An assistant that writes "referral sent" has not necessarily sent a referral. An assistant that creates a referral letter has completed only one part of the process. The useful distinction between a conventional AI scribe and an agent is not how conversational the interface feels, but what observable work happens beyond generating text.
Scribeberry's Berry product page, checked on 16 October 2026, advertises form completion, referral preparation and faxing, patient intake requests and follow-up tasks. It describes clinician review before outward actions, including a review card showing the recipient and content. These are advertised capabilities, not findings from a hands-on iatroX test.
Berry is consequently a useful product through which to examine a practical question: when software starts acting, what must be visible before and after the clinician approves it?
Use a referral as the unit of comparison
Take a fictional request to refer a patient to a specialist. The first task is to produce a clinically appropriate letter. The next tasks are to identify the destination, include the relevant material, obtain approval, transmit it and preserve evidence of what occurred.
A conventional scribe may help with the first task. An agentic workflow claims to carry the instruction further. That extra scope is valuable only when the result is verifiable. A well-written document saved in the wrong place is not a completed referral.
For a local demonstration, define the end state before starting. The clinician should be able to see the approved document, the exact recipient, the included attachments and the transmission outcome. The test should also establish whether successful transmission means the receiving service accepted the request. Those are not automatically the same event.
This approach avoids ranking products by the number of verbs on their homepages. The comparison becomes the proportion of agreed workflows completed correctly, with a known amount of review effort.
Context should be bounded by the task
Berry says it works from visit notes, patient details and the user's request. Its current description distinguishes chart actions associated with Accuro from the wider scribe's compatibility with other systems. Buyers should therefore confirm the exact capabilities available in their own setup rather than assume every advertised function works through every EMR.
The underlying principle is broader than one integration. An agent should have enough information to complete the authorised task, without treating all accessible information as relevant to every output.
For example, a fictional administrative form might need a specific functional history but not unrelated sensitive details from the encounter. A referral might need a recent result but not an obsolete provisional diagnosis. The reviewer should be able to identify both omissions and unnecessary inclusion.
A useful test would ask the agent to prepare two different documents from the same encounter. Review whether each contains the information appropriate to its purpose, rather than the same comprehensive summary with different headings.
Approval must describe the consequence
The Berry page says the user approves actions before sending, scheduling or sharing. This is an important product boundary. The quality of the boundary depends on what the clinician can inspect at the moment of approval.
A meaningful review should display the patient, destination, content and intended action together. It should not require the reviewer to remember details from a previous screen. A change to the recipient or an attachment should remain visible before the final commitment.
Approval should also be granular enough to distinguish actions. Accepting a letter's wording is not identical to authorising transmission. Agreeing that a follow-up task is sensible is not identical to assigning it to a particular colleague.
These are proposed usability criteria rather than assertions about Berry's implementation. Their purpose is to make "human in the loop" operationally meaningful. A clinician cannot exercise useful oversight over an action whose content or consequences are hidden.
Failure is part of the workflow, not an edge detail
A credible agent demonstration should contain a failure. In a fictional referral test, the receiving number could be unavailable or the attachment could exceed a permitted size. The question is what the clinician sees next.
A trustworthy workflow should distinguish a draft, a queued action, an attempted transmission and a confirmed result. It should not collapse those states into one optimistic completion message. Uncertainty about execution should be communicated as uncertainty.
The next question is recovery. Can the user correct the destination and retry without accidentally creating duplicate transmissions? Is the previous attempt retained? Does the system identify which version of the document was actually sent?
This is where agentic capability becomes more demanding than text generation. An incorrect paragraph can be edited before use. An incorrect external action may already have affected another person or system. The appropriate response is not to prohibit all automation, but to require visible status, restrained permissions and a workable recovery process.
Forms, intake and tasks need different acceptance tests
One generic accuracy score would not adequately describe all the workflows Berry advertises. A form should be checked against required fields and the supporting encounter information. An intake request should be assessed for the correct recipient and the appropriateness of the information requested. A follow-up task needs an owner and a meaningful due date.
Consider a fictional instruction to check a result next week. The generated task may be linguistically correct but operationally incomplete if it is saved to an unmonitored personal list. The clinic should decide where such tasks belong and which team members can see them.
Likewise, a submitted intake does not automatically mean its content has been reviewed before the appointment. The workflow needs to make clear whether information has merely arrived or has been incorporated into the clinical process.
Different tasks therefore require different definitions of completion. The product should be tested against those definitions rather than rewarded for producing an output of any kind.
Measure the work that remains after the agent finishes
A sensible trial could compare ordinary documentation support with the proposed agentic workflow using fictional or appropriately authorised test encounters. Measure time to an approved document, time to verified completion and the number of corrections needed along the way.
Separate routine successes from cases involving uncertain destinations, missing information and changed instructions. The harder cases reveal whether the agent reduces coordination work or merely relocates it to an exception queue.
The clinic should also assess whether staff can inspect actions taken by colleagues and whether responsibility is clear when the original clinician is absent. An individual productivity feature needs a team-level operating model before it becomes a dependable service.
This is the concrete version of the question raised in iatroX's AI practice-manager analysis. "Agentic" should describe an observable, authorised and traceable workflow. Berry's advertised scope makes that question testable, but the answer should come from completed work and manageable oversight, not the label alone.
Frequently asked questions
What is Berry within Scribeberry?
Berry is presented as an AI agent for work around the patient encounter, including forms, referrals, intake and follow-up. The product page describes it as part of the Scribeberry experience rather than a separate clinical reference service.
Does Berry act without clinician approval?
Scribeberry describes approval before sending and other consequential actions. A clinic should still verify the specific review screens, permissions and confirmation steps in its own deployment, particularly for external communications.
What is the fairest comparison with a conventional AI scribe?
Compare the entire agreed workflow, including review, transmission, confirmation and recovery from failure. A higher number of generated documents does not necessarily mean more correctly completed tasks.
Explore practical healthcare AI analysis in the iatroX Journal →