An adoption decision starts with a business problem
“We need an AI agent” is not yet a project brief. A useful brief identifies work that is slow, inconsistent or difficult to complete, the people affected and the result worth improving. Then compare process changes, ordinary automation and AI assistance.
An agent can use a model to select steps and tools during a task. A predefined workflow can also use AI at specific points. Anthropic makes this distinction in its architecture guidance. Neither label establishes that a system is suitable for your business.
Anthropic: choosing between workflows and agents
Choose one task with an observable result
Useful candidates are tasks your team can demonstrate and evaluate. The following are illustrative starting points, not claims of completed client deployments or measured savings.
| Team | Bounded first task | Review point |
|---|---|---|
| Lending operations | Draft a submission summary and list missing information. | An analyst checks the sources before progressing the request. |
| Logistics | Extract a delivery enquiry into a draft job request. | An operator confirms locations, requirements and availability. |
| Healthcare and wellness operations | Classify an administrative enquiry and prepare a response draft. | Authorized staff check the message before sending; clinical decisions remain outside scope. |
| Nonprofit operations | Prepare a relationship summary from approved donor records. | A staff member verifies attribution and the proposed follow-up. |
Choose work where the input is accessible, the output can be checked and a mistake has a controlled recovery path. Volume matters, but review effort and error consequences matter too.
Describe the boundaries in the project brief
- Trigger: what starts the task and how duplicate requests are handled.
- Inputs: required records, documents and permissions.
- Output: the exact draft, recommendation or authorized change.
- Review: who checks it, what they see and what requires escalation.
- Exceptions: what happens when information is absent or a tool fails.
- Ownership: who operates the workflow and approves changes.
For work crossing Salesforce, another CRM, a portal and document storage, map the responsibilities of each system. The AI component should not quietly become a second source of truth for customer status or commercial terms.
Compare the pilot against the current process
Record a baseline using the same task definition you will use during the pilot. Include time waiting for information, time spent reviewing output and time correcting errors. A quicker draft does not necessarily mean a quicker completed task.
Keep the initial estimate separate from observed results. A useful cost model includes implementation, model usage, other software, monitoring, human review and maintenance. Released capacity is not automatically cash savings or additional revenue.
After a pilot, report the observation period, volume, types of cases and exclusions alongside results. If the sample is small or unusually clean, make that limitation visible before using it in a sales story.
Agree release criteria before building
Use representative normal and difficult cases, including incomplete information, denied access, conflicting documents and failed integrations. Evaluate whether the workflow achieved the correct result, not only whether the generated text looked convincing.
Anthropic’s agent-evaluation guidance discusses assessing tasks, traces and outcomes through complementary methods. Use that engineering foundation to design checks specific to the business process.
Anthropic: evaluating agent behaviour and outcomes
Agree the human review required at launch, the conditions for pausing the workflow and how work continues if the automated step is unavailable. Expanding authority should require a deliberate review of evidence.
Add complexity only when the evidence calls for it
Several agents may be useful when tasks genuinely require separate work or expertise. They also create more handoffs, costs and failure points. A fixed sequence or one model-assisted step may be easier to support. Make the architecture justify itself against the acceptance criteria.
The delivery estimate depends on data access, integrations, testing and the team’s ability to review the work. Start with a bounded scope and visible demonstrations. Set dates from those dependencies rather than applying the same timeline to every project.
Turn delivery feedback into the next improvement
At each review, show completed examples, corrections and unresolved issues. Keep a record of what changed and why. After release, review exception patterns and user feedback before expanding to another task.
When a project has a supported outcome, ask the customer for permission to share the scope, the result and their feedback. An approved case study helps future buyers understand what was actually delivered. It should distinguish measured results from goals and illustrative demonstrations.
See how a Growbiz engagement works
Explore the readiness scorecard
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.