The AI Adoption Playbook

Adoption Package

What a firm needs before it relies on AI for a process: a map of the process, a policy for using AI in it, a tested proof of concept, and a recommendation. Each answers a question the partner will ask, and together they let the partner check the conclusion instead of taking it on trust.

Process Map

A step-by-step account of a legal process, naming for each step what is done and the rule or form that governs it. Each step gets a call: automate, where the machine does the work and a person checks it; augment, where it helps while a person decides; or human, where a person keeps the task. A firm cannot decide what to hand a machine until it knows what the work is.

Representation by Signer

Under Florida Rule of General Practice and Judicial Administration 2.515(d)(2), each signer represents that the signer has read the filing, that good grounds support it, and that it is not interposed for delay. Since June 15, 2026, the signer also represents that the legal authorities identified exist and are accurately cited. A court may sanction a filing inconsistent with that representation, after notice and a chance to be heard, whether a lawyer or a self-represented party signed it.

AI Use Policy

A firm’s written rules for using AI on client work: which tools staff may use, what client information may go into them, who reviews their output, and how the time is billed. ABA Formal Opinion 512 says a firm’s managing lawyers must establish clear policies on its use of generative AI. Florida Bar Ethics Opinion 24-1 says lawyers must develop policies and practices to verify that their use of it is consistent with their ethical obligations.

Pre-Signature Verification Checklist

The steps a person completes and initials before anyone signs a filing that AI helped prepare. They include opening every cited authority, confirming that it exists and says what the filing claims, and checking each quotation, pinpoint page, and fact against its source. The checklist is what lets a signer make the representation by signer knowing it is true.

Skill Document

A written method an AI tool follows for one task: its purpose, the inputs it needs, the steps in order, the form of its output, and where it must stop for a person. The Commercial Lease Auditor from Activity 5 is one. A firm can review, test, and revise a skill document like any other work product, which it cannot do with a prompt someone typed once and forgot.

Proof of Concept

A small test of whether a tool can do one step of a process well enough to be worth using, run on fictional test cases whose correct outputs are written down before the first run. Its evaluation log records each run: the input, the output, what was right, what was wrong, and how many minutes the checking took. A tool that saves ten minutes of drafting and costs twenty of checking has saved nothing, and the log is where that shows.

Risk Table

A short table of what could go wrong with a workflow, how likely and how serious each risk is, and the control in the policy or skill document that answers it. A row with no control is the most useful one in it, because it tells the partner what the firm would still have to build.

Part 1: Choose the Process and Map It

Read the assignment below. Then read the Florida Small Claims Rules that carry a case from filing to judgment, Rules 7.010, 7.050, 7.060, 7.070, 7.090, and 7.170, on The Florida Bar’s rules page. Skim the Volusia County Clerk’s small claims page and forms list, and the state courts’ account of the court process.

The assignment: You are an intern at Thornton & Associates, P.A., a small litigation firm in Daytona Beach. Senior partner Catherine Reyes has been reading about AI in legal practice and wants to know what is real. She says: “Pick one of our bread-and-butter processes and give me a proof of concept. Show me the whole process, tell me where AI helps and where it absolutely does not, build me something I can test myself, and tell me honestly whether it is worth more of our time.”

The process: Small claims collections in Volusia County Court, which the firm files dozens of times a year. The clients are businesses owed money by other businesses, on an unpaid contract, unpaid services, or a promissory note, for $8,000 or less. The process runs from the client’s first call to a final judgment. With your instructor’s approval you may take the Seaside Imports subpoena below or the review of a Florida residential lease instead.

The alternate, Seaside Imports: The firm represents the plaintiff in Seaside Imports, LLC v. Coastal Logistics, Inc., Circuit Court of the Seventh Judicial Circuit, in and for Volusia County, Florida, Case Number 2025-CA-004892, Division 37. Seaside alleges that Coastal lost or mishandled a shipment of commercial goods valued at $127,000, stored at Coastal’s Daytona Beach warehouse between March 1 and March 31, 2025. Plaintiff’s counsel is Rachel M. Thornton, Esq., Thornton & Associates, P.A., 401 South Ridgewood Avenue, Suite 300, Daytona Beach, FL 32114. Defendant’s counsel is Marcus J. Webb, Esq., Webb Defense Group, 128 East International Speedway Boulevard, Daytona Beach, FL 32118. The process is a subpoena duces tecum for a deposition under Florida Rule of Civil Procedure 1.410, from the decision to take the deposition through service.

The first test case: The continued deposition of Maria Elena Santos, Warehouse Operations Manager, Coastal Logistics, Inc., 450 Mason Avenue, Daytona Beach, FL 32117, at Thornton & Associates, on a date you set. The documents to request are all documents relating to the handling of Plaintiff’s shipment, including receiving logs, inventory records, storage location assignments, temperature and condition monitoring records, transfer or movement logs, shipping and delivery manifests, chain of custody documentation, incident or exception reports, and internal communications concerning the shipment or the Plaintiff.

Write the map in 400 to 500 words, from the client’s first call to the final judgment. Give each step a sentence on what is done and who does it, the rule, statute, or clerk’s form that governs it, and your call: automate, augment, or human, with a one-sentence reason. A table works if each row carries all three. Then score the map. Each step earns a point for its citation and a point for a call with a reason. It loses the citation point if the rule, once you open it, does not say what your step says. Run the Process Mapper in a fresh chat, pasting your map and the text of every rule you cited. It checks and asks questions, and it does not rewrite. Change a step only where a rule supports the change, and keep your score from before and after.

Part 2: Write the Policy and Its Checklist

Read Florida Bar Ethics Opinion 24-1 in full. In ABA Formal Opinion 512, read the sections on confidentiality (pages 6 to 8), supervision (pages 10 and 11), and fees (pages 11 to 14). Then write Thornton & Associates a one-page AI use policy for the process you mapped. It answers five duties, each tied to its rule in the Rules Regulating The Florida Bar. For competence (Rule 4-1.1), it says which tools staff may use and what training comes first. For confidentiality (Rule 4-1.6), it says what client information may go into which tool and when the client’s consent comes first. For supervision (Rules 4-5.1 and 4-5.3), it says who reviews AI-assisted work before it leaves the firm. For fees (Rule 4-1.5), it says how AI-assisted time is billed. For candor to the court (Rule 4-3.3 and Rule 2.515(d)(2)), it ends with a checklist.

That checklist is the firm’s pre-signature verification checklist: the steps a person completes and initials on every filing AI helped prepare, before anyone signs it. Write it so that a signer who follows it can truthfully make each representation in Rule 2.515(d)(2). Then score the policy. Mark each duty met, partly met, or missing, and write beside it the passage of Opinion 24-1 or the page of Opinion 512 that your line answers. Run the Policy Checker, pasting your policy and both opinions, and compare its ratings with yours. Where the two disagree, reread the opinion and decide for yourself, then revise the policy once.

Part 3: Build and Test the Proof of Concept

Pick one step you called automate or augment, such as drafting the statement of claim from the client’s invoice, calendaring the case’s dates, or assembling the papers for a default. Write a skill document for that step, taking the Commercial Lease Auditor from Activity 5 as your model for its parts, not its subject. It states the step’s purpose, the inputs it needs, the steps in order, the form of its output, what it flags for a person, and where it stops and waits for one. Score it a point for each of those six parts it has.

Before the first run, write three fictional test cases and, for each one, the output a correct run would produce. Make one of them hard. One kind of hard case is a claim for $7,600 on an unpaid invoice, with $900 of contract interest and a demand for attorneys’ fees. It still belongs in small claims, because Rule 7.010(b) leaves costs, interest, and attorneys’ fees out of the $8,000 limit. Use only made-up facts, so nothing from a real client goes into any tool. Then run the Proof-of-Concept Reviewer on your skill document and your tests, and make the changes you agree with.

Run the skill in any AI tool that keeps standing instructions or reads an uploaded document, or paste it at the top of a fresh chat, once for each test case. Keep an evaluation log with a row for each run: the test case, what came back, what was right, what was wrong, and the minutes you spent checking it against the rules and forms. Estimate, too, how long the step takes by hand. Then revise the skill document once, for the failure that matters most, and run the case it failed again.

Part 4: Recommend, and Name Five Risks

Write Catherine Reyes a recommendation of about 200 words: go ahead, go ahead only within limits you name, or stop. Rest it on your evaluation log: how many runs were right, how long checking took against doing the step by hand, and what the revision fixed. Under it, add a risk table with five rows. Each row names one risk, rates its likelihood and its severity as low, medium, or high, and names the control that answers it by the line of your policy or skill document where it lives. A risk your package does not answer gets a row that says so. A recommendation not to proceed, well supported, is a good answer.

Score the table a point for each row whose control you can find in your package, and check every number in the recommendation against the log. Then run the Risk Table Reviewer, and decide which of its points to act on.

Part 5: Submit the Package

Put the package in one PDF, in this order: the recommendation and its risk table, the process map, the policy and its checklist, and an appendix. The appendix holds the skill document as first written and as revised, the three test cases with their expected outputs, the outputs, and the evaluation log. Under each part, give your score and one sentence on a change a prompt led you to make, or one you declined, and why. Post the PDF as a new thread in the Project discussion forum, with your process named in the subject line. This submission is worth 100 points. To earn the remaining 20 points, reply to four different classmates. In each reply, point to one call in their playbook, a step, a policy line, or a risk, that you would have made the same way or another way, and say why.

Part 6: Share Chat Link

Include one AI chat link with a 1–2 sentence explanation of what the conversation shows and why you chose to share it.

Process Mapper

Checking your steps against rule text you pasted is source grounding: the model compares two texts instead of reciting small claims practice from memory. Because it asks about your calls instead of making them, the judgment stays with you.

Policy Checker

A table with one row per duty is structured output, and it keeps the model from praising the policy as a whole while one duty goes missing. Because every rating must quote the opinion, you can check the model’s judgment against the source before you change a word.

Proof-of-Concept Reviewer

One question at a time makes you find the gaps yourself, which is Socratic elicitation, and a skill document you have had to defend is one you understand. The question about inputs no test tries is how testers find the case a tool was never built for.

Risk Table Reviewer

Telling the model to attack your table instead of approving it is red-teaming, and it works against a model’s pull toward agreeing with you. The last question holds your recommendation to your own evidence, the same test the partner will apply.