This article is relevant if you are a COO, CFO, or CIO and you are seeking practical guidance on how to address fuzzy business problems with NetSuite-enabled AI tools while ensuring cost-effective, trustworthy outcomes.
TL;DR Summary
This article explains why business leaders (particularly COOs, CFOs, and CIOs) must move beyond hype and understand how NetSuite-enabled AI tools can transform fuzzy, high-cost business problems into measurable advantages. It walks through the progressive journey of adoption, the lessons learned, and provides a framework (including the Prolecto AI Labs Learning Table) that illustrates the pathway from experimentation to trusted operational deployment.
Background
Most readers are aware that I have not rushed to discuss Artificial Intelligence or the surge of Large Language Models (LLMs). Yes, we have been asked, and I have responded to clients judiciously. The marketplace has been saturated with hype; everyone has an opinion, and too often those opinions lack substance. Our NetSuite Systems Integration Practice is built on trust, competence, and grounded delivery. That means we do not join a hype cycle just to appear modern. Only now, after making material investments in this technology, learning from both mistakes and successes, and putting real tools into production, do I feel we have earned a position to share our journey.
This article is intentionally longer than our usual publication length. It is written for the uninitiated business leader and reader who is evaluating whether to commit resources toward AI enablement. My goal is to illustrate the pathway, expose practical learnings, and help you see how previously considered too costly to address, fuzzy business problems can now be tackled in ways that produce measurable advantages.
I naturally expect the costs for getting real-world applications to drop. However, the patterns outlined here are fundamental. Thus, we better prepare for the journey and start to reap the benefits, because if we don’t, others will and we will find ourselves competitively marginalized.
The Fuzzy Problem Domain
Every organization faces what I call the Fuzzy Problem domain. These are challenges that resist easy structuring:
- Institutional knowledge that is scattered across documents and emails.
- Customer interactions buried in notes and tickets (think NetSuite Cases).
- Decision-making that requires context from multiple systems (think NetSuite transaction data and support file documents).
In the past, most firms simply lived with fuzzy problems. They just were too expensive to solve. Only when a problem could be structured and reduced to rules did we attempt to build an automated solution. Firms that successfully transformed fuzzy problems into structured practices often gained a competitive advantage. Things are now changing.
What is exciting about LLMs is the lowering of the cost curve. Problems that once were considered prohibitive can now be revisited. But that does not mean solutions come for free — that’s the hype. They demand investment, deliberate design, and iterative refinement.
Our Prolecto Journey: Building Competence Step-by-Step
We call ourselves a Practice because we recognize the need to continually develop our competence and craft. Thus, here we are.
Led by Chidi Okwuidure and our Technology Practice, and backed by a Nuri H., an outstanding Systems Architect, we committed to learning this space the hard (earned) way. We earned it through solving progressively more difficult applications. We recognized that Oracle is developing its own models, but the integration with NetSuite is still in its early stages and limited. We are confident Oracle’s investments will bear fruit, especially as the Model Context Protocol (MCP) gains adoption, but we could not wait. Instead, we used Amazon’s infrastructure because it offered everything required to advance our competency.
Security and data privacy are important concerns. The major infrastructure players are building capabilities in this area, but we did not aim to solve those policies ourselves. We focused instead on how to tackle fuzzy problems with practical business data. We will let the market evolve here and take guidance from other leading experts.
The following applications illustrate our progression. Each was a step up in difficulty, and each produced insights that shaped our competence.
App #1: Blog Post Summarization
This was our entry-level experiment. We crawled a single blog post on this blog, parsed the text, and passed it through AWS services for summarization. The outcome was modest; it got the gist of an article, but it accomplished something critical. We gained exposure to the infrastructure, tested data pipelines, and confirmed that we could control the output.
The point here was to ensure we connect the infrastructure (beyond Hello World) and begin to understand what the tools need to do their work.
Click on the images to see the application and the architecture in full screen.
App #2: Ask the Blog
The second initiative was more ambitious. Over the last decade, Prolecto has produced one of the most comprehensive bodies of meaningful NetSuite content in the marketplace. We are pleased to be recognized as the #1 resource in the community according to Feedly and Reddit. The problem that everyone recognizes is that blog structures bury older material. Readers rely on category clicks or keyword searches, which are blunt tools. Yes, putting a better search on it helps, but we are moving away from search and toward a prompt. Thus, we need to think and act differently.
With Ask the Blog, we considered a different question: why not load all our articles and allow a curious reader to simply pose a question? Isn’t that more intuitive? The system would return a coherent response and, critically, link back to the original article. This reinforced our authority while making knowledge instantly accessible. We can shape our response to meet the standards we strive to uphold in our service delivery. This was a meaningful objective.
We learned about retrieval-augmented generation (RAG) and its associated limitations. The models are good at “sounding right,” but without careful structuring, answers can mislead. We discovered that defining categories, filtering out invalid questions, and establishing requirements up front were all essential.
Click on the images in this section to see Ask the Blog in use and the architecture behind it.
App #3: PTM Assistant
The third initiative moved us into client-related data that is housed within our NetSuite instance. We reskinned NetSuite Cases into what we call PTM (Prolecto Task Manager). The idea was to parse these PTMs (NetSuite case messages or tickets), extract metrics, and confirm alignment with practice definitions. This is now one of those Fuzzy problems. It is similar to what we witness in email and is effectively “all over the place”.
The lesson here was stark: business data in this domain is messy. Because we never thought the fuzzy problems would be solved in the days of old, we had to get good at clarifying what we cared about (i.e., the requirements). Definitions that seemed obvious turned out to be inconsistent as we began to articulate our thoughts. We learned that noise in the domain data must be managed; otherwise, the signal is lost. Think about the noise in email communication — the chit-chat going back and forth, and finally the discussion on the substance. All of that chit-chat is noise. The PTM Assistant helped us see the importance of data preparation, validation, and iterative feedback when dealing with real operational systems.
App #4: PTMly Chrome Extension
With PTM Assistant, we were now armed. The most advanced initiative so far is PTMly, a Chrome extension that rides directly on top of PTM (our reskinned NetSuite Case system). Everyone in our Practice (analysts and managers alike) uses our PTM system daily, but the quality of case titles and abstracts varies widely. Some are terse. Others are incomplete. This creates friction in reporting and handoffs. If we can make it better, it will help us build greater client trust, and it will help lower the cost of being and staying focused.
PTMly helps change that. It reviews a PTM, intelligently suggests improvements to the title and abstract, and allows the analyst to accept or adjust. The result: consistency, clarity, and efficiency.
This was the leap from experiment to production. We confronted the reality of noisy inputs, built safeguards against inappropriate outputs, and leveraged iteration with user feedback. We also discovered firsthand the strength of browser extensions: they deliver application augmentation without requiring a rewrite of our native App. Once we become satisfied with the model and output tuning, the same language model services can later be embedded into our applications more natively.
Here too is what is essential. The talk today is about AI being agentic (acting on behalf of the user). We now know that with more tuning, we can get this system to work behind the scenes for us without analyst interaction. However, we must build trust in what it is doing, and iterative refinement is clearly the demand in this domain to succeed.
Click the images in this section to see PTMly in use and the architecture diagram that powers it.
Key Learnings From the Journey
The following table outlines our learning. It reinforces the grounding of what we have earned on this journey.
Prolecto AI Labs: Progressive Learning Table
| Initiative/Project | Objectives | Key Learnings | Key Technical Elements | Deliverables | Status | Open Items / Ideas / Limitations |
|---|---|---|---|---|---|---|
| 1. Blog Post Summarization | Get the gist of a single blog post article. The idea is that an article could be fed to produce an abstract summary, which could be valuable when we distribute the article. |
|
We crawled the blog, parsed the text and passed it to the LLM (Titan) with appropriate prompts. We worked with the basic mechanics of the AWS services. | Standalone Research & Development Summarize Blog tool |
|
Shall we integrate it with the blog as a possible tool for users to land on the page and see a summary? The initial idea was to integrate this into the website and make it publicly available. |
| 2. Ask The Blog | Answer free-form questions using all the blog articles (not just one) as a knowledge base |
|
Filter out invalid questions by categorization. For valid questions, vectorize article texts, find nearest “neighbors”, prompt LLM using the top X relevant articles to get answers (i.e., RAG approach).
Note: Vectors are stored in-memory (backed by an S3 bucket) for cost reasons (AWS serverless incurs a base cost) A fixed number of articles is used to drive the answers i.e., the entire blog repository is not used |
Standalone R&D tool with sentiment analysis (based on categorization) | – Light qualitative assessment of results (we would need to do more testing and fine-tuning if we want to deploy this for usage) | Exploring self-tuning of models may be interesting.
Considering what it means to be up and operational with a growing set of knowledge through new articles and reader comments What may happen if we connect it with other trusted sources of information? |
| 3. PTM (Prolecto Task Manager; our skinned version of NetSuite Cases) Assistant | Given a PTM, extract key metrics, confirm alignment (subject/abstract), and extract accomplishments |
|
|
Standalone PTM AI Assistant tool that can use AWS or OCI (via N/llm) as backend.
Note: API tokens are required to run the application. |
Ready for internal usage within the Prolecto team, preferably integrated with our PTM tool so that it showcases the potential. | We have not conducted a qualitative assessment that is required for trustworthy production use.
Case data contains a lot of “chatter” about various social interactions that are effectively not relevant to the specific PTM/case information. This is effectively noise that needs to be diminished in meaning. We are currently not cleaning out email threads and have not assessed the effects on the results (we assume threads may bias the answers in some cases due to repetition) |
| 4. PTMly | PTMly is a Chrome browser extension that provides AI-powered improvement suggestions for the Title and Abstract of a PTM (Project Task Management) in NetSuite. It is a product of the Prolecto AI Labs project.
The goal is to improve our PTM/Case standards for understandability, which will help deepen client trust, enable our team to work with more confidence, and provide a clearer view of status. |
|
|
Chrome Extension (PTMly) that can be installed and targets our specific PTM system. | Deployed and actively in use by the internal team. Watching the metrics and improving the prompts. | Considerations to work now against the native NetSuite UI to help clients understand that it is not limited to our custom NetSuite Case skin.
Framework for centralized Prompt refinements, as this is an ongoing pattern to succeed with these tools. |
Important Takeaways for Those Contemplating the Journey
Through these projects, we captured insights that apply to any firm contemplating AI in NetSuite:
- Clear objectives matter. Without defined outcomes, the work wanders.
- Noise reduction is essential. Domain data is messy; plan for it.
- Safety guardrails must be built in. Consider the risks and do not leave it to chance. Learn the emerging market standards for implementation planning.
- Iteration is the new normal. These systems evolve based on user feedback, not just traditional internal QA. User engagement is paramount in tuning.
- Browser extensions accelerate adoption. They enable augmentation without requiring system rewrites or surgical enhancements. We can ride along, learn and then take back the best thinking to traditional application enhancement.
- Data access strategy is critical. Decide upfront how to feed the models: real-time or batch, and prepare to stitch it up.
- Prompt refinement is a discipline. Build infrastructure for ratings and adjustments. We must listen, refine, and stay organized.
Implications for NetSuite Leaders
The message for CFOs, COOs, and IT leaders is straightforward. LLMs can indeed unlock competitive advantage, but only when pursued with investment, rigor, and a pathway grounded in reality.
Oracle will continue to expose more AI capabilities in NetSuite, and we welcome the progress. But if you want to move today, it is possible. We have walked the path. We know firsthand the challenges and understand the competencies required. It can be realized, but it will require patience and commitment.
Considerations for Embarking on the AI Enablement Journey
If your organization is evaluating AI enablement with NetSuite, this article shows you the journey. Please forgive me again; the article is longer by necessity because the uninitiated must understand the progression. Competitive advantage is not handed to those who chase trends or believe in silver bullets. It is earned by those who invest in competence, step by step, until the value becomes undeniable.
The hype cycle is loud, but it does not build trust. What builds trust is demonstrating competence. At Prolecto, we have moved from simple experiments to production deployment. We have addressed fuzzy problems in real business contexts. We are continuing to invest internally, assembling data from multiple sources to address the classic knowledge base problem that has long resisted a solution. I anticipate producing some videos on these achievements to help inspire others.
If you found this article relevant, feel free to sign up for notifications of new articles as I post them. If you have a fuzzy problem that, if solved, will make a significant difference, and you are ready to address it today, let’s have a conversation.
