# Calculating realistic ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → when hours aren't the metric
A mid-size firm rolls out an AI contract-review tool. The vendor demo promised 70% time savings on first-pass review. Six months later, the managing partner pulls the numbers: associates are indeed flagging clauses faster, but partner review time has gone up, not down, because the tool occasionally misses a nonstandard indemnification clause and everyone now double-checks everything. Net time saved: close to zero. This is the trap of measuring AI ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → (return on investmentreturn on investmentReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète →) in raw hours, and it is the norm, not the exception, across professional services in 2026.
Vendors sell AI contract-review, due-diligence, and research tools on throughput: pages per minute, clauses flagged per hour. These numbers are real, but they measure the tool in isolation, not the tool inside your actual workflow.
Three costs rarely make it into vendor pitch decks:
The honest ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → equation looks like this:
Net value = (Hours saved x billed/loaded cost per hour) − (Oversight hours x reviewer cost) − (Expected error cost x probability) − (Tooling cost) − (Training and change-management cost)
Assume a mid-size law firm evaluating an AI contract-review tool for NDA (non-disclosure agreement) and vendor contract review, a high-volume, lower-complexity document category. All figures below are illustrative estimates for 2026, not vendor-verified numbers, use them to build your own model with your firm's actual rates.
Baseline (no AI):
With AI tool:
Cost per contract with AI: $112.50 + $30 + $15 = $157.50
Nominal saving: $300 − $157.50 = $142.50 per contract (about 47%), well short of the vendor's advertised 70%.
Now add error correction. Assume 1 in 40 contracts has an AI-missed issue caught downstream (by a client or in litigation discovery), costing an estimated $4,000 in rework, partner time, and risk mitigation. Amortized: $4,000 / 40 = $100 per contract.
Realistic net saving: $142.50 − $100 = $42.50 per contract.
Still positive, but a fraction of the headline claim, and this excludes training time, change management, and the intangible cost of a near-miss client relationship. This is the calculation clients and finance committees rarely see, and it's the one that determines whether a tool earns its subscription renewal.
Not all professional services tasks carry equal error cost. Calibrate oversight intensity to stakes, not to a blanket policy.
| Task type | Error cost if AI misses something | Appropriate oversight level |
|---|---|---|
| NDA/vendor contract review | Low-moderate | Spot-check sampling |
| M&A due diligence document review | High (deal risk, client capital) | Full partner review of AI-flagged items |
| Regulatory compliance filings | Very high (fines, license risk) | Full review, possibly dual review |
| Internal research memos | Low | Associate self-check only |
A tool that delivers strong ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → on NDA review may deliver negative ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → on M&A due diligence if forced into the same light-touch oversight model. Firms that apply one oversight standard across all use cases either overspend on low-risk work or underprotect high-risk work.
When evaluating an AI vendor for contract review, research, or drafting, ask for evidence beyond throughput claims:
1. Error rate on your document types, not the vendor's benchmark set. Run a pilot on your own historical contracts with known outcomes.
2. Time-to-detect for the tool's own errors. Does it flag its own uncertainty (confidence scores), or does it present everything with equal confidence?
3. Audit trail quality. Regulators and courts increasingly expect firms to show how AI-assisted work was produced and reviewed. The EU AI Act (entered into force 2024, phased obligations through 2026-2027) classifies certain professional-service AI uses as requiring transparency and human oversight documentation.
4. Total cost of ownership, including seat licenses, integration with document management systems, and training hours, not just the per-document price quoted in sales calls.
A short technical detail worth understanding even for non-technical readers: many contract-review tools use retrieval-augmented generation (RAG), where the AI searches your firm's own clause library or precedent database before generating an answer, rather than relying purely on its general training. RAG-based tools tend to hallucinate less on firm-specific language but are only as good as the underlying document library, garbage in, garbage out still applies.
# Simplified ROI check, adapt inputs to your own rates
baseline_hours = 2.0
baseline_rate = 150 # $/hour, loaded cost, illustrative
ai_review_hours = 0.75
partner_check_cost = 30 # allocated per document, illustrative
tool_cost = 15 # per document, illustrative
error_rate = 1/40
error_cost = 4000
baseline_cost = baseline_hours * baseline_rate
ai_cost = (ai_review_hours * baseline_rate) + partner_check_cost + tool_cost
amortized_error = error_rate * error_cost
net_saving = baseline_cost - ai_cost - amortized_error
print(round(net_saving, 2)) # illustrative output: 42.5Vérification des acquis
1. In the law firm scenario, why did the AI contract-review tool fail to produce net time savings despite associates working faster?
2. What is the core flaw in using vendor-provided metrics like 'pages reviewed per minute' to estimate AI ROI?
3. A firm is deciding whether to adopt an AI tool for due diligence. According to the realistic ROI framework, which factor is most important to quantify before rollout, beyond raw hours saved?
4. Select ALL correct answers about hidden costs that vendor pitch decks typically omit when promoting AI tools for professional services.
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers about why 'hours saved' alone is an unreliable metric for evaluating AI tools in regulated, liability-heavy fields.
Sélectionnez toutes les réponses correctes.
Vanity metrics describe the tool. Real ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → describes the firm's outcome. Watch for these tells:
Consulting and accounting firms face a parallel version of this with AI research and drafting tools: a McKinsey or Deloitte-style research memo generated in minutes still needs partner-level judgment applied before it reaches a client, and that judgment time rarely shrinks proportionally to drafting time. For sector-wide context on where adoption is actually paying off versus stalling, see McKinsey's periodic State of AI survey which tracks realized value versus expectations across functions, useful as a sanity check against vendor promises.
🎬 [VIDEO: "How Law Firms Are Actually Using AI (and Where It Fails)" - youtube.com - search for recent legal-tech panel discussions covering real deployment outcomes and oversight costs, useful for grounding vendor claims against practitioner experience]