Mage vs. Kira: How They Compare for M&A Diligence

Key Takeaways

•Kira pioneered the supervised-classifier approach to contract clause extraction; Mage uses LLM-based reasoning over structured extraction.

•Kira's strength is its mature library of pre-trained clause models built over a decade. Mage's strength is workflow ownership end-to-end (memos, schedules, redlines).

•On modern diligence work, Mage handles amendment chains and multi-document reasoning natively; Kira's classifier-per-clause architecture is built for single-document extraction.

•Both have legitimate places in the market. The choice depends on whether the team wants extraction tools or end-to-end workflow.

Kira (now part of Litera) was one of the original legal AI companies. Founded in 2011, the platform built its reputation on supervised-learning clause classifiers — a deep library of clause models trained over a decade of labeled examples. For a long stretch of the 2010s, Kira was the default answer to "what does AI for contracts look like?"

The category has moved. LLM-based reasoning over structured extraction is the modern approach, and the products built natively on that architecture (Mage among them) are differently shaped than the supervised-classifier tools that came before.

This is an honest take from a vendor on the LLM-native side of that line.

Architecture, briefly

Kira's foundational technology is a library of supervised-learning models. Each clause type has its own classifier, trained on labeled contract examples. To find termination clauses across 500 MSAs, the platform applies the termination-clause classifier to each contract; the output is the extracted text plus a confidence score. The classifier library is the moat: hundreds of pre-trained models built over more than a decade.

Mage uses frontier LLMs for clause-level reasoning, with structured extraction passes that handle multi-document context. To answer "what is the current operative termination provision in this multi-amendment MSA?", Mage processes the original agreement and each amendment in sequence, tracks which provisions have been added or modified at each step, and produces a resolved view. The architecture is less about clause-by-clause classification and more about reasoning over the full document graph. We have written about the architecture in detail in Amendment Chain Resolution: The Hardest Problem in Legal AI.

Both work for finding clauses in single documents. They diverge on multi-document reasoning, novel clause patterns, and end-to-end workflow.

Strengths and tradeoffs

Kira's strengths: Mature, well-tested, deterministic outputs. The supervised-classifier approach has well-understood error modes — when it misses a clause, it misses in predictable ways the team learns to compensate for. The clause library is deep, and for the clause types it covers well, accuracy is strong.

Kira's tradeoffs: Single-document orientation. The architecture treats each contract as its own classification target. For multi-amendment commercial contracts (the dominant case in real M&A data rooms), the team has to manually piece together what the current state of any given provision is. The platform is also primarily an extraction tool; downstream work (memo drafting, schedule synthesis, redline review) typically happens outside it.

Mage's strengths: Native multi-document reasoning. Amendment chain resolution is built in. End-to-end workflow ownership: the same platform that finds the issues drafts the memo, prepares the schedules, and runs redline review. Configurable per-deal risk lists rather than a fixed library. Modern LLM-based language handling for novel and edge-case clause patterns.

Mage's tradeoffs: The category is younger. The pre-trained library of "every clause we've ever seen" is shallower. Hallucination risk is non-zero in any LLM-based system; we have specific architectural guards against it (covered in LLM Hallucination in Contract Analysis) but the issue is real and the engineering investment to manage it is ongoing.

Workflow fit

For a firm's typical M&A deal, the question is not "which tool extracts clauses better?" but "which tool gets the deal team to a partner-reviewable memo and schedule fastest, with which level of confidence?"

Mage is built around that question end-to-end. Kira is positioned as a strong extraction component within a workflow the firm builds around it. Both are valid; they are different products.

For more on the workflow shape, see AI Due Diligence: An Operational Playbook.

How firms tend to choose

A firm with a mature M&A practice that already has a heavily-customized Kira deployment will weigh the migration cost against the workflow upside. We have seen both directions: firms migrate to Mage when the M&A workflow is the bottleneck; firms stay on Kira when the existing extraction workflow is well-tuned and the team isn't asking for end-to-end ownership.

A firm starting a legal AI rollout fresh today, in our view, is making a category-shaped decision rather than a vendor-shaped one. LLM-native tools are the modern category. Picking inside that category is the second-order question.

How to evaluate honestly

Same framework as our other comparison content: run the same deal through both tools, in parallel with the manual workflow. Compare deliverables against ground truth. We have laid this out in Evaluating Legal AI Tools.

The bar should include: amendment chain handling, end-to-end deliverables (not just extraction), output voice that matches the firm's house style, and time to partner-reviewable memo. The tool that wins on those metrics on real deals is the right tool for the firm.

If you would like to see Mage on a real deal: request a demo. Bring the data room. We will run end-to-end diligence and walk through the result against your manual workproduct.

For our broader competitive view, see Legal AI vs. Harvey vs. Generic AI.

Frequently Asked Questions

What is Kira's architectural approach?

Kira's foundational technology is supervised-learning clause classifiers. Each clause type has a model trained on labeled examples; the platform applies the relevant classifier suite to each contract and returns extracted clauses with confidence scores. The approach is mature (Kira was an early mover, founded 2011) and has a deep library of pre-trained classifiers across hundreds of clause types.

How does Mage's architecture differ?

Mage uses frontier LLMs for clause-level reasoning, layered with structured extraction passes that handle multi-document context (amendment chains, cross-referenced provisions, schedule-to-agreement linkage). The advantage of LLM-based reasoning is flexibility on novel clause types and language nuance; the cost is you have to design carefully for hallucination prevention. We discuss our approach in [LLM Hallucination in Contract Analysis](/blog/llm-hallucination-in-contract-analysis).

Is supervised classification 'old' technology?

Not exactly — it's mature technology, with strengths LLMs do not have (deterministic output, well-understood error modes). The category has shifted toward LLM-based reasoning, though, because LLMs handle the long tail of clause variations better and let the same engine serve multiple workflows (extraction, drafting, summary).

Which is better for M&A specifically?

Depends on the workflow. For pure extraction (find every change-of-control clause across 1,000 contracts), both work. For end-to-end deal workflow (extract, reason about amendment chains, draft memos and schedules, review redlines), Mage is built for that loop; Kira historically positions as the extraction component within a workflow you build around it.

Can a firm run both?

Theoretically, yes. In practice, running two contract analysis tools in parallel multiplies operational complexity (deal data in two places, two sets of access controls, two vendor relationships) without a clear additional benefit. Firms tend to pick one.