A useful SAP Joule agent pilot starts with one bounded business problem, a measurable baseline, controlled access, and a clear escalation route. Begin with recommendations that people review. Expand execution authority only after the agent demonstrates dependable behavior on representative cases. Ninety days is a planning framework, not a guarantee that every organization can reach production within that period.

For U.S. SAP leaders, the opportunity is to remove avoidable work without creating a faster route to operational mistakes. The first pilot should answer a business question: does this agent improve a specific workflow enough to justify its cost and support requirements?

Separate platform momentum from deployment readiness

SAP's May 2026 announcement outlined its Business AI Platform direction. That direction is relevant to strategy, but it does not establish which agent capabilities are available in your edition, region, or subscription today. [1]

SAP and Anthropic also announced plans to expand Claude's role across SAP Business AI. Treat the announcement as a roadmap signal. Confirm current availability and terms before making a feature central to a project plan. [2]

Build a capability checklist with your SAP account team and implementation specialists. Include required applications, environments, permissions, service entitlements, integration methods, and restrictions. A compelling demo does not replace that checklist.

Pick a workflow small enough to evaluate honestly

Good first candidates have frequent cases, clear evidence, identifiable owners, and bounded consequences. Examples include summarizing an invoice exception, preparing an order status explanation, or drafting the next step for an internal service request.

Avoid beginning with a workflow that combines supplier changes, purchasing authority, and payment decisions. Multiple permissions and irreversible consequences make evaluation harder.

Choose a case where the output can be checked against a known record. Ask what a competent employee needs to inspect before deciding. Those records become the agent's grounding requirements and the reviewer's evidence packet.

Also define what the agent cannot do. An agent that summarizes an invoice exception does not need permission to change supplier banking details. Restricting tools makes both failure analysis and support more manageable.

Days one through fifteen establish the baseline

Document the current workflow before introducing automation. Measure handling time, queue age, escalation frequency, rework, and recurring error categories. Use a representative sample rather than the easiest available cases.

Record the human decision criteria. A written procedure may omit steps experienced employees perform instinctively, such as checking whether a delivery receipt is current or recognizing an unusual currency combination.

Create a pilot charter with the process owner. State the business objective, allowed actions, prohibited actions, evaluation method, stop conditions, and incident owner.

NIST's AI Risk Management Framework organizes risk work around governance, context, measurement, and management. Use that structure to connect the pilot's evaluation to business consequences rather than merely rating fluent answers. [3]

Days sixteen through thirty build the evidence contract

Define the records the agent must retrieve and how it identifies them. Specify company code, document identifier, timestamps, and any required business context.

Require outputs to distinguish observed facts, missing information, and proposed actions. The agent should not silently fill gaps with plausible explanations.

SAP's architecture guidance addresses AI integration, security, ethics, and governance. Apply those considerations to identities, data access, and the systems the agent can reach. [4]

Give the reviewer an efficient view of the evidence. If checking the agent takes longer than doing the original work, the workflow needs redesign even when the answer is usually correct.

Before live use, decide where prompts, tool results, approvals, and exceptions are recorded. Protect sensitive information and keep retention aligned with organizational requirements.

Days 31 through 50 test the uncomfortable cases

SAP's Joule Studio documentation includes guidance for testing an agent. Use the supported test tools in your environment, then extend evaluation with business scenarios and independent review. [5]

Include incomplete records, conflicting statuses, stale data, revoked access, tool failures, and unexpected user instructions. Test whether the agent stops when evidence is insufficient.

Do not evaluate only the final text. Review tool selection, arguments, source records, and proposed actions. A correct-looking answer produced from the wrong company's invoice is a failure.

Create a holdout set that developers do not repeatedly tune against. Keep it representative and refresh it when the business changes. Otherwise, an improving test score may simply reflect familiarity with the examples.

Test recovery as well as failure detection. The agent should explain what happened, preserve the relevant evidence, and route the case to someone who can act.

Days 51 through 70 run in shadow mode

Consider a hypothetical U.S. accounts payable team that handles blocked invoices. Its pilot agent retrieves the invoice, purchase order, and receipt status, then drafts an explanation and suggested next step.

In shadow mode, the employee continues the normal process while the agent's recommendation is evaluated separately. The agent cannot release a block or post a financial transaction.

A case may look simple: the invoice matches the purchase order, but the goods receipt is missing. The correct next step might be contacting receiving, not declaring the shipment complete because a supplier attached a delivery note.

Compare the recommendation with the employee's decision and the underlying evidence. Record disagreements by cause: missing source data, retrieval error, reasoning error, policy ambiguity, or reviewer inconsistency.

This is an illustrative pilot design, not a reported GBSI customer result. Its value comes from exposing failure categories before granting execution authority.

Days 71 through 90 decide what to scale

Move to an approval-based workflow only if evidence supports it. Define which recommendations can enter that workflow and which remain manual.

Evaluate business improvement alongside control performance. Faster handling is useful when rework, inappropriate actions, and review effort remain acceptable. Track the cost of model usage, integration, monitoring, and support.

Make the decision reversible. The owner should be able to disable the agent's tools or route cases back to the existing process without disrupting normal operations.

The final decision may be scale, extend testing, narrow scope, or stop. A pilot that reveals an unsuitable use case has produced useful information. Calling every pilot a success weakens investment decisions.

Write the incident plan before the first live recommendation

A pilot needs an incident plan even when it begins with read-only access. Incorrect advice can still cause an employee to make a poor decision, expose information, or waste time investigating the wrong record.

Define who receives an issue, how it is classified, and who can suspend the pilot. The business owner should participate when the error concerns process interpretation. The technical team should investigate retrieval, permissions, tool behavior, and execution logs.

Preserve the relevant evidence without spreading sensitive data through informal messages. Record the request, retrieved identifiers, recommendation, reviewer action, and eventual business outcome. Access to those records should follow organizational policy.

Separate immediate containment from permanent correction. Disabling a tool may prevent further mistakes while the team investigates. A prompt adjustment alone may not fix an authorization problem or an ambiguous business rule.

Define the restart criteria. The owner should know which tests must pass and what approval is required before the agent returns to the workflow. This avoids a rushed restart based only on a convincing demonstration.

Review incidents as learning evidence, not as a reason to hide failures. A small pilot is valuable precisely because it can expose weaknesses before wider deployment.

Also plan for absence. The designated reviewer may be unavailable, and the queue may grow outside business hours. The process needs a backup owner and an acceptable manual route.

A good pilot report includes the incidents, near misses, unresolved limitations, and the costs of operating the controls. Leaders can then judge whether the improvement is worth maintaining.

This discipline makes the scale decision more credible. It also gives employees a practical reason to trust the workflow: they can see its boundaries, report problems, and return to a known process when necessary.

Make those instructions easy to find within the workflow itself. Rehearse a suspension and restart with the owners before live recommendations begin, and keep the tested manual route available for employees to use.

Frequently asked questions

Is a Joule agent pilot the same as a chatbot pilot

No. An agent may select tools, retrieve business records, and propose or execute workflow steps. Evaluation must include those behaviors and their permissions, not just conversation quality.

What should the first pilot measure

Measure handling effort, evidence quality, incorrect recommendations, escalation behavior, rework, and operating cost. Use the current process as a baseline and inspect difficult cases separately.

When should the agent ask a human

Escalate when required evidence is missing, records conflict, access fails, policy is unclear, or the proposed action exceeds its authorized scope. The trigger should be explicit and testable.

Make the first agent useful enough to keep

Genius Business Solutions (GBSI) provides SAP Business AI and process consulting services. Bring a candidate workflow, sample exceptions, and a baseline to a GBSI discussion. The aim is a controlled pilot that gives leaders evidence for the next investment decision.