First-draft generation for routine filings

The Legal and Administration team, made up of four fee earners and one paralegal, produces routine legal filings every working day. The trigger is straightforward: a request arrives with a set of structured inputs, typically the names of the parties, a monetary amount, and the type of filing required, and the team must produce a reviewed, approved first draft ready to go to the client or counterparty. The three document types in scope are demand letters, payment reminders, and GDPR responses. At 1,800 runs per year, this is genuinely high-volume work for a small team.

The current process is entirely manual. A team member receives the request, identifies the correct template and the statutory references that apply, drafts the document inside the document management system, and then circulates it by email for internal review. Any errors caught during review trigger a correction cycle before the document is finalised. The whole sequence takes roughly thirty minutes on average, which adds up to approximately 900 hours of effort across a year. About 85 to 90 percent of cases follow the standard template pattern closely. The remaining 10 to 15 percent involve unusual facts or edge cases that require genuine legal judgment and cannot be handled by a template alone.

The team operates under significant compliance obligations. Legal professional privilege, GDPR, and the firm's duty of care all apply. A draft that leaves the firm with an incorrect statutory reference or a wrong legal conclusion could expose the firm to negligence liability. Client data cannot be processed by unvetted external tools. The team has no developer resource and no appetite for custom code, and the existing technology environment is Microsoft-based, centred on Outlook and a document management system.

The recommended path is an AI Workflow, specifically a no-code AI drafting and routing workflow built within the firm's existing Microsoft 365 environment, most likely using Microsoft Copilot for Microsoft 365 or Power Automate with AI Builder. The case for this approach rests on three things working together: the process is highly repetitive and structurally consistent for the vast majority of cases, the volume is large enough to make automation genuinely worthwhile, and the compliance obligations demand a controlled, auditable solution rather than an ad hoc one.

The fit with the team's stated technology preference deserves direct acknowledgement. The client has no developer resource, no appetite for custom code, and a firm requirement that client data stays within vetted infrastructure. A Microsoft 365-native tool satisfies all three constraints at once. It requires no developer to deploy, it integrates naturally with Outlook and the document management system the team already uses, and it keeps client data within the Microsoft cloud tenancy that the firm has presumably already assessed for GDPR and privilege purposes. That last point is not a minor convenience: it removes the data-residency risk that would accompany any third-party AI drafting service and is likely a condition of getting any tool approved internally.

The mandatory human approval gate is not a concession to caution; it is structurally correct for this process. The 10 to 15 percent of cases that fall outside the standard pattern cannot be reliably identified by the AI before a draft is produced, which means a fee earner must review every output, not just the flagged ones. More importantly, the negligence exposure attached to legal documents means the firm cannot delegate the final sign-off to an automated system regardless of how good the drafts are. Building the approval step as a hard gate in the workflow, one that prevents a document from being finalised without an explicit fee-earner action, means compliance is enforced by the system rather than relying on individuals to remember to check.

The alternatives were considered carefully. Hybrid scores close to AI Workflow in the assessment, and the honest answer is that a well-designed AI Workflow effectively is a hybrid: it generates drafts automatically and routes exceptions to humans. The distinction in the scoring reflects the fact that a more elaborate hybrid architecture, with multiple escalation tiers and complex branching logic, is not warranted for a team of five at this stage of tooling maturity. Start with the simpler implementation and extend it once the team has confidence in the system. RPA was considered and ruled out because it can populate templates with structured data but cannot generate natural-language legal drafts or handle the variability in wording that even the standard cases involve. It would also require a specialist to configure, which the client has explicitly ruled out. An AI Agent, meaning an autonomous system capable of multi-step reasoning without a mandatory human gate, would conflict directly with the firm's duty of care and is not appropriate here. The process does not need autonomous reasoning; it needs reliable draft generation with controlled human oversight. Education alone, training the team to use AI tools themselves on an ad hoc basis, would deliver some benefit quickly but would not produce the consistency, audit trail, or approval routing that compliance requires.

The ROI case is clear. At 1,800 runs per year and thirty minutes each, the process currently consumes around 900 hours annually. A conservative 60 percent saving, accounting for the ongoing approval step and the exception cases that will still need full manual effort, implies roughly 40,500 pounds of annual savings against a build cost estimated between 4,000 and 12,000 pounds. The payback period is approximately three months. Those numbers are based on a blended rate of 45 pounds per hour and should be validated against the firm's actual cost structure, but the order of magnitude is robust even with significant variance in the assumptions.

The two strongest candidates for this process are AI Workflow and Hybrid, and in practice they are close relatives. Both centre on AI-generated first drafts with a mandatory human approval gate before anything leaves the firm. The reason AI Workflow edges ahead is simplicity: the team has no developer resource, no confirmed APIs, and no appetite for complexity they cannot maintain themselves. A Hybrid path, as scored here, implies a more elaborate escalation architecture for the 10-15% exception cases, which adds real configuration and ongoing maintenance overhead that is not yet warranted for a five-person team. The AI Workflow approach handles exceptions adequately by flagging them clearly to a fee earner rather than attempting to route them through a separate automated track, and that is sufficient at this stage.

RPA scores noticeably lower and is worth addressing directly, because it might look attractive given the structured inputs. The problem is that RPA can populate a template with field values, but it cannot generate natural-language legal text, handle the variation between a demand letter and a GDPR response, or adapt when a case sits slightly outside the standard pattern. It would also require a specialist to configure and maintain, which the client has ruled out. The gap between what RPA can produce and what a qualified first draft actually needs is wide enough to make it the wrong tool here.

Education as a standalone path would deliver some benefit quickly, and training the team to use AI drafting tools is genuinely useful as a complement to the recommended approach. On its own, though, it produces inconsistent outputs, no audit trail, and no systematic approval routing, all of which are hard requirements given the firm's privilege and negligence obligations. The value of the AI Workflow path over ad hoc education is precisely that it builds compliance controls into the process itself rather than relying on individual discipline each time.

The starting point is a template audit, and it matters more than it might appear. Before any tool is configured, every filing type the team produces, demand letters, payment reminders, and GDPR responses, needs to be catalogued with its exact input fields, its variable clauses, and the statutory references that must be locked. This work typically takes one to two weeks and is done by the paralegal and one fee earner working through the existing document library. The output is a clear specification: which fields are always present, which clauses change depending on the case type, and which references cannot be altered without a deliberate template update. This specification is what the AI workflow is built from, and gaps caught here are far cheaper to fix than gaps caught after go-live.

Tool selection and data-security sign-off happen in parallel with the later stages of the template audit. The natural candidate for this team is Microsoft Copilot for Microsoft 365, specifically because the firm already runs on Outlook and a Microsoft-hosted document management system. Copilot sits within that existing Microsoft 365 tenancy, which means client documents do not leave the firm's vetted infrastructure to reach an external AI service. Power Automate with AI Builder is a close alternative and offers more explicit workflow logic for routing and approval steps, though it requires a slightly higher configuration effort. Either way, the responsible partner or data protection lead must sign off on data residency and GDPR compliance before any live client data is processed. This step is non-negotiable and should not be treated as a formality.

The pilot build focuses on a single filing type, whichever has the highest volume, and runs for two to three weeks alongside the existing manual process. In practical terms this means the workflow receives the same structured inputs the team currently enters by hand, names, amounts, filing type, relevant dates, generates a draft document using the approved template and AI text generation, and routes it to the designated fee earner via an Outlook task or Power Automate approval action. The fee earner reviews the draft within the document management system, makes corrections if needed, and approves or rejects. Running twenty to thirty live cases in parallel with the manual process before switching over gives the team confidence in accuracy and surfaces edge cases before they become live filing errors. The approval gate is configured as a hard stop: the workflow cannot advance to finalisation without a recorded approval action from an authorised fee earner.

Full rollout across all filing types follows the same pattern established in the pilot, using the validated template structure and the approval routing logic already in place. The final and ongoing stage is monitoring. A simple log, which Power Automate can generate natively, records each draft generated, the approval outcome, any corrections made, and whether the case was escalated as an exception. This log serves a dual purpose: it gives the firm an audit trail for compliance purposes, and it provides the data needed to refine templates over time. Exception rates and the most common corrections made during review should be reviewed monthly for the first quarter and used to improve the underlying templates. The team member responsible for template maintenance needs to be named explicitly, along with a calendar reminder for periodic statutory reference checks, because this is the control that prevents outdated legal text from being reproduced at scale.