AI Lease Abstraction: Kira vs Luminance vs Custom Agent

By Jude Lee · · Comparison

Lawyers and a paralegal comparing lease documents against extracted key terms on a laptop in a small firm conference room

Abstraction is not review, and the distinction matters

Contract review asks is this term acceptable? Abstraction asks what does this document say, in fields I can sort? A commercial lease abstract typically captures parties, premises, commencement and expiration dates, base rent and escalations, CAM/operating expense treatment, renewal and termination options with notice windows, assignment and change-of-control restrictions, insurance minimums, and guaranty terms.

That output is a row in a table, not a memo. Which is why tools optimized for redlining — Spellbook, LegalOn and similar review assistants — often disappoint on abstraction work, and vice versa. If your real estate, franchise, or business practice ends every diligence engagement with someone retyping dates into a spreadsheet, abstraction is your automation target, not review.

The few fields that carry most of the value

There’s an old management heuristic lawyers borrow constantly — the 80/20 rule, the idea that a small share of inputs drives most of the output. Applied here, it’s the most useful planning move you can make: in most abstraction projects, a handful of fields carry nearly all the downstream risk and client value. Dates that trigger notice obligations. Rent and escalation math. Options and their exercise windows. Assignment consent. Everything else is nice to have.

So scope the pilot to those fields, get them near-perfect, and add the long tail later. Firms that try to abstract sixty fields on day one usually get sixty mediocre fields.

8–15
high-value fields to abstract first — illustrative planning range, build your own list
Illustrative, not a benchmark

Alongside that, three ground rules we’d argue for on any abstraction pilot:

What the purpose-built platforms actually do

Litera’s Kira (originally Kira Systems) and Luminance both grew up as document-analysis platforms for diligence: you point them at a document set, they identify provisions using trained models, and you review extractions in a side-by-side interface. Kira is known for a library of pre-trained provision models plus the ability to train your own; Luminance markets analysis across the contract lifecycle.

They aren’t the only shelf options. Extraction also shows up natively in document and contract systems your firm may already license — iManage’s Extract/Insight line (built on the former RAVN technology), DocuSign Insight (built on the former Seal Software analytics), and CLM platforms such as Ironclad, Icertis, or Evisort (now part of Workday). Naming and packaging in this market change fast, so verify current product names, seat minimums, deployment options, and pricing directly with each vendor’s documentation rather than trusting any blog’s numbers, including this one.

What you’re buying is a mature review UI, provenance built in (click the field, see the clause), and a support relationship. What you’re often not buying is a clean automated write-back into your matter records, or pricing that makes sense for a firm doing twenty leases a quarter rather than two thousand documents a deal.

The custom path: a skill plus MCP

The alternative is assembling the same capability from general-purpose parts. Concretely, that means three things.

A skill — a written, versioned instruction set that teaches an AI assistant to abstract your way every time: the field list, the definitions, the tie-break rules (which document controls when a lease has three amendments), the required output format, and the rule that every field must cite a page and section or come back marked “not found.” This is the same pattern as reusable skills for standardized document review, applied to extraction.

MCP — the Model Context Protocol, an open standard for giving an AI assistant governed access to specific tools and data. It’s what lets the assistant read the lease from your document management system and write the abstracted fields back into the matter, instead of you copying JSON around. If you’ve already looked at connecting Claude to Clio through MCP, this is the same plumbing pointed at a new job; a firm with its own document store may need a custom MCP server over its matter data.

An agent loop — the assistant works document by document: locate the operative lease, gather amendments, extract fields, flag conflicts between the original and later amendments, write the record, and post an exception list for a human.

Be honest with yourself about what standing this up costs: expect real developer time to wire and secure the MCP connection to your DMS and practice management system, plus a named person inside the firm — usually a senior paralegal or practice-group lead — who owns the skill, the field definitions, and the quarterly re-scoring. That ownership isn’t free, and if nobody has capacity for it, buy instead.

An abstract that can’t point to the page it came from isn’t an abstract — it’s a rumor with a due date attached.

Comparing the two builds honestly

Off-the-shelf extraction platform
Fastest to first output; no engineering. Provenance UI and reviewer workflow already solved. Vendor handles model updates. Weaknesses: quote-based enterprise pricing that can overwhelm a small firm’s budget, field taxonomies you adapt to rather than define, and integration into your practice management system that may be manual export/import.
Custom agent (assistant + skill + MCP)
Your field schema, your tie-break rules, output written directly into Clio/NetDocuments matter records. Can be cheaper per document at low volume once the build time is amortized — but only you can model that, because the offsetting variables are the initial development cost and the ongoing hours your internal owner spends on QA and skill maintenance. Reusable for adjacent jobs (franchise agreements, MSAs, loan docs). Weaknesses: you own quality assurance, access controls, and there’s no vendor to escalate to.

And a third option people skip: keep it manual. If you abstract a dozen leases a year, a good paralegal with a well-designed template will beat any tool on total cost and probably on accuracy. Automation earns its place at repetition.

Where these systems break

The honest disadvantages of AI in this workflow are specific and predictable:

Building the pilot

  1. Freeze a field list

    Write the 10–15 fields, with definitions and controlling-document rules. This document is the spec for every option you evaluate.

  2. Build a gold-standard set

    Have a person abstract 15–25 real documents by hand, including two ugly ones. This is your scoring key — without it, every vendor demo looks great.

  3. Score both paths on the same set

    Run the platform trial and a prototype agent against identical documents. Count wrong values, missing values, and hallucinated values separately.

  4. Require citations and confidence flags

    Any field the system can’t tie to a page and section comes back flagged, not guessed. Route flags to a reviewer queue.

  5. Wire the write-back last

    Only after accuracy is acceptable should the agent write into matter fields. Keep an audit trail of what it wrote and who approved it.

  6. Re-score quarterly

    Models and documents both change. A one-time validation ages badly.

Modeling the payback without making up numbers

Don’t take anyone’s hour-savings claim, including a vendor’s. Model it yourself:

(baseline minutes per abstract − reviewed-agent minutes per abstract) ÷ 60 × documents per year = hours recovered

Then decide what those hours are actually worth to you. If the work is billed flat-fee on diligence, the gain shows up as margin. If it’s hourly, it shows up as capacity — and capacity only converts to revenue if you have demand waiting. Multiply recovered hours by your realized rate, not your rack rate, and subtract the build or subscription cost plus your owner’s maintenance hours. Firm economics vary enormously by practice area, geography, and leverage, so plug in your own realization and utilization rather than any headline figure. Our law firm automation ROI walkthrough covers the arithmetic in more detail.

There is no single best tool — pick the task first

So what is the best AI program for a law firm? There isn’t one, and any list that names a single winner is selling something. Abstraction, research, intake, and billing are different jobs with different tools. As of 2026, the pattern that holds up for small and mid-size firms is narrower than the hype: pick one repetitive, structured, verifiable task; instrument it with provenance; keep a human on the money-and-deadline fields; and expand only after the first workflow is boring. Buy the platform when volume is high and your documents are conventional. Build the agent when the output has to live inside your own systems, your field list is genuinely yours, and you have someone who will own it. Do neither when a paralegal and a good template already clear the work.

Where is your firm losing billable hours?

Get a free automation audit: we map your intake-to-invoice workflow and show you exactly what's worth automating — before you spend a dollar.

Get a free automation audit

Related guides