Agentforce readiness assessment

Decide what should become an agent

A convincing demonstration does not establish that a business process is ready for an agent. The example may use ideal data, unrestricted access and a person quietly correcting the output. Emerge’s readiness assessment identifies the conditions needed for a useful production workflow and the gaps that should be resolved before release.

The assessment starts with a specific task, audience and business owner. Answering an approved service question, preparing a sales brief and changing an account record have different requirements. We compare an agent with simpler alternatives such as search, a structured form or deterministic automation, then recommend the approach that fits the actual problem.

Describe the task and its current cost

We map the current process using representative examples. The inventory records the trigger, inputs, decisions, systems, handovers and evidence of completion. It identifies how often exceptions occur and which steps require judgement or authority that cannot be delegated casually.

Value is expressed in the terms the operation can measure: useful work completed, time spent reviewing, delays waiting for information or avoidable repeated contact. We distinguish a measured baseline from an estimate. A proposed saving is not presented as an achieved result, and automation volume is not treated as a benefit when it creates more correction work downstream.

The first candidate should have a clear boundary and a receiving team willing to own the outcome. A narrow task with reliable sources can be a better starting point than an ambitious workflow crossing several unowned systems. The assessment makes that tradeoff visible to the sponsor.

Inspect knowledge and data readiness

We review the actual sources an agent would need, including ownership, currency, permitted audience and retrieval structure. A large document collection can still lack the approved answer to a common customer question. Conflicting policies and unexplained exceptions become content work items before they become model behaviour problems.

Structured data receives a similar review. We inspect identifiers, required fields, quality and freshness, then define what information may enter the conversation. The ability to connect a database does not establish that every record is relevant or authorised for every visitor.

The output is a source map tied to the proposed task, with representative supported and unsupported questions. It identifies where a person must supply judgement, where an answer needs a citation and where the safest useful response is a handover.

Assess actions and integration boundaries

Each proposed tool operation is classified as a read, a proposal or a system change. We identify the executing identity, required permission, input validation and destination result. Consequential actions need an explicit approval or policy boundary independent of the model’s wording.

Integration readiness includes duplicate handling, timeouts, retries and the ability to verify an uncertain outcome. A destination that can accept a request but cannot report its later state may limit what the agent can promise. Those constraints are reflected in the experience and release scope.

For Salesforce, the review covers the actual agent and messaging configuration, available actions and enabled products. It does not assume that a demonstration environment’s feature set matches the customer’s production contract. Unresolved access or entitlement questions become named dependencies in the plan.

Create an evaluation set before a pilot

Representative tasks are converted into evaluation cases with expected behaviour. The set includes ordinary questions, ambiguous requests, missing sources, restricted information and failed tools. It also contains attempts to override instructions or misuse retrieved content, using safe examples that do not expose real customer data.

Acceptance criteria distinguish factual support, permission correctness, action accuracy and appropriate escalation. The business owner defines which errors are unacceptable and which can be reviewed during a controlled pilot. An overall score should not hide a serious access failure inside a large number of easy answers.

We identify the traces and operating measures required to investigate a failure. Logging must support review without indiscriminately retaining sensitive conversations. Data handling, reviewer access and retention are agreed with the responsible owners.

Estimate operation in useful units

Consumption is modelled around the complete task, including retrieval, model calls, tool activity, retries and human review. Provider pricing and the customer’s actual Salesforce arrangements are checked when commercial estimates are prepared. We avoid presenting a generic per-message price as the cost of a completed workflow.

The assessment also identifies who will maintain knowledge, approve changes, respond to incidents and decide whether the pilot should expand. These responsibilities can determine readiness more strongly than the availability of a model or connector.

Leave with a concrete release decision

Emerge delivers a prioritised use case, source and action inventories, readiness gaps, evaluation examples and a staged release plan. Each gap has an owner and an acceptance condition. The recommendation can be to proceed with a bounded pilot, repair a specific dependency first or use a simpler solution for the selected task.

That gives the organisation a reviewable investment decision and a practical path into Agentforce, grounded in the work the agent will actually be expected to complete.