A purchased AI assistant can arrive with an interface, model access and product updates. That is useful scope to buy. It still leaves a firm with decisions about eligible work, approved sources, attorney review and what happens when quality changes. I would make that boundary visible before comparing a vendor license with the cost of an internal build.
Draw the boundary around responsibilities
For a fictional chronology workflow, the vendor may operate the core application. The firm defines which matters can use it, how reviewers verify the output and who supports the service internally. Integration and incidents often cross the boundary. Their division depends on the agreement, so the product decision needs explicit responsibilities rather than an assumption that a purchase transfers them all.
Application
Contracted service
Product updates
Access integration
Incident response
Change evaluation
Eligible scope
Attorney review
Adoption & ownership
Follow the time beyond the first draft
Consider 1,000 attempted work items in an illustrative month. The existing process takes 30 minutes each, or 500 hours. Assisted preparation takes eight minutes per attempt: about 133 hours. Human review takes another six minutes on every attempt: 100 hours. If 20% of outputs are rejected and require a complete manual redo, that adds another 100 hours.
The assisted workflow therefore consumes about 333 hours and releases about 167. Looking only at preparation would suggest a much larger 367-hour benefit. The missing 200 hours are exactly the work a polished demonstration can make easy to overlook: verification and the cost of starting again.
133 hReview
100 hRedo
100 hReleased
167 h
1,000 attempts · 30-minute baseline · 8-minute preparation · 6-minute review · 20% full redo. Rounded hours; fictional example.
Decide which benefit can actually be realized
Those 167 hours are capacity, not automatically cash. If spending remains unchanged, cash savings can be zero. A fixed-fee team may use the time to serve additional demand, but the demand and contribution need evidence. Under hourly billing, faster work can also displace previously billed revenue. The same released hour cannot both remove a cost and serve a new matter.
License, processing, support, evaluation and implementation costs belong in the calculation. Shared foundations must be assigned consistently rather than charged twice across multiple opportunities. I would show the unpriced assumptions too: ramp-up, additional staffing when review gets slower, and work that has no credible demand to absorb the released capacity.
The final decision has two separate tests. Does the bounded workflow meet its quality and operating requirements? And do the economics justify the next investment? Positive economics cannot resolve a critical omission. Passing a quality check cannot create customer demand. The operating agreement and the funding recommendation should make both conditions explicit.
An original decision framework with authored examples. These are not measured model results or descriptions of a particular firm.