Harvey AI workflows: make repeatable work reviewable
Define the inputs, output and review points for a repeatable Harvey task. Confirm the workflow or agent available to your organization before building around it.
Updated 5 min read
Harvey AI workflows use guided inputs for structured, multi-step work, according to its support hub (opens in a new tab). To make one useful, define the input, the expected result and the points where a person must review the work. fenbs can preserve an approved generic procedure between sessions, while client material and confidential outputs stay in firm-approved systems.
Start with one repeatable task that a reviewer can explain and check. A workflow should make that task easier to perform consistently. It should also make a missing input, an uncertain finding or an unfinished review visible before the output is treated as complete.
Are Workflows and Agents the same instructions?
Harvey’s public help hub retains an Agents & Workflows topic and a Workflows Overview article. Its current Agents page (opens in a new tab) describes agents, Agent Builder and Review and Approve Work. Those product labels are related, but they do not establish that an older workflow uses the current builder or has identical controls.
Ask your administrator which feature is available to your organization and whether you may use or customize it. The public workflow article requires authenticated access, so it does not supply a verified creation-menu sequence for this guide. Use your organization’s current instructions for the feature you actually have.
If you are still selecting a platform, see Claude vs Harvey. For the broader organization of reusable procedures, read the AI workflow library guide. The task here is to define a repeatable Harvey process whose result another person can assess.
What should you define before using or building one?
Write a short task description in your approved working system. State what starts the work, what information is required and what a satisfactory result contains. Make exclusions explicit. A process that knows when to stop is easier to review than one expected to fill every gap on its own.
| Part | What to write |
|---|---|
| Input | The approved material and required fields, with a way to identify missing information. |
| Task | One outcome the assistant should work toward, with clear boundaries. |
| Output | The requested structure and supporting evidence for the reviewer. |
| Review | The checks a named person performs before the work moves forward. |
| Stop point | The uncertainty or missing approval that must pause the task. |
For example, a practice task could compare a short set of synthetic clauses against a generic checklist. The output might identify the relevant passage, the checklist item and what still needs a person’s interpretation. This example describes a task specification, not a promise that Harvey includes that exact template.
Where should a person review the work?
Put review where a mistake could affect the next stage. Before execution, check that the intended sources and scope are correct. After extraction or analysis, check the evidence. Before any result is reused or delivered, confirm that the unresolved points have reached the person responsible for them.
The Agents page describes inspecting a plan, changing its scope and approving work before execution. It also describes steps and citations that support review. Confirm how those controls work for the particular feature in your deployment. Do not assume a public description of Agents proves the controls of every older workflow.
An approval should answer a specific question. “The input set is correct” differs from “the final analysis is ready.” Name the review being completed so that one approval is not silently treated as both. A person remains responsible for judging whether the evidence supports the intended use.
How do you test a procedure before repeating it?
Use a small public or synthetic sample your firm permits and establish the expected checks manually. Include an ordinary example, an incomplete example and an example with an exception. The purpose is to find where the procedure needs clearer instructions before anyone relies on it for a wider set.
- Confirm the approved inputs and the expected output for the sample.
- Use the workflow or agent your organization provides, following its authenticated instructions.
- Inspect each stage that can change the conclusion, including source passages and unresolved questions.
- Record where the output differs from the task specification and decide what needs correction.
- Repeat the affected checks after a change and obtain the required human review before extending use.
Changing a prompt can change more than the sentence you meant to improve. Revisit the relevant sample after each material change. Mark a check as unfinished when it has not been repeated. The AI agents for research guide discusses giving research work a clear finish line.
What if the result is incomplete or unsupported?
Keep it open. Identify the missing input, unsupported inference or unresolved interpretation and send it to the appropriate reviewer. A well-structured table or polished memo can still contain a gap. Do not treat completion of the steps as evidence that the conclusion is sound.
Use the Harvey legal research guide when the gap concerns a proposition and its source. Use the Harvey Vault guide when it concerns the collection of documents being analyzed. A repeatable procedure should send these problems to the right check rather than hide them in a final summary.
What should you preserve for another run?
Keep the approved general procedure, the reason for its review points and the safe lessons from testing. Harvey’s Library & History topic (opens in a new tab) describes access to saved workflows, prompts and previous queries. It does not establish an export format or guarantee that material can move to another AI tool.
Client confidences and deal data stay out of any tool your firm has not approved. Keep client inputs, source text, matter-specific findings and legal advice in approved firm systems. A generic lesson such as “pause when a required input is missing” can be useful without copying the confidential example that prompted it.
The AI at work guides cover the habits around individual tools. For an approved generic procedure, fenbs can keep the next improvement and the point where testing stopped, so a later assistant does not have to guess which checks remain.
Keep the prompts that work
Save your best prompts in fenbs and use them in any AI, on any device. No more digging through old chats.
Create a free accountFree for up to 5 people and 50 open tasks · No credit card · Works with Claude, ChatGPT, Cursor and more
How to keep a reusable procedure with fenbs
Keep the safe procedure so you can carry on from exactly where you left off, on any device and in another connected AI. fenbs is not the right place for client work product or confidential workflow examples. Check firm approval before saving even a general method elsewhere.
- Create a free account.
- Write the approved steps in plain words, including what to check and when to ask a person. Use public or synthetic examples only.
- Connect Harvey to fenbs through Integrations only if your firm approves and its administrator confirms and configures a supported connection. A ready-made Harvey connector has not been verified.
- Save which general checks passed, which still need work and the next step. Keep client facts and confidential outputs out.
Harvey’s Connector Library (opens in a new tab) supports custom MCP connectors under workspace administrator control; its public page does not provide a full setup sequence. The fenbs overview explains how another approved assistant can read the saved work when you return.