Legal workflow automation: choose your first project
Use a practical legal workflow automation scorecard to compare tasks, review effort, integration and operating costs before choosing a first project.
Choose a recurring task with a clear finish, accessible source material and a person able to review the result. A promising automation saves useful effort after preparation, checking and correction are included.
Begin with a shortlist of actual tasks from the firm. Examples might include assigning an enquiry, preparing a document register or locating an approved policy. Confirm the problem with the people doing the work before selecting a tool.
What legal workflow automation includes
Legal workflow automation connects repeatable steps such as assigning tasks, requesting documents and updating a verified status. Some steps use ordinary rules; an AI step may classify a document or propose a draft which a person checks.
The decision is which task warrants implementation first. Use this worksheet after the introductory AI guide, and before the separate workflow evaluation.
Record a candidate in one sentence: “When [verified event] occurs, prepare [checkable output] for [named role], with [failure route].” If the team cannot fill those fields, clarify the process before buying a tool. The automation service describes the wider implementation approach; this scorecard does not promise a particular saving or project scope.
Record a baseline the team recognises
Observe representative work and record preparation time, waiting, review, correction and failures. State what counts as completion. If the current process is poorly recorded, begin with a short measurement exercise rather than inventing a baseline.
Keep the data proportionate. The purpose is to understand the task, not to gather unnecessary client material in an analytics system.
Compare candidates using evidence
| Criterion | Question | Evidence to bring |
|---|---|---|
| Recurrence | Does this happen often enough to justify ownership? | Actual task counts and examples |
| Source quality | Can the output be checked? | Permitted sources and expected results |
| Review capacity | Who can assess the proposed output? | Named reviewer and realistic workload |
| Integration | Can the workflow fit the existing systems? | Demonstrated supported access and failure behaviour |
| Consequence of error | What could go wrong, and who is affected? | Specific failure scenarios |
| Reversibility | Can staff stop or complete the work manually? | Fallback and recovery procedure |
| Value | Does the completed process improve? | Preparation, review, correction and operating costs |
This is an Ampliflow decision aid, not a validated scoring instrument. Mark each criterion as evidenced, uncertain or unsuitable for the proposed trial. Do not add the marks into a single number that hides a serious unresolved issue.
Work through a capacity example
Suppose a fictional team completes 80 comparable document registers in a month. Its current process takes an average of 45 minutes per register, including checking. In an illustrative assisted process, preparation takes 8 minutes, review takes 12 and correction takes 5. These are invented inputs for the calculation, not a benchmark or an Ampliflow result.
| Calculation | Illustrative result |
|---|---|
| Current staff time: 80 × 45 minutes | 60 hours |
| Assisted staff time: 80 × (8 + 12 + 5) minutes | 33 hours 20 minutes |
| Difference before ongoing administration | 26 hours 40 minutes |
That difference is a capacity estimate. Subtract staff time spent maintaining the workflow, resolving failed runs and managing the supplier. Include failed attempts in the averages or record them separately; excluding them makes the comparison misleading. Setup and training also belong in the investment decision.
Now change one assumption: if checking takes 30 minutes, the assisted total becomes 43 minutes per register. The same 80 registers release only 2 hours 40 minutes before ongoing administration. The quality of the draft matters because review effort can consume most of the apparent benefit.
Use the firm's measured inputs in place of these figures. A cash-saving case also needs an actual cost that falls or demonstrable additional work; multiplying these hours by a charge-out rate does not establish either.
Apply the stopping conditions first
Defer a candidate if its accountable reviewer cannot be named, source material cannot be used appropriately, permissions cannot be enforced or a failure has no workable human route. A frequent task is not automatically a suitable AI task.
For SRA-regulated work, the supervision guidance calls for appropriate human scrutiny and an authorised individual retaining responsibility for AI-assisted legal services. That responsibility belongs in the choice of task as well as in the final review. SRA: effective supervision.
Compare the simplest viable options
For each candidate, consider a process change, an existing software feature, a rules-based integration and an AI-assisted step. A routing rule may solve a handover problem. AI becomes a candidate where interpretation or extraction is useful and can be checked.
The practice-management guide helps assess the existing system. The legal AI product comparison helps shortlist tools after the task is defined.
Write a bounded evaluation brief
- 01Observe the current task
- 02Check stopping conditions
- 03Compare simple options
- 04Define a permitted sample
- 05Measure completed work
- 06Proceed, revise or defer
Specify the input, proposed output, permitted actions, reviewer, failure route and success evidence. Use the synthetic evaluation method before considering live matter information.
Record the decision as proceed, revise or defer, with reasons and unresolved questions. A decision to evaluate is not approval for unrestricted production use.
The legal hub provides candidate workflows across practice areas. Bring the scorecard, one permitted example and the current process to a discussion; those give the next decision a concrete basis.