The Second Half of Agentic AI

Why a framework to build agents is necessary, but not the same thing as AI doing insurance work.

Almost every agentic AI announcement in insurance this year has described the same thing: a framework to build, deploy and govern agents. Model-agnostic runtimes. MCP servers. Low-code studios. Developer assistants. Registries and audit trails.

This is real progress, and it matters. An insurer that cannot build, govern and observe its own agents has no durable AI program.

But a framework is the means, not the outcome.

It answers the question “how would we build an agent?” It does not, on its own, answer the question a claims leader or a chief underwriter actually asks: “what work is the AI doing for my team on Monday morning?”

This is an AI business value gap.

That gap between a platform to build agents and AI already doing insurance work is the second half of the agentic story. It is the second half that is harder to demo, harder to build, and far more valuable to the business.

It is where Majesco chose to start.

The framework is becoming table stakes

It is worth being clear-eyed about the framework itself, because Majesco provides one and we are not arguing against it.

A credible agentic platform needs a way to define agents without hard-coding them, a way to expose enterprise systems as governed tools, a way to route across models, and a way to see and control everything at runtime. Majesco’s AI Core provides these: a versioned agent and tool registry, a model catalog that selects the right model per task, an MCP gateway that publishes governed insurance tools with allow-listing and access control, a low-code Agent Studio, framework APIs for code-first teams, and an SDK to embed the experience into any application.

None of that is the differentiator, never was and never will be.

Industry analysts have started saying so directly. Celent’s recent architecture research frames the shift precisely: agentic AI is “an inversion of control,” where the AI is no longer inside a hard-coded software loop but is itself the loop, calling classic services and databases as tools. In that same research, the recurring conclusion for incumbents is not that they need a greenfield agent platform. It is that the assets they already have, well-governed APIs, a modern core, identity and access controls, become “load-bearing” in the agentic era. The agent sits above the systems of record and calls them; it does not replace them.

If the framework is becoming a shared expectation, then the question that separates AI programs is no longer “can you build agents?” It is “how quickly does the AI start doing real work, and how much of that work did you have to assemble yourself first?”

Primitives are not a process

Here is the distinction that gets lost in most agentic pitches.

A framework gives a carrier the primitives: a way to make an agent, a way to give it a tool, a way to watch it run. What it does not give is the insurance process itself. Someone still has to decide what the underwriting-intake agent should extract, how the appetite check should reason, when an exception should route to a human, and how the whole sequence carries a submission from intake to a bindable decision. That design work, the assembly of primitives into a governed insurance workflow, is most of the effort and almost all of the risk.

This is why so many agentic pilots stall. The framework works. The demo works. And then the carrier discovers that turning a demo into a production workflow across underwriting, policy servicing, billing and claims is a multi-quarter build, repeated for every process, before a single measurable outcome appears.

The result is what many teams now call pilot-and-POC sprawl: a lot of agents, a lot of governance, a lot of resources expended, and not much real insurance work actually being done or real business value achieved.

Majesco has the framework. But Majesco’s focus and approach started on the other side of that build.

We deliver ready-to-use agentic workflows: 22 P&C agents available today, expanding to 36 by Fall ’26, alongside a growing L&AH portfolio, built on more than 100 governed insurance tools and grounded in the Majesco Intelligent Core. These are not primitives waiting to be assembled. They are insurance work already assembled into workflows, governed on the same core a customer would use to build their own. A carrier can put AI to work first, creating business value and organizational trust and acceptance, then extend and build, rather than building for a long period of time before any work gets done.

Immersive AI: the experience the framework does not produce

There is a second important thing a build framework does not give you, and it is the one I care about most.

Most agentic architectures still treat an agent as a thing you invoke: you ask, it acts, it returns. That is useful, but it is not how skilled insurance work happens. An underwriter assessing a complex commercial submission is not issuing commands to a bot. They are holding a case in their head, asking “can I proceed, what is blocking me, what changed, what should I do next,” and moving fluidly between judgment and analysis.

Immersive AI is Majesco’s answer to that. The user and multiple AI capabilities work the same case together. As a commercial submission moves from intake through assessment to decision, the AI reads and classifies the submission, validates the risk data, checks appetite, assesses the risk and loss profile, and surfaces coverage and pricing considerations with confidence and rationale. The underwriter never has to decide which agent, which model, or which capability to invoke. They review, adjust where their judgment differs, and make the call. The AI does the gathering, analysis and preparation. The human keeps the judgment and the decision.

Getting there requires confronting something the industry has been slow to admit: the core system’s user experience was designed for human-only work.

Every screen, form and navigation path assumes a person is doing each step. Drop agents into that same interface and the design starts working against you, because a UX built to walk a human through twenty screens is the wrong surface for work an agent has already done. Human-agent workflows need their own design, one where the AI carries the case forward and the interface is organized around review, exceptions and decisions rather than data entry.

This is why Immersive AI is an approach, not a feature. It is a rethink of the operating surface itself, so that human and agent are working the same case in a shared space rather than the human operating software the agent happens to also touch.

Designed this way, the experience changes what the human is for. When the AI handles the gathering, the reconciling and the first-pass preparation, the person is freed to do the work only a person can do: exercising judgment on the hard cases, guiding the AI where context is missing, owning accountability for the decision, and building the relationships with brokers, policyholders and colleagues, that no agent will ever own.

The point of taking work off the human is not to remove the human. It is to move the human up to the parts of the job that actually need one and create trust and value to the customer and business.

There is a second, quieter benefit, and it may matter more for adoption than any single capability. When AI arrives inside the tools people already use, they do not have to learn a new way of working to get value from it. Adoption stops being a training program and starts being the natural way the existing job gets done. That is how AI scales across an organization and delivers value faster: not by asking thousands of users to migrate to a separate AI tool, but by bringing the AI to where they already work. The experience meets people where they are, which is precisely why it spreads.

That experience is not something a framework emits as a byproduct. It has to be designed, from the outcome backward, and embedded into the operating surface where the work already happens. It requires business knowledge and expertise. It is the hardest thing to replicate and, for the people doing the work, the most valuable.

One core, or two disconnected products

The strongest argument for how Majesco is built is also the least visible on a feature and function list: the ready-to-use agents and the agents a customer builds run on the same AI Core.

Same execution engine. Same orchestration. Same insurance context and governed tools. Same systems of record. Same governance spine, identity, entitlement, approval, attribution, audit, telemetry and metering, applied uniformly to every agent, model and tool. A capability defined once is reused by Majesco-built agents, by customer-built agents, and by external AI clients reaching Majesco tools through MCP. Improve a tool once and it improves everywhere it is called, under the same controls.

This matters because the alternative, a set of pre-built agents on one stack and a build-your-own framework on another, forces carriers to reconcile two governance models, two audit trails and two integration surfaces.

Celent’s work on identity in agentic systems makes the cost of that fragmentation concrete: when a decision is questioned, an insurer has to answer who did what, on whose behalf, and within what limits, and it has to answer across every agent in the chain. That is only tractable when there is one governed core underneath, not two.

Where this leaves an insurer

The agentic framework question is largely settled. Carriers should expect a model-agnostic runtime, MCP, a registry, a studio and full observability from any serious core platform, and they should hold Majesco to that standard. We meet it.

But the differentiating question for a core platform is about delivering governed, insurance AI with ready-to-use agentic workflows.  

To accurately separate AI programs and how they are different, and they are the ones I would put to any vendor, including Majesco is:

• How much insurance work is already assembled into governed workflows I can turn on, versus primitives I have to assemble myself?
• Is there an experience where my people and the AI work the same case together, or only a bot I invoke?
• Is the AI where my people already work, or does adoption depend on them learning new tools? 
• Do ready-to-use agents and the agents I build share one governed core, or am I reconciling two products?
• Does this run across my business, P&C, L&AH and Retirement & Pension, on one foundation?

A framework is how you build. It is necessary.

But the business value shows up when the AI is doing insurance work, in the flow, governed, on day one, and when the experience is good enough that the people doing the work would not go back.   

That is when trust and adoption accelerate.  That is when business operating metrics improve.  That is when real business value is achieved.  That is when the promise of AI becomes reality.

That is the second half of agentic AI. It is the half worth measuring, and it is where the next round of the AI conversation should be.