I automate aggressively across my firms. Deadline calculation is one of the few places I have deliberately left alone, and the reason is specific rather than general caution.
Most AI errors announce themselves. A weak draft reads weak. A bad summary feels thin. You notice, and you fix it. A wrong deadline looks exactly like a right deadline. It is a date. It goes on a calendar. Nobody discovers the error until the day it matters, and by then the remedy is a call to your carrier.
Why deadline computation is unusually hard
It looks like arithmetic and it is not.
Which event starts the clock is frequently the actual question, and it is a legal judgment. Service, filing, entry of the order, and receipt can all be candidates, and picking the wrong one moves everything.
Counting conventions vary. Calendar days against court days, whether the triggering day counts, how weekends and holidays are treated, and what happens when the last day falls on a closure. These rules differ across jurisdictions and sometimes within them.
Local rules override general rules, and they are the least well documented material in the system. A federal district’s local rules and a particular judge’s standing order can each change the answer.
Service method can add days. Rules change. A model trained on material from two years ago will confidently apply a rule that has since been amended, and it has no way to tell you it is out of date.
None of that is beyond a computer. It is exactly what rules-based docketing software does, because someone is paid to maintain those rule sets when courts change them. That maintenance is the product, and it is what a general model has no equivalent of.
Where AI does help
The calculation is off limits. The work around it is not, and there is more of that work than people realize.
Reading orders and extracting obligations. Feed it a scheduling order and ask for every date and obligation in a structured list. It is good at this, and it is faster and more thorough than a person skimming at 6pm. A human then confirms each date against the docketing system, which is a verification step rather than a computation.
Drafting the internal summary. What this order requires of us, in plain language, for the file.
Surfacing what is coming. With a read-only calendar connector as described in MCP servers for law firms, you can ask which matters have deadlines in the next ten days, or which have none scheduled at all. That second question is the useful one, because a matter with no upcoming deadline is often a matter someone forgot to docket.
Catching the absence of an entry. AI is genuinely good at noticing that a case with a filed complaint has no answer deadline calendared. That is pattern matching against your own data, not rule interpretation.
The division that works
| Task | Tool |
|---|---|
| Computing the date | Rules-based docketing software |
| Reading an order and listing obligations | AI, human verifies |
| Entering the date | Human or docketing integration |
| Reminding before it arrives | Case management |
| Finding matters with no deadlines set | AI against your calendar |
| Deciding what a rule means | A lawyer |
If you are a solo without docketing software
Many small firms compute deadlines manually and would be tempted to use AI as an upgrade. It is not an upgrade. It is a lateral move to a more confident version of the same risk.
The better path is one of the established rules-based products, which are not expensive relative to a single missed deadline. If that is genuinely not possible right now, the minimum is a written calculation checklist and a second person confirming every date, which is what the software would be replacing.
You can still use AI to read the order and produce the candidate list. Just do not let it do the counting, and make the verification step explicit rather than assumed.
Do this today
Take the last scheduling order you received. Ask AI to extract every date and obligation into a list.
Then compare that list against what is actually on your calendar for the matter. Firms that run this comparison usually find at least one obligation that was never calendared, and that is the real value here: not computing dates, but noticing what nobody entered.