Assistance or integration: how AI enters the firm's workflow
Two operating modes for the same toolbox. One puts AI inside the existing work step, the other rebuilds the workflow around the very same tools.
Most firms adopt generative AI the way they once adopted a spell checker or full-text search: as a tool inside a work step that a person still starts. That works, it is the normal case today, and it is the honest starting point for any discussion of benefits and limits.
There is a second operating mode alongside it. It uses the same models and the same tools but rearranges the workflow around them. The difference is not a question of software but of work organisation, and it shows most clearly where patent work is most sensitive anyway: in the handling of deadlines. This article describes both modes, shows the difference on an office action with a four-month term, and names what shifts as a result and what expressly stays the same.
Two operating modes
Mode A
Assistance
AI supports the person inside the existing step.
AI takes over individual tasks: translation, search, correspondence, text blocks for drafts. Trigger, sequence and division of labour stay as they are. The person starts every step, carries it and remains its author.
Mode B
Integration
The workflow is rebuilt around the same tools.
The incoming event starts the work, preparation runs before anyone opens the file, and the human decision moves to the front. Every step with legal effect stays with a person.
The point where this comparison is often misread: mode B is not better technology. Same models, same failure modes, and the duty to check is identical in both cases. What differs is how the work around them is cut.
Same office action, same term, two runs
The starting position is an examination communication with a four-month term, as the European Patent Office regularly sets (Rule 132(2) EPC allows two to four months for periods set by the Office). Same firm, same file, once in mode A and once in mode B.
Track A. The incoming mail is docketed and the term noted. After that the file sits, because the client instruction is pending and nobody starts without it. What then happens, happens in the last third: search, reconstructing the case, drafting with AI support, approval, filing. The tools are good, they just all sit in a window that has no slack left.
Track B. The same event automatically produces the proposed term and a working file before anyone opens it: the communication, the cited documents, the claims on file, the previous correspondence, a first feature comparison. The strategy decision is taken early and not under time pressure. Drafting and review run separately. Filing happens at the end, and what remains is term buffer.
The amount of work is comparable on both tracks. What moved is where in the term it sits and when the decisions are taken. The real return is therefore not the hour saved but the buffer: time in which a query to the client, a second look at a claim, or a technical failure is still absorbable.
What shifts, and what does not
Kept apart soberly, so the operating mode does not promise more than it delivers:
| Shifts | Stays the same |
|---|---|
| When preparation happens (before the file is opened rather than after) | Hallucinations. A model with tool access still invents, and with tools it acts on what it invented |
| Where the decisions sit (early rather than shortly before expiry) | The duty to check every adopted statement against the source |
| The buffer available before the deadline | Responsibility for term calculation, claim wording and filing |
| Sequence and trigger within the workflow | The confidentiality requirements for the model in use* |
The most important entry is in the right-hand column. Tool access makes assembling the context easier, it does not discharge the duty of evidence: an agent that cites a passage may still have summarised it wrongly, mixed several sources, or dropped a field. How to check that in detail is set out under Duty of evidence; what an agent technically is, and why its isolated context is both a strength and a source of error, under AI agents.
What mode B requires in practice
Less software than expected, and more organisation. Four points decide whether the rebuild holds:
- An instruction regime that covers preparation. If the working file is produced before the client has responded, the engagement terms must cover that preliminary work, including who bears it when the instruction turns out to be "do not pursue".
- A review function separated from the drafting. The same model that wrote the draft is a poor reviewer of its own text. Separate review means a different pass, a different prompt, a different model, or simply a person with the original document next to them.
- A deliberate approach to confidentiality. Event-driven preparation means case material passes through tools without a case-by-case decision. What leaves the firm in the process has to be settled beforehand, not afterwards.* On the local route see Prompting for local models.
- Deadlines stay in the docketing system. An automatically calculated term is a proposal. Recording and monitoring stay in the established procedure, with four-eyes review and no special path for machine-generated dates.
Technical detail: what "event-driven" actually means
The trigger usually exists already: the docketing or case management system records the arrival of an official communication anyway. Event-driven only means that at this point a preparation run is also started, instead of waiting until a person opens the file.
Technically this needs three building blocks: a trigger (a new entry in the case system, a file arriving in a watched folder, a notification from the office), access to the relevant material (file, claims, cited documents, registers), and a model that turns it into a structured working file. The access part is the laborious one, not the model. What such a connection to patent registers looks like via the Model Context Protocol is described under MCP and patent data.
None of this requires a new product. The difference to mode A is that the work hangs on an event rather than on free capacity.
Where mode B has no business
The line does not run between important and unimportant files but between preparation and legal effect. What can be discarded at any time can be prepared automatically. What acts outwards, or contains an assessment a person stands behind, is not automated: the wording of a claim, the filing, setting a term towards the client, any statement to an office or a court.
That is not a cautionary formula but the condition for the rebuild to remain permissible at all. An agent that sends mail or files documents on its own creates actions nobody oversees any more, and responsibility for them stays with the professional.
A realistic entry point
The rebuild pays off as a slice, not as a programme. One event type is enough, for instance the examination communication. For that single type it is defined what the preparation contains (communication, cited documents, claims on file, previous correspondence, a feature comparison table), who releases it, and when the strategy decision is taken.
The useful yardstick is not time saved but distance to the deadline: how many days before expiry the approvable draft was ready, measured over ten files before and ten after. That number describes exactly what mode B is supposed to deliver, and it is hard to massage.
* The processing of confidential content is subject to professional-conduct confidentiality requirements, among others. These can be met by a locally hosted model or, on a case-by-case basis, by a trusted or self-hosted cloud instance. This does not constitute legal advice.