Editor’s note: The following is a guest post by Rohit Kapoor, chairman and CEO of AI data platform and services provider EXL.
There is a version of AI adoption that looks transformative from the outside and changes almost nothing on the inside. A summarization tool here. An assisted drafting feature there. Automated nudges layered onto workflows that were designed for a different era and never fundamentally reconsidered. It’s like a digital veneer was added to an otherwise analog business model, and it’s becoming increasingly common.
McKinsey recently found that despite significant enterprise AI spending increases, growing use of AI across functions and widespread reports of individual workers increasing their productivity, 37% of businesses have yet to link bottom-line financial gains to the technology. AI is being used in pockets while organizations struggle to scale projects across core workflows.
The scaling problem hinges on change management. Technology acquisition is just the first step in a process that requires organizations to put existing workflows under the microscope and find ways to reinvent legacy processes that are no longer relevant or necessary. ROI from enterprise technologies challenges organizations and their partners in the channel to assess underlying operating models.
Rigid model headaches
One of the first decisions most businesses encounter when they apply AI within enterprise workflows is whether to take a centralized or federated approach to deployments. Each has advantages and pitfalls.
In a centralized model, a single team, typically a center of excellence reporting to a CIO, CTO, CDO, or chief AI officer, owns AI for the entire enterprise. The team handles most or all of the lifecycle, from identifying use cases, building and training models and managing data pipelines, to setting governance and risk controls, deploying into production and maintaining what's live. Business units bring requests to this team rather than building AI capability themselves.
While centralization brings a level of consistency and standardization to an AI roll-out, it can also create bottlenecks because every business function needs to be routed through a team that may lack the domain expertise required to manage the details of specialized operations.
The alternative is to leave each various departments within an organization to develop an approach that works best for its goals. There are downsides to this approach, including tool duplication, app sprawl, shadow AI and inconsistent controls. Most organizations swing between these two modes without finding a third option.
That third option is a federated model. More than just a compromise between the extremes, the federated approach allows independent ownership over interconnected capabilities built atop a shared foundation. Accordingly, this model preserves independent ownership while removing the need for independent reinvention.
Federated models put to the test
The insurance industry is the ultimate real-world test bed for the federated approach. Heavy on repeatable actions and systematized operations, but also layered with dozens of nuances, insurance claims workflows require both centralized and decentralized functions to work properly. Failure to recognize these two different scenarios and design AI operating models to handle each of them differently is where many enterprise AI initiatives are failing.
Certain transactional processes, such as the claims management workflow, involve the same steps run thousands of times a day against a defined set of rules. These are ideal candidates for AI embedded directly into the workflow. The business specialists — not the IT team — need to identify the use cases, because only the people running claims know where the real friction, leakage and customer pain sit. They can design new processes rather than automating the old ones and add new roles, new handoffs, new decision rights and new metrics to the operating process.
Other aspects of the insurance workflow involve more knowledge work done at the individual level where the task can change based on the question being asked. Examples of this include financial planning, actuarial analysis and strategy work. Here, individual productivity tools developed by specialists for specialists will be more effective than an embedded AI workflow developed by a center of excellence. However the value of those individual tools will only compound when business users turn what they have learned into repeatable, reusable skills.
A flywheel for future innovation
Domain experts working on the ground in the business are essential for establishing the design and the key parameters for AI uses, but once the design is settled, the build should move to a central IT team, which owns the agent harness, model selection, integrations and controls. This is where standardization and centralization pays off, with the same harness, observability and guardrails serving every business unit.
Managing the balance between decentralized processes and centralized controls is the challenge, and that’s where federated models yield the greatest value. What is common across the enterprise is a platform and model access, sling with the skill catalogue and the security scan, the agent harness and the assurance framework. Everything that requires domain knowledge is distributed, giving each business function decision power over which skills to build, which processes to redesign and how to judge whether an agent is doing the right thing. The business units own the outcomes. IT provides the foundation.
How to establish a federated model
Getting the balance between centralized and decentralized functions right can be tricky. Autonomy and guardrails should not be distributed evenly across all parts of the business. They should be earned.
Here are some rules of the road:
- Federate asymmetrically: Not every spoke gets equal autonomy at the same time. Autonomy is granted against demonstrated maturity, risk, talent, adoption discipline and outcome ownership.
- Scale centrally: As models mature, more aspects of their operation should be able to be centralized. The federated approach is designed to demonstrate proof points, but scaling happens in the centralized environment.
- Share costs: A core AI platform should be centrally funded. Don’t hide platform cost inside a single use case budget.
3 key questions
Before choosing an AI operating model, IT should run assessments to determine how to proceed. Here are some questions to ask as the journey begins.
- How is the business model structured today, who holds the mandate for individual business functions and who measures what’s working and what’s not?
- Is your time-to-value for each new AI use case shortening, or staying flat?
- What has to be common across the enterprise before you distribute AI decision-making, and is it actually built and reused, or just declared in a slide?
Remember, if a capability only helps the team that built it, it’s not federated. What you then have is a fragmented business that thinks it’s making progress on AI when it is actually standing still
