santosh rachamalla

Industry: Property & Casualty Insurance Process: End-to-end insurance claims (from FNOL: First Notice of Loss to Settlement) In most insurance companies I’ve worked with, the conversation about claims delays usually starts — and ends — the same way: “We don’t have enough adjusters.” On the surface, it makes sense. Claims are piling up, cycle times are long, customers are unhappy, and adjusters are clearly stretched thin. So leadership reacts by hiring more adjusters, outsourcing work, or pushing harder on productivity targets. But in reality, that’s often not the real problem. The Constraint People Commonly Miss In one large auto and property claims organization, we dug deeper and realized the real constraint wasn’t adjuster capacity at all. It was decision delays caused by rework and unclear policies. What was actually happening looked like this: Claims were coming in fast — FNOL was already well automated. Adjusters were picking up claims quickly. But a large number of claims kept getting stuck waiting for: Supervisor approvals Coverage clarifications Missing documents from customers Estimates from body shops or inspectors Every time a claim paused, it created side effects: Adjusters had to context-switch Claims got reassigned or reopened “Simple” claims came back multiple times Many claims were touched 3 to 5 times before resolution. That created a huge amount of hidden work that never showed up in staffing models. From the outside, it looked like adjusters were the bottleneck. In reality, the system itself — policies, approvals, and variability — was slowing everything down and multiplying the workload. Why Humans Have a Hard Time Seeing This Even very experienced claims leaders struggle to spot this kind of constraint, for a few reasons. First, the work is fragmented. No one sees the full claim journey end to end. Adjusters see their own queues, supervisors see their teams, and operations looks at averages and dashboards. The delays live in the cracks between those views. Second, the metrics are misleading. Average handling time hides how uneven the work really is. Productivity metrics reward “touching” claims, not resolving them. Reopened claims often aren’t treated as new demand, even though they consume just as much effort. Third, there’s natural human bias. We focus on what’s visible — busy people and growing queues — and overlook what’s invisible, like waiting states, policy friction, and handoffs. Humans are good at reasoning about cause and effect. We’re not great at doing it across millions of transactions and tiny delays that add up. How AI Can Help Identify the Real Constraint This is where AI can genuinely add value. 1. Seeing the Real Process, Not the Designed One Using system data, AI can reconstruct how claims actually flow: When FNOL happens When claims are assigned Every touch, pause, approval, and reopen That makes it very clear: Where claims spend most of their time waiting How often they move backward Which rules or decisions trigger repeated rework In one case, this analysis showed that about 25% of total cycle time was spent waiting on clarifications tied to just three policy rules — something no one had explicitly called out before. 2. Exposing Hidden Workload Inflation AI can also estimate the true workload by accounting for: Reopens Rework The probability that certain claim types will bounce back Humans tend to plan capacity in straight lines. AI shows how small increases in ambiguity or exceptions can cause backlog growth that’s completely out of proportion. 3. Testing “What If” Scenarios Safely AI can simulate questions like: What if fewer claims required supervisor approval? What if documentation requests were triggered earlier? What if very simple claims were auto-settled? Those simulations help answer an important question: does the constraint actually move? If adding adjusters doesn’t improve cycle time in the model, staffing isn’t the constraint. If changing one rule cuts reopens by 30%, the bottleneck was policy — not people. This kind of experimentation is almost impossible to do safely in the real world. Where AI Clearly Beats Humans AI is especially strong at: Finding hidden waiting time Quantifying rework loops Identifying policy-driven bottlenecks Seeing patterns across huge volumes of claims Challenging long-held assumptions with data Where Humans Still Matter More AI still struggles with: Understanding why a policy exists (regulation, fraud risk, brand protection) Deciding whether a change is acceptable, even if it improves speed Anticipating how people will change behavior once metrics shift Balancing speed with fairness and customer trust AI can tell you what is constraining the system. Humans have to decide whether and how to fix it. Final Thought In insurance claims, the most damaging constraints usually aren’t the obvious queues. They’re buried in policies, handoffs, and variability that quietly slow everything down. AI doesn’t replace human judgment. It forces uncomfortable clarity about where work and value are actually getting stuck. And from what I’ve seen, that clarity is often the turning point for real process improvement.

AI-to-AI Collaboration in Banking: What Mortgage Processing Might Actually Look Like I’ve been spending time thinking about what happens when AI systems don’t just assist humans inside one organization, but actually talk to other companies’ AI systems to complete an entire business workflow. Mortgage lending seems like the most obvious place where this will become real first. Not in a futuristic way—just in a very practical, operational sense that product and business teams will end up building toward. I wanted to share how I’m thinking about it and get reactions from others who are exploring similar ideas. A Practical Scenario: AI Agents Coordinating a Mortgage End-to-End Right now, a mortgage application requires coordination across multiple organizations—banks, credit bureaus, payroll providers, title companies, insurers, and even public records systems. Most of this is still handled through emails, PDFs, and humans chasing down missing information. In the near future, I see this moving toward AI agents acting on behalf of their organizations and coordinating the same steps with far more precision and speed. Here’s the workflow I imagine: 1. Customer applies → Bank’s mortgage AI kicks off the process Bank A’s AI assistant (“LoanBot-A”) reaches out to: Credit bureau AIs for credit data Payroll/HR AIs for income and employment verification Other banks’ AIs to confirm assets (with customer consent) Title AIs to check ownership, liens, and disputes Insurance AIs for property coverage and risk Public Records AIs to scan for active legal issues Fraud/AML AIs for identity and risk assessment But instead of PDFs, these systems exchange structured, machine-readable messages—almost like business email threads, but optimized for machines. How These Conversations Might Actually Look To make this more tangible: LoanBot-A → PayrollAI: “Customer consent validated (token good for 8 hours). Please provide monthly gross income, bonus history, and a 12-month employment continuity estimate.” PayrollAI → LoanBot-A: “Confirmed. Gross income: $11,200/mo. Bonus range: 8–14%. Employment stable with 94% continuation probability. Flag: quarterly bonus volatility.” Or: TitleAI → LoanBot-A: “Ownership verified. One past lien (closed last year). No open disputes. Structured summary attached.” What takes 30–45 days today could realistically shrink to a few hours. Collaboration + Competition in a Shared AI Ecosystem What’s fascinating is that this model is both: Highly collaborative All lenders, AIs, and data providers must interoperate to complete the mortgage. Highly competitive Each bank still differentiates through: speed of underwriting quality of risk modeling customer experience pricing sophistication The analogy that keeps coming to mind is telecom: Everyone interconnects, but they still aggressively compete. Risks (and the mitigations we will need) As exciting as this is, there are serious considerations: 1. Errors can propagate extremely fast If one AI misinterprets income or debt, that mistake could flow into every downstream agent. Mitigation: multi-agent validation, confidence thresholds, human review triggers. 2. Two “black box” models agreeing without explainability That’s a governance and audit nightmare. Mitigation: explainability requirements structured “reasoning summaries” logs for internal and regulatory audit 3. AI identity & security A spoofed AI agent could cause major damage. Mitigation: cryptographic agent identity short-lived consent tokens zero-trust interfaces 4. Unintentional pricing coordination If AIs observe each other’s moves, lending markets could start behaving in synchronized ways. Mitigation: randomness in exploration, regulatory transparency access. 5. Larger privacy footprint More cross-system collaboration = more risk exposure. Mitigation: purpose-bound queries auto-purge after closing customer-facing audit trails (“who accessed my data and why?”) Questions I Keep Asking Myself As I map this out, a few internal questions keep coming up: What communication standard will all these AI agents eventually use? Will we see one unified protocol emerge, or a patchwork of industry-specific formats? How do we design AI systems that collaborate on data exchange but still maintain competitive differentiation? Especially in underwriting logic and pricing. What level of oversight will regulators expect? Do they eventually want read-only access to AI-to-AI communication logs the way they audit emails today? What happens when a key ecosystem player refuses to join this model? Do we need fallback paths, or will market pressure force participation? Does this ultimately create a winner-take-most environment? Where lenders with the smartest, fastest AI underwriting engines gain a disproportionate share of the market? These are the areas I’m still working through as I think about how ecosystems will evolve once AI becomes both the collaborator and the competitor.

How Transparent Should AI Be Across an Entire Pharma Ecosystem? (My 2 cents from inside a biotech commercial team) So I work in a biotech commercial org where we rely heavily on omnichannel and all this “Next Best Action” stuff everyone’s buzzing about. Basically our AI is always trying to figure out: which HCP, what channel, what message, and when… using past behavior, macro-industry events, engagement trends, all that. Sounds clean on paper. In reality, it’s like a dozen different AI systems—ours + vendors + media agencies + field tools—trying to make decisions at the same time without fully knowing what the others are doing. Kind of like having a bunch of “smart” traffic lights that don’t talk to each other. That’s where the transparency question hits hard. How much do we actually need to share? Personally, I don’t need to know anyone’s secret sauce. I don’t need algorithms, model weights, IP, any of that. But I do need to know the intent of the decision. Like: Why was this HCP flagged today? What signal pushed this moment? Why a rep visit vs an email? Why this message over that one? When AI just spits out “Call this doctor now,” it doesn’t feel like intelligence—it feels like nagging. Especially for field reps. Field force orchestration is where transparency matters the most Reps are human. They have context the AI doesn’t. If they don’t trust the recommendation, it’s dead. They’ll ignore it and go back to their own strategy. And honestly, reps are a goldmine of qualitative insight: “This doc hates morning calls.” “This clinic has been slammed lately.” “Dr. Z is only responding to digital right now.” AI needs that stuff in the loop. But reps aren’t going to provide feedback if the system feels like a black box. Give them simple reasoning—nothing fancy—just enough for them to go: “Okay, makes sense.” Digital media is its own wild west The paid media algorithms are out here doing their own thing… meanwhile, CRM and email suppression rules are doing something else entirely. And then we’re all surprised when an HCP gets bombarded on one channel and ghosted on another. A little transparency across partners would go a long way. Again, not trade secrets. Just: channel saturation signals, predicted responsiveness, what’s currently running or paused. Otherwise it’s algorithm vs. algorithm, with the customer stuck in the crossfire. And then there’s HCP support + content recommendations Medical info portals, chat assistance, patient support services—these all have AI baked into them now too. HCPs aren’t asking for deep AI explainability. They just want the “why” behind what they’re getting: “Why am I seeing this content?” “Why did you prioritize this topic?” “Why this format?” Basic stuff. A bit of transparency here actually builds trust. What about being transparent with HCPs or patients themselves? This is where Pharma gets nervous, but I think we actually need more openness. Not a creepy, “We tracked your clicks for 18 months.” More like: “We recommended this topic because you showed interest before.” “We adjusted your cadence based on your preferences.” “If you want X instead of Y, you can update it here.” It’s respectful. And honestly, HCPs appreciate when we don’t pretend everything is random. The feedback loop is the part we’re missing the most Right now, feedback is basically: a rep note, a “Was this helpful?” click, or some unstructured survey input. But imagine if: HCPs could say “send me more of this, less of that” reps could flag when the AI missed context patients could rate which support steps were actually useful content topics could be voted up/down like Reddit … and the AI actually learned from it. That’s how it becomes a shared system, not a top-down command machine. But again, transparency is what gets people to even give that feedback. If they feel like it disappears into a void, they won’t bother. And the boundaries? We still need them. Pharma is regulated. Some things absolutely cannot move across orgs: no personally identifiable data sharing, no detailed targeting logic, no promo triggers that look like profiling, no partner-to-partner leakage of commercial strategy. The AI ecosystem should be transparent enough to collaborate but still siloed enough that privacy, compliance, and competitive advantage are protected. We don’t need “show me your code” transparency. We need “help me understand what you’re doing and why” transparency. Especially: for field reps, across digital channels, within HCP support, and even directly with HCPs/patients in a respectful way. And if we want the system to actually improve, we need to build feedback loops that people trust enough to use. That’s my take. Curious how others are handling this mess in their orgs.

Process Chosen: Omnichannel Marketing for HCP Engagement In pharma marketing, we often reach HCPs through multiple channels—emails, texts, social, portals—alongside in-person rep visits. While the intent is good, the reality is that many customers feel overwhelmed. Similar messages in different formats from various pharma can create fatigue, and the most relevant information sometimes gets lost in the noise. Often times marketers also fell in the trap of delivering more content, more creative and more channels instead thinking from Customer perspective. For eg, How many other pharma marketers send the emails/messages/brochures reg their products. For an oncologist or any specialty, 90% of those messages goes unread. Here’s where AI can really help—but only if it’s done thoughtfully. Instead of acting as a separate “digital rep,” AI should work as a partner to field teams, not a competitor. All the information between digital and sales rep activities should be consolidated and orchestrated to create an efficient and valuable customer journey. Imagine AI helping reps by: Predicting what content matters most for each HCP and timing it right. Ability to gather feedback on relevancy as well. Consolidating messages so they feel personalized, not repetitive. With AI and automation it's not the volume that matters, it's the relevancy and timing matters. Measuring true customer experience—not just clicks or opens, but how supported and understood the HCP feels, gather feedback if the content is insightful or met the need of their hour. The risk? If AI oversteps—sending too much, misreading intent, or replacing human touch—it erodes trust. HCPs value first-hand info from reps and the ability to ask questions. So, we need balance: Use AI for efficiency and low-complexity tasks. Keep humans front and center for strategic or nuanced conversations. Offer an easy path to connect with a real person when needed. Ultimately, tech should fade into the background while customer experience stays in the foreground. If we combine smart segmentation, CX frameworks, and real-time human support, AI can strengthen relationships instead of straining them.

The objective of any technology is to meet business goals in an efficient and secured manner. When it comes to AI systems, the accuracy of outcomes, timeliness and the relevancy of models behind the scenes to the business needs at that point of time. If accuracy keeps dropping despite fine tuning, or if data and concept drift make predictions unreliable, it’s time to take a hard look. High maintenance costs compared to the value delivered are another red flag. Sometimes, the issue is simply technological obsolescence; newer models may offer better performance and easier integration. If the system no longer supports current objectives or creates friction with integrating with cross platforms, that’s a sign too. Bias or lack of transparency uncovered during audits is a serious concern, and losing user trust is the ultimate deal-breaker.