1. The workflow and its limits are clear
Someone can describe the trigger, the expected result and the conditions that require a person. Choose a complete task rather than a broad goal such as “improve sales”. For example: prepare a draft summary of an inbound request, then put it in an analyst’s review queue.
Evidence to bring: a process map, representative examples and a list of decisions outside the pilot. If the team disagrees about the process, resolve that first. AI will not settle competing ownership rules.
2. The required data is usable
Identify the specific fields and documents the task needs. Check whether they are current, traceable and consistently linked to the right customer or request. A sample of clean records is not enough if difficult cases routinely contain missing or conflicting information.
Evidence to bring: an approved test set covering normal work, incomplete records, duplicates and exceptions. Record where each important fact comes from. If the task cannot proceed without guessing, define a request-for-information or escalation path.
3. Access, actions and review are agreed
Define the user or service identity, permitted records, allowed tools and approval requirements. Test those boundaries with the identities that will actually use the workflow, including attempts to access another customer’s information.
Evidence to bring: an access matrix, named approvers and examples of actions the system must refuse. For Salesforce Agentforce, review the platform’s shared-responsibility guidance alongside your own configuration.
Salesforce: permissions and agent guardrails
4. There is a testable definition of success
Agree what a correct result looks like and what would make a release unacceptable. Include the quality of the output, the action taken and the final system state. A polished response is not a successful result if it updates the wrong record.
Evidence to bring: reviewed examples with expected outcomes, failure cases and criteria for stopping a pilot. Anthropic’s evaluation guidance explains why task results, execution traces and evaluation methods need attention when assessing agents.
Anthropic: evaluating AI agents
| Situation | Expected behaviour |
|---|---|
| A required document is absent | Identify the missing item and leave the request for review. |
| Two sources disagree | Show the conflict without selecting an unsupported value. |
| A document includes instructions to ignore rules | Treat document text as input data, not permission to change the workflow. |
| A CRM update times out | Make the failure visible and avoid duplicate updates on retry. |
| The user lacks access | Do not retrieve or expose the restricted information. |
| The summary is accepted | Record the reviewed version and the permitted next action. |
5. Someone can operate and improve the workflow
Name the business owner, technical owner and person responsible for handling exceptions. Decide who can change prompts, models, rules or integrations, and what must be retested when those changes occur.
Evidence to bring: a monitoring plan, escalation route and rollback procedure. The team should be able to pause the automated step and complete the work through an agreed fallback.
What a gap means
A gap is a work item, not an automatic rejection of AI. Missing ownership calls for an operating decision. Poor records call for data work. An unclear process calls for discovery. Some of those improvements may solve the original problem without an agent.
A high ticket volume, slow response time or a previous project overrun does not establish AI readiness. Those are symptoms to investigate. Avoid scoring them as proof that an agent will pay for itself.
Use the scorecard to organize the discussion
Our ten-question scorecard helps record the foundations you have and the areas that need attention. It shows the result before asking for contact details. The score is self-reported and is not a certification, an ROI forecast or a review of your Salesforce org.
Complete the AI readiness scorecard
For a discovery conversation, bring one workflow and the evidence you already have. Do not share customer records or credentials in the website form. We can agree the next investigation and the appropriate way to exchange material.
Scope the next step
Bring one workflow, the systems it touches and an example of where the current process breaks down. We can discuss whether to improve the process, use conventional automation or scope an AI pilot. A delivery estimate follows the agreed requirements.