Infosys Topaz connects AI agents with existing enterprise services

Infosys Topaz is an enterprise AI portfolio that combines services, solutions and platforms. Topaz Fabric organizes reusable agents, models and services around existing enterprise systems, while Agentic Foundry addresses the work of building and operating agents. For a buyer, the key question is how those components improve a specific service without obscuring who owns the data, approves actions and remains accountable when the automated path fails.

01 / ProductSeparate the portfolio from the implementation components

The Topaz overview describes an AI-first set of services, solutions and platforms using generative AI. It is a portfolio name, not a single general-purpose chatbot subscription. A useful discussion with Infosys should identify the exact component and delivery service being proposed, along with the business process it is intended to change.

Topaz Fabric is presented as a composable stack of agents, services and models. The current page distinguishes digital workers that run deterministic workflows, AI workers handling more complex tasks, and human workers responsible for judgment and governance. That separation matters because different tasks need different forms of control rather than one blanket claim of autonomy.

The current enterprise AI page places Fabric within Topaz and describes it as suitable for clients building on existing AI investments. It also presents Topaz AI Next as a broader business operations platform associated with EdgeVerve. Buyers should avoid treating those related names as identical products or assuming a Fabric proposal includes every part of the wider portfolio.

02 / AudienceStart where the existing service has recurring coordination work

Infosys is relevant when AI must work across an established estate rather than a new standalone application. An incident response team may already have monitoring, ticketing, a service catalog and runbooks. Its recurring work is reconciling those records, identifying the responsible service owner and deciding whether an approved remediation applies to the current situation.

Topaz Fabric for Operations describes observation, diagnosis, remediation, learning and human analysis. It names integrations with enterprise platforms such as ServiceNow and Dynatrace. An integration logo establishes a relationship to investigate; it does not prove that every version, permission model or custom workflow in the customer’s environment is supported without additional work.

Read the ServiceNow blueprint when the service record and approval process are central. The Dynatrace blueprint is relevant when the difficult question is understanding application behavior and dependencies. Infosys may connect work across those environments, so compare the additional service responsibility being purchased rather than assuming their existing roles disappear.

03 / WorkflowA proposed incident triage workflow with a narrow action boundary

Consider an application support team handling repeated failures in an overnight order-import job. This proposed workflow prepares an incident assessment and a suggested next step. Begin with one job and its documented dependencies. The aim is to reduce repeated evidence gathering while preserving the operator’s ability to distinguish a known failure from a new data-integrity problem.

Connect monitoring events, job execution records and recent deployment changes. Give the workflow stable identifiers for the application, job run and affected dataset. A generated summary should preserve the original event times and links, because two similar errors can have different causes when one occurs before a deployment and another after it.

Use deterministic rules to establish known facts such as whether the upstream file arrived and whether the same run has already been retried. An AI component can group relevant evidence and explain possible causes. Keep those roles distinct: a plausible explanation should not silently overwrite a failed status or turn an incomplete import into a successful one.

Agentic Foundry describes framework-based agent services, pro-code solutions, graph-based plans and human interaction. It also discusses evaluation and agent operations. Use the discovery phase to decide which approach fits this incident process. A fixed sequence for a known retry condition may be more appropriate than an agent dynamically inventing a recovery plan.

Require the assessment to name its evidence and the conditions of the approved runbook. If it proposes a retry, show whether that retry is safe for duplicate records and whether downstream work has already started. These are application-specific requirements to implement and test. A general promise of self-healing does not resolve them.

Keep execution under an operator’s approval during the pilot. The approval record should identify the run, proposed action and current evidence state. If the evidence changes while the decision is pending, refresh the assessment instead of executing an old proposal. This protects against a common operational race: a person fixes the problem while an automated workflow still holds an earlier plan.

Evaluate with old incidents containing missing files, schema changes, partial imports and duplicate retries. Compare the time to assemble evidence, the correctness of the proposed runbook match and the number of cases appropriately escalated. A system that quickly recommends the wrong retry is less useful than one that clearly explains why it cannot safely choose.

04 / PricingBuy a defined service and clarify the software boundaries

Infosys’ public pages provide demonstration and service-request routes, but the consulted material does not publish a standard Topaz Fabric subscription price. The commercial unit must therefore be established in a proposal. Do not infer a per-agent fee, unlimited model use or a particular cloud allowance from the phrase services-as-software.

For the order-import example, ask the proposal to distinguish initial integration, evaluation work and continuing service operation. Identify who provides monitoring access, maintains the runbooks and pays for model or cloud consumption. Those costs may be packaged differently across engagements, so a comparison needs an explicit scope before a headline price is meaningful.

The operating agreement should also specify how exceptions are handled. A reduced support workload is not established by the existence of an agent. Budget for incidents the system cannot resolve, connector changes, review of proposed actions and periodic checks against source-system changes. These responsibilities are part of the service being evaluated.

05 / DistinctionsThe worker model encourages a useful division of responsibility

Topaz Fabric’s distinction among deterministic, AI and human work is useful when it guides implementation. In the import workflow, software can verify file arrival, an AI component can organize evidence, and a person can authorize a recovery with business consequences. That division is more precise than labeling the entire process autonomous and then adding a generic approval screen.

The Responsible AI Suite is organized around Scan, Shield and Steer. Infosys presents it as a set of accelerators and solutions for AI governance, security and ethical use. It creates a basis for asking which controls will be included in the engagement. It does not independently establish that a customer application satisfies every applicable requirement.

The portfolio’s emphasis on existing investments is also relevant to the practical economics of adoption. If the current service platform already holds the approval and incident history, retaining that system of record can avoid a second disconnected queue. The implementation should make responsibilities visible across systems rather than create another place where staff must reconstruct what happened.

06 / QuestionsConfirm the actual deployment and the claims being measured

Which specific agents and connectors are available for the proposed environment today? The public pages cover a broad range of industries and service types. Request a demonstration using the intended source systems and permission boundaries. A prebuilt agent that requires substantial customization may still be useful, but that work belongs in the delivery estimate.

What information leaves the customer environment, and what is retained for evaluation or learning? The portfolio pages discuss responsible operation without establishing a single data-handling arrangement for every engagement. The proposal should document model providers, hosting locations, retained traces and the process for deleting or correcting information in operational records.

How will the organization interpret the operations page’s improvement figures? They are vendor statements, not a forecast for the selected application. Establish the existing incident baseline, distinguish evidence preparation from total resolution time and retain unresolved cases in the evaluation. Excluding difficult incidents would make an automation result look stronger while concealing the remaining workload.

07 / DecisionEvaluate one service before expanding the AI operating model

Infosys Topaz is worth considering when an organization needs AI integrated with established enterprise services and a delivery team that can work across those systems. Select a recurring process whose evidence, ownership and permitted actions can be defined. Then ask the proposed Fabric configuration to make that process easier to inspect and maintain.

For incident triage, a useful first outcome is an accurate assessment that preserves source links and identifies when an approved action is appropriate. Expansion should follow demonstrated operational usefulness and a clear maintenance agreement. The number of available agents is less important than knowing which responsibility each agent takes and which decisions remain with staff.