The AI bill many law firms haven’t priced in

# What changed Most legal AI today is sold on a per-seat licence. Vendors are moving toward consumption pricing — charging for compute or token use. When a tool runs unsupervised across datasets, access (a seat) and compute (an agent consuming tokens) stop being the same product. That shift makes costs variable and potentially much larger.

# Why the bill will grow

# Where law firms are exposed Boards are unusually engaged because AI affects knowledge work directly and client demand is strong. That combination rewards speed over cost discipline. Three market drivers are pushing rapid adoption: client expectations, fear of missing out, and lawyers' recognition of their exposure as knowledge workers. Those drivers incentivise fast deployment and make it easy to overlook variable billing risk.

# A practical rollout strategy Start with the boring, high-volume, low-risk tasks where baselines exist and errors are recoverable. Examples include:

• File opening and conflict checks
• AML and KYC document handling
• Medical records sorting and chronology building
• Disclosure de-duplication and first-pass relevance
• Court form population and bundling/pagination
• Time-narrative cleanup against billing guidelines
• Inbound enquiry triage

These tasks are already done by paid employees, so you can measure current unit cost and compare it to projected token consumption. Consumption per output in these areas tends to be stable and forecastable. Use those learnings to build an internal token curve before applying AI to drafting and contract review, which are partner-visible, higher-value, and harder to verify at scale.

# Financial planning and governance Assume AI costs will become variable. Plan budgets and approvals as you would for any metered utility: forecast volume, set thresholds for review, and assign a governance owner who can explain usage to the board. If a vendor does not yet meter by tokens, plan as if they will. Treat token consumption as the primary risk variable rather than unit price alone.

# The Jevons insight applied to AI Efficiency gains can increase total consumption. As per-token costs fall, it becomes rational to use more tokens. That pattern means cheaper inference can paradoxically raise total spend. Firms should model both price and volume changes together when forecasting AI costs.

# Practical next steps for firms

• Audit current manual processes to identify high-volume, low-risk pilots.
• Baseline current cost per unit for those tasks.
• Run controlled pilots and measure token consumption per output.
• Establish reporting and approval thresholds for consumption spending.
• Delay wide deployment on drafting and contract review until the firm understands its consumption curve.

# Bottom line