Q1. Could you start by giving us a brief overview of your professional background, particularly focusing on your expertise in the industry?

I'm the Founder and CEO of Synottic, where our mission is to put people at the heart of AI. We help organizations prepare their teams for the future, create strong AI governance, and adopt responsible AI practices at scale. Before starting Synottic, I led global learning and capability-building programs at PepsiCo, and I bring over 20 years of experience across technology, learning, leadership development, and digital transformation.

My background combines software engineering, enterprise learning, executive coaching, and AI strategy, allowing me to integrate technical implementation with business outcomes. I have trained more than 50,000 professionals worldwide on AI, leadership, and future skills and regularly advise organizations on enterprise AI capability building. My current research focuses on AI, neuroscience, and human behavior, helping organizations design AI systems that improve human intelligence rather than replace it. I believe successful AI transformation is fundamentally a people transformation supported by technology, governance, and continuous learning.

Q2. Based on your experience with Model Context Protocols (MCPs) and Glean, what are the most common unpredicted points of failure or data latency bottlenecks during real-time data syncs?

The biggest misconception is that latency is caused only by the LLM. In reality, most delays originate before the model even starts generating an answer. In MCP-based architectures and enterprise knowledge platforms such as Glean, the primary bottlenecks include:

• Connector synchronisation delays
• Inconsistent permissions across systems
• Stale embeddings
• API rate limits
• Excessive sequential tool calls

Another common issue is context inconsistency. Different enterprise systems regularly maintain different versions of the same information, making retrieval unpredictable. Authentication refresh failures, network latency between cloud regions, and large document indexing cycles also introduce delays that users perceive as AI failures.

The most effective approach is to treat retrieval as an engineering problem rather than an AI problem. Event-driven synchronization, incremental indexing, intelligent caching, asynchronous retrieval pipelines, and continuous supervision of connector health significantly improve performance. Enterprises also need clear observability dashboards that track retrieval latency, data freshness, failed connector calls, and context quality instead of only monitoring model response time.

Q3. When deploying multi-agent environments where agents cross-validate work, what typically happens to token consumption, and how do you prevent exponential compute cost spikes?

Multi-agent systems tend to use more tokens simply because each time agents interact, they need to reason, share context, and cross-check each other’s work. If you don’t coordinate these interactions carefully, token usage can spiral much faster than the actual value you’re getting. I’ve seen many organizations surprised by how quickly costs add up when agents keep passing around the same information over and over.

The answer isn’t to cut down on the number of agents—it’s about making sure they work together as efficiently as possible. Each agent should have a clear job, only the information they really need, and just enough memory to do the task. Rather than passing around full conversation histories, agents should share concise summaries or references. Lighter-weight models are great for daily tasks, while you can save the heavy-duty models for the tough problems.

I also suggest using budget-aware orchestration—set up your workflows so they’re always keeping an eye on token usage, cost, and response quality. Smart routing, caching, compressing context, and cutting out unnecessary checking steps all help keep costs from spiraling while still producing accurate results. The most successful multi-agent systems strike a balance between being smart and being cost-effective.

Q4. We often hear generic metrics like 'hours saved.' If we are underwriting an investment based on AI-driven margin expansion, what concrete, ground-level KPIs prove an architecture is driving sustainable ROI?

Measuring AI success by 'hours saved' doesn’t tell the whole story, since it rarely leads directly to financial value. Instead, investors should look for real operational and business metrics that show clear, measurable impact.

The most valuable indicators include cost per transaction, cycle time reduction, revenue generated per employee, process automation rate, first-time accuracy, customer resolution time, quality improvement, compliance adherence, and decrease in operational exceptions. AI adoption rates, active usage, employee productivity improvements, and model dependability are equally important because unused AI delivers no return.

I also recommend tracking the ratio of AI-generated outputs demanding human correction, infrastructure cost per workflow, and the percentage of business decisions supported by AI. Sustainable ROI comes from consistent advances in efficiency, quality, and scalability rather than isolated productivity gains. The strongest AI programs connect technical metrics directly to financial performance and business outcomes.

Q5. When an enterprise automates a major workflow like an AI-powered Learning as a Service model, what do the day-to-day maintenance and support costs look like 12 months post-deployment once model drift and API updates set in?

After the first year, AI systems shift from being newly built to running as part of day-to-day operations. At this stage, maintenance is less about big development projects and more about keeping things running smoothly—think platform upkeep, governance, and ongoing tweaks for better performance.
Teams should plan for regular work like checking how models are performing, fine-tuning prompts, updating API versions, reviewing security, maintaining connectors, keeping dashboards up to date, and supporting users. Knowledge sources need to be refreshed all the time, and workflows might need tweaks as your business changes. Plus, organizations have to stay on top of regulatory rules and updates from vendors to make sure everything is still working as it should.

For mature AI setups, yearly operating costs usually land somewhere between 15% and 30% of what you spent getting started, depending on how complex your setup is and how many systems are connected. Companies that build with modular designs, strong monitoring, good governance, and automated testing right from the start find their long-term maintenance costs are much lower than those who go with tightly connected, custom solutions.

Q6. For an enterprise deploying autonomous agents in regulated spaces, what specific technical logs must be captured at the database or pipeline level to satisfy a regulatory auditor looking for an auditable AI workflow?

An auditable AI system requires complete traceability from user request to final outcome. Every interaction should generate a structured audit trail.
Essential logs include user identity, authentication records, timestamps, prompts, retrieved context references, model version, system instructions, tool invocations, API calls, workflow execution paths, confidence scores, human approvals, final outputs, and any post-generation modifications. Database logs should also record data access events, permission checks, document versions, retrieval sources, and retention history.

Just as crucial are governance logs that record things like policy violations, rejected requests, bias detection, security events, and any override decisions. It’s important for organizations to keep audit records that can’t be changed after the fact. Regulators now expect companies to show not just what decisions were made, but also how the AI reached those decisions—and whether the right people were involved at every step along the way.

Q7. If you were an investor looking at companies within the space, what critical question would you pose to their senior management?

My first question would be: When foundation models get even stronger, cheaper, and more common, what unique advantage will your company still have that sets you apart from the rest?

The world of AI is always changing, and any technological edge usually doesn’t last long. What really sets a company apart over time is having unique data, deep industry know-how, strong customer relationships, solid governance, seamless workflow integration, and—most importantly—the power to deliver real, measurable business results.

I'd also want to see if the company actually has strong AI adoption metrics, not just flashy demos. Are customers using the solution more over time? Is it really woven into day-to-day business operations? Can the company reliably boost productivity or profitability? On top of that, I’d look at whether the leadership team can adapt as the AI world keeps changing. In this space, long-term winners are those who deliver results, create real value for their customers, and stay flexible—much more than just having access to the latest technology.