Consider a 100-page case file. You can use AI for four very different jobs with the same 100-page case file: sort it, pull the dates and parties, build a chronology, or find where the client interview contradicts the contract.
The file stays the same, but the cost of a miss changes with each task. Choose the model based on what a missed detail could cost and who is responsible for catching it.
One file, four assignments
| Model | Think of it as | Give it | If it misses |
| Haiku | Paralegal | Tagging and extraction across hundreds of documents | A wrong date or party, caught on spot-check |
| Sonnet | Associate | Chronologies, privilege logs, review summaries | A gap in the draft that changes how the team reads the facts |
| Opus | Senior associate | Conflicting evidence, exposure analysis | A missed inconsistency that changes strategy |
| Fable | Partner | Planning and running a multi-step review | A task lands on the wrong step and nobody notices |
Haiku: Start with the stack
When 500 documents need the same questions asked, Haiku can complete the first pass quickly and consistently. It sorts files, pulls dates and parties, and summarizes deposition transcripts.
Its job is to make the record cleaner, not to decide what matters.
Two habits make it trustworthy:
- Require a page citation for every extracted date and party, so reviewers can verify it without rereading the whole document.
- Spot-check a sample before relying on the batch. If the error rate is too high, fix the instructions before moving up a tier.
Sonnet: Give the findings shape
Once the facts are out of the documents, someone has to arrange them. Sonnet turns a client interview into a chronology, organizes review findings, and builds a privilege log from flagged files.
It works best from the cited extraction created in the first pass, rather than the raw documents. A draft is only as reliable as the information it includes.
The attorney’s review should focus on what is missing, since omitted facts can distort a chronology. The model can assemble the log, but the privilege call stays with the attorney.
Opus: Connect the pieces
Say a contract sets a 30-day window, an email thread suggests someone agreed to 45 days, and an interview note says the extension never came up. A summary of each document won’t catch that. Opus can read all three together, surface the conflict, and organize what it may mean for the client’s exposure.
Ask it to cite each conflict, say how confident it is, and flag where the evidence is thin. Attorneys still decide what the record means and what it means for strategy. Opus gets them to a sharper view of the record before that conversation starts.
Fable: Keep the review moving
Some assignments look less like a task and more like a team in motion.
Fable can plan the work, divide it, coordinate research and document analysis, check the results, and back one consolidated package.
The benefit is fewer handoffs. The risk is fewer human checkpoints. Decide up front where it has to stop and ask for approval: before it touches privileged material, before anything leaves the matter, and before the final package goes out. Attorneys approve the work. Fable runs the logistics.
They work as a sequence
On a 100-page file, the models are stages, not alternatives. Haiku extracts and cites. Sonnet turns that extraction into a chronology. Opus compares the chronology with the contract and emails to surface conflicts.
The attorney reviews two checkpoints: a sample of the extraction and Opus’s issues list.
Move up a tier when the output tells you to:
- Reviewers are rewriting the draft instead of editing it
- The chronology has gaps no one can explain.
- Extractions disagree across documents that should match.
The mistakes can run in both directions. Using the top model for a first pass across 800 documents wastes money. Using the lightest model for exposure analysis wastes attorney time and may put the case at risk.
Five questions before you hand over the file
- How much material is there? Volume favors the lighter model for the first pass.
- What kind of work is it? Extraction, organization, and analysis each map to a different tier.
- What does a miss cost? This matters more than the other four questions. One wrong date in a routine review is a small problem. A missed inconsistency before a deposition is not.
- Who reviews it, and how much time do they have? Output no one has time to check is a risk on any model.
- Where does the data go? Sensitive matter information, or access to another system, means permissions must be settled before choosing a model.
Then count the whole cost: model use, review time, corrections, and what gets missed. The cheapest model is not cheap if the reviewer has to redo the work.
Build it once: a starter prompt
Once you know which model fits each job, an orchestrator can route extraction to the lighter model, chronology building to the middle tier, and conflict analysis to the strongest. It can also pause for attorney review at the checkpoints you set.
Use the starter prompt below to design that workflow. Fill in the brackets and test it on a closed or dummy matter, never on live client files. Automatic model routing requires Claude Code or an API-based build.
<prompt box starts here>
| You are helping me design a document review agent for a law firm. The agent sends each stage of the work to the right Claude model and stops for attorney review at set points. Design it, then build it in the environment I name below. Matter type: \\\\\\\[e.g., commercial contract dispute] Typical file size: \\\\\\\[e.g., 100 pages to 800 documents] Environment I can use: \\\\\\\[Claude Code, API, or “not sure, tell me what fits”] Who reviews output: \\\\\\\[role, and how much time they have] Build the agent with these stages:
|
<prompt box ends here>
Three things to change before you use it: the sample size at the first checkpoint, the rewrite threshold that triggers a move up a tier, and the approved system. Those are firm decisions, not model decisions.
The model is the smaller part
A capable model can’t build a dependable legal workflow on its own. It needs reliable source material, secure systems, matter-level permissions, and a review process with names attached. In practice, that means page citations back to the source, access limited to the matter, and a record of who checked what.
That connective work is where we spend our time with law firms. Sikich connects AI to the systems and processes the firm already runs on, so attorneys can stay focused on legal judgment and client service.
The compressed version
Match the model to the cost of a miss. Use lighter models for volume and organization, and stronger ones for analysis and coordination. Set review checkpoints before you start, then build the routing into an agent once the pattern holds. A model fits when the workflow around it is secure, and someone is accountable for the output.
Want to see how Sikich connects AI to your law firm’s systems?
Reach out to our team of legal tech experts today!
This publication contains general information only and Sikich is not, by means of this publication, rendering accounting, business, financial, investment, legal, tax, or any other professional advice or services. This publication is not a substitute for such professional advice or services, nor should you use it as a basis for any decision, action or omission that may affect you or your business. Before making any decision, taking any action or omitting an action that may affect you or your business, you should consult a qualified professional advisor. In addition, this publication may contain certain content generated by an artificial intelligence (AI) language model. You acknowledge that Sikich shall not be responsible for any loss sustained by you or any person who relies on this publication.
