
Because the enterprise will not let you.
I have seen teams attempt to re-engineer entire functions in one move. HR. Finance. Customer Service. The result is always the same: architectural sprawl, unclear ownership, endless dependencies, and nothing that survives production scrutiny.
The rebuttal that changed outcomes:
Transformation starts with pain, not ambition.
Winning teams begin with one workflow that already hurts. One metric leadership already cares about. One process people are desperate to improve.
Resume screening. Invoice matching. Claims triage. Sales research.
Solve one problem properly, and something shifts. Trust replaces scepticism. Governance becomes concrete. A reusable agent pattern emerges.
Actionable takeaway:
If you cannot name the single workflow causing the most friction today, you are not ready to scale AI.
Mistake #2: We Hid Behind “Safe” Pilots
❌ “Let’s run a six-month pilot.”
✔️ “Let’s prove ROI in six weeks.”
Long pilots feel responsible. They buy time. They reduce political risk. They create the illusion of progress.
They also quietly kill urgency.
I have rarely seen a pilot longer than eight weeks make it to production. Executive attention drifts. Business priorities shift. Technical complexity compounds. The pilot becomes a science experiment no one wants to own.
The rebuttal that forced clarity:
Short pilots expose the truth.
Teams that shipped ran six-week pilots with success metrics defined on day one. Not model accuracy. Not technical elegance.
Business outcomes.
Cycle time reduction. Cost per transaction. Throughput improvement. Risk reduction.
Actionable takeaway:
If you cannot demonstrate tangible business value quickly, the problem is not the technology. The use case is not ready.
Mistake #3: We Waited for Perfect Data
❌ “We need to clean all our data first.”
✔️ “Let’s start with 80 per cent clean data.”
This mistake comes from fear.
No leader wants AI making decisions on imperfect data. Waiting feels prudent. Responsible. Defensible.
It is also fatal.
I have seen organisations spend years preparing data that never becomes “ready,” while competitors ship with imperfect inputs and improve in production.
The rebuttal that unlocked progress:
Data maturity follows value creation, not the other way around.
Agentic AI systems are designed to operate with imperfect information, provided guardrails, human oversight, and continuous learning are built in.
Actionable takeaway:
If data quality is the reason you have not started, you are already behind.
The Turning Point: From AI Assistance to Agentic Workflows
For years, many organisations have confused AI assistance with transformation.
Early-stage AI answered questions, drafted content, or suggested next steps. It looked impressive in isolation. But humans still carried the cognitive load, stitched outputs together, and managed exceptions manually.
The moment complexity increased, the system collapsed.
That is not scale.
That is augmentation with a ceiling.
True scale begins when organisations move from assistance to agentic workflows.
Agents generate near-final outputs. They coordinate subtasks across systems. They invoke tools. They apply policies. They escalate only when judgment is required.
At scale, more than 80 per cent of task execution can be handled by AI, while people retain accountability and decision authority.
This is how AI survives enterprise reality.
Why Architecture Determines Who Scales and Who Stalls
Agentic workflows only work if the architecture beneath them is open, modular, and governed.
This is why IBM Watsonx Orchestrate matters.
Rather than locking intelligence into a single model or assistant, watsonx Orchestrate is designed as an open orchestration layer for the enterprise AI stack.
It enables:
-
Multi-agent orchestration across complex workflows
-
Pre-built, custom, and IT-engineered agents operating together
-
Governed access to models and tools via AI and API gateways
-
Seamless integration with ERP, BI, and automation platforms
-
Model-agnostic flexibility that future-proofs investment
This is not about picking the best model today.
It is about building an execution fabric that survives change.
From AI as a Feature to AI as Infrastructure
In this architecture, assistants still exist. They are simply no longer the centre of gravity.
The centre of gravity becomes multi-agent orchestration, sitting between users and enterprise systems. Agents operate across data platforms, RAG infrastructure, enterprise applications, and automation tools under a shared governance model.
This is the shift from AI as a feature to AI as infrastructure.
And infrastructure, when done right, disappears into how work gets done.
The Enterprise Reality Check
When AI is designed this way:
-
Complexity is absorbed by the platform, not pushed onto people
-
Governance is enforced centrally, not retrofitted later
-
Integration strengthens existing systems instead of fragmenting them
-
New agents can be added without re-architecting everything
This is how organisations move from impressive pilots to durable production systems.
Not by adding more assistants.
But by orchestrating intelligence across the enterprise with an architecture built to last.
Why We Invite Organisations Early
Early engagement is not about buying technology sooner.
It is about shaping outcomes before constraints harden.
By joining early, organisations gain:
-
A private discovery and co-creation workshop focused on one high-impact workflow
-
Hands-on design of an agentic workflow aligned to real systems and policies
-
A six-week execution roadmap with clear success metrics
-
Direct access to patterns learned from over 100 enterprise deployments
This is applied experience, not theory.
The Only Question That Matters Now
The bar for enterprise performance keeps rising. Customers expect seamless experiences. Teams are under pressure to deliver more with less. Governance expectations continue to intensify.
Agentic AI is no longer optional.
If you have identified a use case and want to avoid the three mistakes that quietly kill most initiatives, the next step is deliberate.
Book a consultation.
Start with one workflow. Build it properly. And move from experimentation to production with confidence.
Hard Lessons From 100+ Deployments Across IBM’s Ecosystem What I Wish I Had Understood Earlier – by Steny Sebastian, Principal Data and AI Platforms Over the past year, I have been deeply immersed in enterprise transformations. Not slideware. Not proofs of concept. Real systems operating in production across HR, finance, insurance, sales, and core operations. Many of these lessons emerged through close collaboration with IBM Product Managers, IBM APAC Technical and Sales Leader, IBM Client Engineering, and Octane AI engineers. Working together on real delivery challenges, sharing ...
.webp)
.png)
.png)
.png?width=595&height=341&name=From%20Insight%20to%20Impact%20(1).png)
.png?width=488&height=281&name=Future%20of%20finance%20(1).png)
