Every legal AI budget conversation eventually arrives at the same question: what are we actually getting for this? The honest answer requires measurement discipline that most rollouts never set up: not because it is hard, but because nobody owns it. This framework is the setup: what to measure before rollout, the four channels where value actually appears, worked examples for a firm and an in-house team, and the mistakes that quietly corrupt the numbers.
Why legal AI ROI measurement usually goes wrong
Three structural problems sink most attempts. First, no baseline: if nobody recorded how long a first-pass data-room review took before AI, the "we're faster now" claim has nothing to stand on. Second, wrong unit of value: hourly-billing firms fear that efficiency destroys revenue, so they measure nothing rather than confront the question: missing that realization, capacity, and fixed-fee margin are where the value lands. Third, selection effects: measuring the enthusiastic pilot team and extrapolating to the whole firm inflates every number, because pilot volunteers adopt harder than the median lawyer. The framework below is built to survive all three.
Baselines first: measure before you roll out
Pick the two or three workflows the platform is supposed to improve: the same ones you scoped in your evaluation: and record current-state numbers for each: typical hours per task, who does it, and at what rate or cost. Keep it honest and rough; four to six weeks of matter-level data beats a perfect study that never happens. Examples of useful baselines:
- First-pass review of a standard contract batch: hours per document, by seniority.
- A research memo with verified citations: hours from question to memo.
- Due diligence extraction for a data room of a given size: associate-days.
- A litigation chronology from a document set: hours, plus how often deadlines were caught late.
- Turnaround time on routine contracts (in-house): request to signature.
These baselines are the denominator of every claim you will make later. Without them you are quoting vendor marketing back to your own CFO.
The four value channels
Value from legal AI arrives through four distinct channels, and conflating them is how numbers get gamed: or dismissed:
| Channel | What it is | How to measure it |
|---|---|---|
| 1. Time on task | The same work, done faster | Baseline hours minus measured hours on comparable tasks, times a realistic cost rate |
| 2. Capacity absorbed | More work with the same team | Matters or contracts handled per period, before vs after; external spend avoided |
| 3. Risk avoided | Errors and omissions caught | Deadlines flagged, adverse clauses caught, coverage of documents actually reviewed: tracked as counts, valued conservatively |
| 4. Revenue enabled | Work you could not offer before | Fixed-fee matters won on efficiency; realization improvement on research-heavy matters; new service lines |
Channels 1 and 2 are where measurement starts because they are cleanest. Channel 3 should be reported as events, not invented currency ("the timeline flagged a limitation date that had been missed in manual review" is worth more in a partner meeting than a fabricated dollar figure). Channel 4 shows up last but compounds.
Use external research as context only after checking its sample and method. A survey of expected savings is not measured savings in your practice. NIST’s AI RMF can help organise evaluation responsibilities; the financial model here is our own illustrative method, not a NIST return-on-investment formula.
Worked example: a mid-sized firm
Invented arithmetic example, not a customer result: assume a document review takes 45 hours manually and 18 hours with AI assistance including verification; a memo takes 8 versus 3.5; and a chronology takes 12 versus 4. These are teaching inputs, not measured Judicio outcomes.
| Workflow | Hours saved per task | Monthly tasks | Illustrative capacity |
|---|---|---|---|
| Document review | 45 − 18 = 27 | 2 | 54 hours |
| Research memo | 8 − 3.5 = 4.5 | 6 | 27 hours |
| Chronology | 12 − 4 = 8 | 3 | 24 hours |
| Total | 105 hours |
The arithmetic yields 105 hours of capacity in this invented example. Capacity is not cash saved or revenue earned. Measure whether the freed time is redeployed, removes a real cost or affects a fixed-fee margin, then deduct software, training and administration costs without double-counting verification already included above.
Worked example: an in-house team
Invented in-house example: a team compares request-to-completion time before and after a pilot. Record comparable task groups, their size, total elapsed time, reviewer effort, corrections and work referred externally. Exclude no failed task merely because it makes the result less attractive.
If a task remains in-house, compare the avoided external invoice with the actual additional internal effort and service costs. If turnaround improves, report the measured change with its period and sample. Do not attach a monetary value to a hypothetical avoided dispute or assume that every hour of capacity becomes a budget reduction. The pilot worksheet supplies blank measurement fields.
The costs side: count all of it
Credibility comes from counting costs as carefully as savings: subscription and usage fees (model a realistic month: transparent credit pricing makes this easy; opaque schemes are themselves a cost signal); onboarding and training time at real hourly value; the ongoing verification time that professional use requires (it is real work: count it in the measured "after" hours, which is also why citation-first tools that make verification fast score better on true ROI); and administration: template setup, policy, usage review. A platform that publishes exactly what each tier includes simplifies this arithmetic; one that hides audit trails or key workflows behind unlisted enterprise pricing buries costs in the fine print.
The mistakes that corrupt the numbers
- Phantom savings: valuing "saved" hours that were never billable or never happened. Tie hour claims to the baseline tasks.
- Pilot-team extrapolation: your keenest adopters are not the median. Report adoption rate alongside per-user savings.
- Ignoring verification: quoting AI drafting time without lawyer review time. The honest metric is time-to-verified-output.
- One-quarter verdicts: portfolio effects (realization, spend) lag a quarter or two behind task effects. Set expectations accordingly.
- Trophy metrics: "documents processed" and "queries run" measure activity, not value. Hours, cycle time, coverage, and spend are the currencies that survive a CFO's questions.
How Judicio supports ROI measurement
Judicio gives the measurement programme its raw data. Project analytics in Collaboration (an Organisation-tier feature) show usage by feature, by member, and over time (so adoption (the number that predicts everything else) is visible, not guessed) and the activity trail records who ran what, on which files, and when. The workflows you baseline map one-to-one onto platform features: Document Review for batch first passes, Research for cited memos, the Review Matrix for extraction, and the Timeline Builder for chronologies with deadlines flagged.
Record verification time directly; citations can make source access easier but do not establish how long review will take. Include usage, subscription, training and correction costs using the current plan details. A small trial is a starting point for measurement, not evidence of an organisation-wide return.
Worked examples are illustrative methods, not promised outcomes. Third-party figures are attributed to their publishers; your results depend on workflows, adoption, and measurement discipline.