The short version
- Start with one scoped workflow and approved access.
- File preparation can include extracting dates, drafting a chronology and flagging possible gaps.
- A fee earner checks sources, omissions and output before use.
- Private hosting is one control; it does not eliminate confidentiality risk.
What an agent could do
An agent can use connected tools to carry out agreed steps after a person or an event starts the workflow. Its permissions and approval points determine what it can do. Some steps may use ordinary automation; others may need an AI model or human judgement.
A useful first pilot prepares work for review and flags exceptions. It should not be described as completing every job without interaction or supervision.
The job worth giving one first
File preparation is one candidate because the steps repeat and the output can be checked against the source documents. Its suitability depends on document quality, access and the effort needed to review the result.
An illustrative pilot for a new matter could be scoped to:
- Read the approved bundle, with scanned-document extraction and file formats assessed during scoping.
- Pulls out the key dates and the parties and puts them in one place.
- Drafts the chronology in the firm’s own format rather than a generic one.
- Lists what is missing or inconsistent for a fee earner to verify. A search failure does not prove that a document is missing.
The proposed output is a draft chronology with source references and a list of potential gaps. The fee earner checks the original documents, corrects errors and confirms omissions. This is an example to evaluate, not a measured customer result.
What it must not do
It does not advise, it does not decide, and it does not produce anything that leaves the firm unsupervised.
The SRA’s warning notice on misuse of AI emphasises confidentiality, effective supervision and accountability for outputs. A workflow with human review still needs appropriate safeguards; its design alone does not establish compliance.
The line that matters
Anything requiring judgement stays with the fee earner. The agent does the sorting, the extraction and the list-making that happens before judgement starts.
Confidentiality needs more than private hosting
Both free and paid AI tools need appropriate contractual, technical and organisational safeguards for client information. Review the actual architecture and terms rather than assuming a particular product or hosting label makes the use safe.
A private deployment can reduce exposure to external services, depending on the architecture. We agree where processing occurs, restrict access, review supplier and training terms, and test the handling of documents, prompts and logs. The firm remains responsible for how the system is used and supervised.
Discuss document boundaries and a private AI pilot.
How to tell whether it is worth it
Three questions, answerable in an afternoon without buying anything.
- Is there a job with a right answer? Define a checkable output and representative test cases. Do not assume that a repeatable task can be automated accurately; test it and keep judgement with the responsible person.
- Does it happen often enough to matter? A weekly two-hour task beats a dramatic one that happens twice a year.
- Where does the data actually live? Not where it is supposed to live, but where it is. This decides how hard the build is, and most firms have not asked it.
If the third answer is uncomfortable, that is worth knowing before you spend anything. It is often the real fix.
Measure preparation time, corrections, missed cases and review effort against your current process before deciding on wider rollout.