There is no universal return-on-investment figure that tells a law firm or in-house team whether legal automation will pay off. To calculate a defensible case, define one workflow, measure its current cost and quality, include human review and exceptions in the assisted state, count the full cost of ownership, and show how any released capacity becomes useful or financially realised. Saved minutes are evidence of a process change; they are not automatically cash savings.

Start with the decision, not a vendor percentage

Write down what decision the ROI model must support. Is the team deciding whether to run a pilot, renew a tool, replace an outsourced intake provider, expand a contract-review workflow or build internally? The decision determines which costs and outcomes matter. A firm evaluating a fixed-fee review workflow will not value capacity exactly like an in-house team trying to reduce a backlog.

Define the unit of work. Examples include a contract request from submission to routing, review of a standard supplier agreement, extraction of a defined clause set, or a legal intake matter from submission through accepted assignment. Do not benchmark “legal work” as a whole.

Set the cohort, period, complexity bands, start and end events, eligible and excluded work, reviewer roles, output criteria, error severity, exception route and system version before collecting data. The NIST AI Risk Management Framework and its Playbook provide risk-management activities teams can adapt; they do not prescribe an ROI figure or replace legal and finance review.

Separate six different outcomes

OutcomeWhat is measuredWhat evidence supports it
OperationalActive time, elapsed cycle time, volume, backlogWorkflow timestamps and representative samples
CapacityQualified time made available for other workNet time difference, adoption and a named use plan
CashSpend actually avoided or reducedBudget, invoice, contract, staffing or procurement evidence
Revenue or marginCollectible additional work or improved economicsMatter, billing, collection, pricing and attribution records
Quality and riskCompleteness, errors, rework, escalation and impactReviewed outputs with severity rules and incident data
Service and adoptionPredictability, requester experience, use and abandonmentService data, user feedback and workflow telemetry

Do not label capacity as cash. A legal team may use reclaimed hours to reduce backlog or improve response quality, which may be valuable without reducing payroll. A firm may increase capacity but have no additional collectible work. State which outcome the model measures and which remain hypotheses.

Build a baseline from the current workflow

Select tasks that reflect actual work, not only the cleanest examples. Include ordinary, difficult and exception cases: incomplete intake, ambiguous contract language, scanned documents, unusual counterparties, missing approvals, cross-references, urgent requests and matters that should be escalated or declined.

For each task, record:

  • work type, complexity, source, jurisdiction and cohort;
  • active requester, lawyer, reviewer and administrator time;
  • elapsed time from accepted request to agreed completion;
  • handoffs, queue time, clarification rounds and approval delay;
  • rework, correction, exception, escalation and fallback;
  • outcome quality, severity of any error and acceptance;
  • outsourced, software, model, storage and service costs; and
  • the staffing and process conditions that could affect comparison.

Use a documented sample rather than one memorable matter. Compare medians and ranges by complexity; an average can conceal a small number of very slow or costly exceptions. Record the sample size and uncertainty. Do not claim causation if staffing, intake mix or policy changed during the same period.

Measure the assisted process end to end

Calculate net assisted effort as production plus prompt or configuration work, source review, legal verification, correction, exception handling, communication, and manual fallback. Include abandoned runs where users tried automation but returned to the old process. Measure elapsed time as well: faster drafting may not shorten the cycle if approvals or missing facts remain the bottleneck.

For a legal intake tool, count initial requester effort, clarification questions, legal triage, conflict checks, sensitive-route review, assignment, approval and closure. A form that reduces email volume but creates more missing-information follow-up may not reduce total effort.

For AI contract review, count the time to provide context and the correct contract version, extract issues, verify citations to clauses, identify missing schedules, resolve escalations, generate a usable review note, and approve delivery. Include false positives and missed issues in quality measurement. Do not count model output as completed legal work until the responsible reviewer accepts it.

How to calculate ROI for an AI contract review tool

Use the same period for benefits and costs. First separate a cash case from a capacity case, then apply a transparent formula:

ROI (%) = (attributable benefit over the period − total cost over the period) ÷ total cost over the period × 100.

For a first-year cash model, total cost includes implementation and transition as well as the recurring licence, integration, administration and reviewer effort that are incremental to the baseline. Count an avoided expense only where the team can explain the actual budget or invoice change. If the benefit is released lawyer time, label the output as a capacity-value scenario and state how that time will be used. Keep incremental costs consistent so review effort is not counted twice.

Fictional arithmetic example, not a Gotham result or market benchmark: assume a finance-approved scenario attributes ₹120,000 of annual expense reduction to a workflow, with ₹60,000 of annual incremental recurring cost and ₹20,000 of one-time implementation. First-year total cost is ₹80,000; net benefit is ₹40,000; first-year ROI is 50%. A later year with the same benefit and ₹60,000 cost would model 100%. The difference comes from the period and one-time cost, not a new productivity claim. Replace all assumptions with the organisation's evidence and revisit them if the work mix changes.

Alongside the percentage, show eligible contracts, actual adoption, total assisted review minutes, exception rates and the quality threshold. A spreadsheet with only licence cost and headline minutes saved is missing the work required to reach an approved deliverable. Use the legal AI ROI calculator for a scenario and retain the input sheet, period and source of each assumption.

Use the capacity-to-cash bridge

Move from task-time savings to a value claim through explicit steps:

  1. Eligible volume: how many tasks meet the approved workflow scope?
  2. Observed net difference: what is baseline active time minus assisted total effort, including review and exceptions?
  3. Adoption: what share of eligible work actually uses the workflow?
  4. Sustainable capacity: what remains after training, admin, support, governance and workflow maintenance?
  5. Capacity use: which backlog, service, outsourced work, revenue work or staffing decision will change?
  6. Financial realisation: what invoice, budget, headcount or collected revenue proves the economic value?

Fictional example, not a benchmark: A firm completes 80 eligible contract reviews a month. Baseline median active effort is 100 minutes. Assisted production, verification, correction and exceptions total 70 minutes, so the observed net saving is 30 minutes per accepted review. Adoption is 50%, yielding 20 hours of gross monthly capacity (80 × 50% × 30 minutes). If training and operations consume five hours monthly, 15 hours remain. If the firm cannot redirect these hours to paid work, avoid a vendor invoice or reduce backlog, it has measured capacity, not cash savings. Replace each assumption with observed cohort data.

Do not multiply reclaimed time by a standard hourly billing rate unless that value is achievable and attributable. A loaded salary figure may estimate capacity cost, but it does not prove a budget saving. Finance should approve cost allocation and any claim of realised savings.

Suppose a company pays an outside provider a monthly fee to receive and classify routine legal requests. The team considers internal automation. Compare the current provider cost with the complete alternative: software, implementation, integration, security review, internal triage, escalation, training, support, and remaining external overflow. Count the provider tasks that are actually eligible; complex, sensitive or conflict-prone requests may still need specialist handling.

Measure whether routing accuracy, completeness, response time and requester effort remain acceptable. A lower invoice is not a net saving if Legal must spend more time repairing misroutes or collecting missing facts. A fictional calculation might show annual fee reduction of ₹6 lakh but incremental internal, software and transition costs of ₹5 lakh; the provisional cash difference is ₹1 lakh before quality effects and one-time implementation. This illustrative number is not a market benchmark or promise. Include sensitivity if volumes, exceptions or contract terms change.

Example: AI contract review ROI

For an AI contract review tool, choose one contract family and a defined set of clauses. Create a baseline sample that includes normal and difficult agreements. Record reviewer minutes, negotiation rounds, missed or misunderstood provisions, source verification and time to final approved summary. In the assisted sample, include time spent uploading, correcting document structure, reviewing source passages, resolving uncertainty and checking the final work product.

Assume in a fictional pilot that 40 eligible agreements per month take 90 minutes each today. With the tool, total human work is 65 minutes per agreement, including verification; 30 of the 40 agreements are actually completed through the assisted process. Gross capacity difference is 12.5 hours monthly (30 × 25 minutes). If the licence, integration and support consume more than the value the team can realise from those hours, the case may still support a service or quality improvement, but it is not yet a positive cash ROI. If a missed clause or unsupported summary creates material risk, quality guardrails may veto scale regardless of time savings.

Count the full cost of ownership

Include one-time and recurring costs, internal and external work, and the transition period:

  • procurement, legal, security, privacy and risk review;
  • configuration, integration, data preparation and migration;
  • training, workflow design, change management and user support;
  • reviewer, administrator, supervisor and system-owner time;
  • model, usage, storage, hosting, API and vendor charges;
  • evaluation, monitoring, audit, incident response and revalidation;
  • exceptions, manual fallback, duplicate work and remediation;
  • export, retention, deletion, switching and exit costs; and
  • taxes or other finance-approved expenses.

The legal AI total cost of ownership guide, build-versus-buy framework and legal AI RFP template help structure these inputs before procurement.

Use the legal AI ROI calculator to compare your measured baseline with assisted effort, adoption, verification costs and implementation assumptions. Treat the output as a scenario model and replace example inputs with observed values before making a spending decision.

Set quality gates and sensitivity cases

Define disqualifying failures before the pilot: unsupported legal conclusion, missed material clause, wrong party or version, unauthorised disclosure, deadline error, untraceable citation or failure to escalate a sensitive request. A favourable average time score should not erase a severe failure. The NIST AI Risk Management Framework's Measure function discusses testing, documenting metrics and monitoring; apply the relevant controls to the actual system and workflow.

Model conservative, expected and upside cases. Vary eligible volume, adoption, baseline time, assisted review, exceptions, implementation duration, recurring cost, quality remediation and capacity use. Show break-even volume and the assumptions with the greatest impact. If the result works only when adoption is near universal, reviewers do no correction and every hour becomes cash, the model is not a credible expected case.

Use a decision table set before seeing results:

ResultEvidenceNext decision
StopDisqualifying quality failure or no plausible value pathEnd or redesign the use case
ExtendSample or confidence is too limitedRun a bounded additional test
RemediateSome value, but adoption, controls or integration failFix named gaps and measure again
Proceed conditionallyGuardrails pass and expected case supports a narrow scopeLimit use and monitor defined outcomes
ScaleProduction evidence remains stable across users and complexityExpand in controlled stages

ROI checklist

  • The decision, workflow, cohort and outcome are defined.
  • Baseline includes active effort, elapsed time, review, rework and exceptions.
  • Assisted measurement includes verification, fallback and abandoned attempts.
  • Quality gates and stop conditions were set before results were known.
  • Capacity, cash, revenue, service and quality are reported separately.
  • Outsourced cost comparisons include internal replacement effort.
  • Full one-time, recurring, transition and exit costs are included.
  • Conservative, expected, upside and break-even cases are visible.
  • Finance, legal, security, workflow and user owners validate assumptions.
  • System or process changes trigger remeasurement.

The strongest business case is a model another reviewer can reproduce: defined work, measured evidence, disclosed uncertainty, full costs and a realistic value-realisation path. This article is general operational information, not financial, legal or procurement advice. Validate assumptions with qualified finance, legal, risk and technology owners. See legal AI pilot planning, accuracy evaluation, or contact Gotham to discuss a workflow baseline.

Questions about legal automation ROI

How do I evaluate ROI of legal technology and AI for my law firm?

Choose one workflow and cohort, measure baseline and assisted total effort, include review and exception handling, assess quality, count full costs, and identify how capacity becomes paid work, reduced expense or better service. Keep capacity separate from realised cash and test the result under conservative assumptions.

How do I calculate ROI for an AI contract review tool?

Compare the same contract types and complexity before and after. Include document preparation, AI use, lawyer verification, corrections, escalations and final approval. Measure both time and issue quality, then include licence, implementation, training, integration, support and monitoring costs. Do not treat unreviewed output as completed work.

How should outsourced legal intake ROI be calculated?

Compare the current provider's full fees and service outcomes with the complete alternative cost, including software, internal triage, exception handling, support and overflow. Measure routing quality, requester effort and time to assigned owner. A vendor fee reduction is not net value if internal repair work absorbs the difference.

Does time saved equal money saved?

No. Time saved first indicates capacity. It becomes cash only when it changes a budget, invoice, staffing decision or other realised financial outcome. Otherwise report capacity and specify how the team plans to use it.

Which tasks should be included in the pilot?

Include eligible ordinary work and representative hard cases, exceptions and refusals. Exclude work outside the approved scope but track how often it occurs. A clean demonstration-only sample will overstate likely production performance.

Should ROI use average time or median time?

Use measures appropriate to the distribution and show ranges. Medians can describe a typical task when a few outliers distort the average; averages may matter for total workload. Segment by complexity and publish sample size so neither metric hides the tail.

What if the tool saves time but makes errors?

Apply pre-set quality gates. Classify errors by severity and remediation, and stop or narrow the use case when material failures exceed the approved threshold. A positive time result does not offset every legal or operational risk.

This article is general evaluation guidance, not financial, tax, procurement or legal advice. Use organisation-specific baselines and validate assumptions with accountable finance and legal owners.
How Gotham researches and reviews field guides

Share this guide