A Better Tuesday: One Small Workflow at a Time
How a Florida firm can turn one recurring irritation into a controlled, measurable process using software it may already own.
A former Big Law corporate equity partner at Steptoe (NYC Managing Partner) and Seyfarth Shaw • Florida-focused • Independent consultant • No software to sell
Last reviewed: September 30, 2026. This page is updated when workflow practices, product capabilities, Florida guidance, or linked resources materially change.
Most firms begin the AI discussion upside down. Someone sees a product demonstration. The partners ask what the tool can do. A committee produces a list of possibilities. Six weeks later, the firm has another license and the same Tuesday afternoon.
Start with the work instead. Find one recurring task that consumes attention, creates delay, invites retyping, or forces a lawyer to reconstruct information the firm already has. Define the beginning and end. Decide where judgment belongs. Then test whether an approved AI feature can assist with one part of it.
The target is not an “AI firm.” The target is a slightly better workday: fewer avoidable handoffs, less clerical reconstruction, faster client service, and more lawyer time for work that requires a lawyer.
The short version
Six rules for a first workflow.
- Choose a recurring bottleneck, not a fashionable tool.
- Keep the first task narrow enough to describe on one page.
- Test with fictional, public, or otherwise approved information.
- Name the person who reviews the result before the pilot begins.
- Measure production, review, corrections, and rework, not demo speed.
- Set stop conditions. A pilot may earn a no.
Start with one workflow worth fixing.
Five workflow pilot cards, a readiness checklist, tool-approval and verification checklists, sample client language, an incident-response protocol, and a 90-day roadmap. Plus the Small Law Firm AI Workflow Brief.
Get the Free Florida AI Starter Playbook →On This Page
Start With the Bad Tuesday, Not the AI
A tool is not a business problem. “We bought Copilot,” “we have Gemini,” or “Lexis includes AI” describes a license. It does not identify work worth changing. The better opening question is: “Which recurring process makes capable people do avoidable work?”
Listen for complaints with verbs: chase, copy, retype, sort, summarize, rename, compare, remind, locate, reformat, and reconstruct. Those verbs often reveal repeatable work. They do not prove AI is the answer. They tell you where to look.
The best first pilot is usually boring. It happens often. The input is recognizable. The output has a familiar shape. A reviewer already knows what good looks like. An error can be caught before it affects a client. Boring is underrated. Boring can be measured.
The six-part first-workflow test
| Test | What it means |
|---|---|
| Frequent | The task occurs weekly or more often. Repetition creates enough examples to test and enough volume to matter. |
| Painful | The current process consumes time, causes delay, creates rework, or produces inconsistent service. |
| Repeatable | The task has recognizable inputs, steps, and outputs. Every matter need not be identical. |
| Reviewable | A lawyer or trained staff member can tell whether the result is usable, incomplete, or wrong. |
| Recoverable | A failed test can be discarded or corrected before it creates material harm. |
| Measurable | The firm can compare old and new time, review effort, corrections, cycle time, or service quality. |
A task that passes five or six tests is a strong candidate. A task that passes two is not a bold pilot. It is a future incident report with better branding.
The Monday status chase passes all six.
Every Monday, a partner emails the team asking for case status. Run it through the test.
- Frequent. It happens weekly across every active matter.
- Painful. Each lawyer loses 20 to 30 minutes compiling an update, and the partner spends more time chasing stragglers.
- Repeatable. The same fields repeat every time: status, next deadline, open question, risk flag.
- Reviewable. The matter owner already knows what the correct status is and can tell a good draft from a wrong one in seconds.
- Recoverable. A wrong first draft gets corrected before it reaches a client or the partner. Nothing external is at stake.
- Measurable. Preparation time and the number of partner follow-up emails are both easy to count before and after.
Six for six. That is why the weekly matter report shows up again later on this page as one of the six workflows worth testing first.
What a Workflow Actually Is
A workflow is a repeatable path from a defined input to a defined result. It identifies who does what, in what order, using which information, with which review, and where the final record belongs. AI may assist with one step. The workflow remains the larger system.
A prompt is not a workflow. Neither is a chat transcript, a saved template, or the sentence “Ask the AI for a summary.” Those may be ingredients. The workflow explains how the work begins, how exceptions are handled, and who owns the result.
The nine parts of a usable workflow
- Problem. State the recurring friction in one sentence.
- Trigger. Identify the event that starts the work.
- Input. Specify the approved source material and where it comes from.
- Steps. Describe the human and software actions in order.
- Output. Define the expected format and where it will be used.
- Review. Name the reviewer and the checks that person performs.
- Exceptions. List the conditions that require escalation or a manual path.
- Record. Decide what is saved, where it belongs, and who can access it.
- Measure. Define what will determine whether the workflow continues.
If the process cannot be explained, it is not ready to be automated.
Automation accelerates the process you give it. A confused process merely becomes confused at machine speed.
Run a 15-Minute Friction Audit
Do not ask the entire firm for an AI wish list. Ask three people what wasted their time last week. A managing partner, an experienced lawyer, and the person who actually keeps the process moving will usually describe three different versions of the same problem. If you want a faster read on where to start, the 2-minute readiness self-check asks ten questions and takes no email.
Questions worth asking
- What did you have to find before you could begin?
- What information did you copy, retype, sort, compare, or summarize?
- Where did the work wait for someone else?
- What did you check because the first version was often incomplete?
- Which client question required reconstructing the matter from several places?
- Which task produced a second round of work because instructions were unclear?
- What would you gladly never do manually again, but still know how to verify?
Write down the answers in plain language. Group similar complaints. Choose one problem with volume, a recognizable output, and a clear reviewer. Do not begin with the most expensive or prestigious work. Begin where the firm can learn safely.
Poor first-pilot candidates
- A task that occurs twice a year and changes every time.
- A process no one owns or can describe consistently.
- A workflow whose first live test would expose sensitive client information.
- A task that turns on legal strategy, privilege, settlement authority, or professional judgment.
- A process with no reliable examples of good work.
- A project selected because a vendor showed it on a stage.
Six Workflows Worth Testing First
These are starting points, not endorsements of a product. A firm should evaluate the exact system, account plan, terms, settings, data, task, and reviewer before using client information. The first test should use fictional, public, or otherwise approved material.
1. Intake issue list
Turn a completed fictional intake form into a structured list of people, dates, documents, stated goals, missing information, and questions for the lawyer. The system does not decide whether to accept the matter, run the conflict check, or give legal advice. Measure preparation time, missed fields, and lawyer corrections.
2. Meeting notes to assigned tasks
Convert approved meeting notes into a draft task list showing owner, due date, dependencies, and unresolved questions. A participant checks the list before anything is assigned. Measure cleanup time, missed commitments, mistaken assignments, and the number of follow-up emails avoided.
3. Client-update first draft
Using an approved matter summary or fictional record, draft a short update covering what happened, what comes next, what the client must do, and when the firm will communicate again. The responsible lawyer verifies every fact, removes strategy that should not be stated, and approves tone. Measure drafting plus review time and client follow-up questions.
4. Email chronology
From an approved set of test emails, create a table of date, sender, recipient, event, promised action, and source reference. The reviewer compares entries with the original messages and checks omissions. Measure assembly time, correction count, and whether the chronology makes later work faster.
5. Document-request checklist
Turn a firm-approved blank form, procedure, or fictional matter into a tailored draft checklist grouped by source, purpose, and status. A lawyer or senior paralegal removes unnecessary requests and adds matter-specific items. Measure preparation time, missing-document follow-up, and client confusion.
6. Weekly matter report
Convert approved status inputs into a standard internal report: activity, next deadline, owner, open question, client contact, billing issue, and risk requiring attention. Matter owners confirm the report. Measure preparation time, stale items, overdue actions, and partner time spent asking for status.
Need more examples? The Small-Firm Workflow Playbook covers intake, document collection, matter management, drafting, client communication, calendar, and billing workflows built around common small-firm software.
Use a One-Page Pilot Canvas
A first pilot should fit on one page. The document is not bureaucracy. It is the shared answer to “What exactly are we testing?” Without it, each user conducts a different experiment and the firm learns nothing.
| Canvas field | Decision to record |
|---|---|
| Problem | What recurring friction are we trying to reduce? |
| Scope | Where does the pilot start and stop? What remains manual? |
| Approved input | What information may be used in this exact tool and account? |
| Expected output | What format should the system produce? |
| Reviewer | Who checks the work, against which source, before what use? |
| Hard stops | Which conditions require rejection, escalation, or the manual path? |
| Test set | Which ordinary, difficult, ambiguous, and deliberately flawed examples will be used? |
| Measures | Which time, quality, rework, adoption, or service measures will decide the result? |
| Owner and date | Who owns the pilot, and when will the firm decide continue, revise, or stop? |
A Four-Stage Testing Path
Do not move from a product demonstration to live client work. Use stages. Each stage should earn the next one.
Stage 1: Establish the baseline
Collect five to ten recent examples produced through the current process. Record total elapsed time, hands-on time, common corrections, waiting points, and the reviewer’s effort. Use the median rather than the best or worst example. You need a fair comparison, not a prosecution exhibit.
Stage 2: Test with fictional or public material
Create representative test inputs that include ordinary, incomplete, ambiguous, and deliberately flawed examples. Test the instructions, output format, omissions, and failure modes. Record every material correction. A single impressive result proves only that one result was impressive.
Stage 3: Run in shadow mode
Shadow mode means the pilot runs beside the existing process, but its output is not used externally. Compare both results. Measure production time, review time, corrections, and anything the new method missed. Use only information approved for the exact system and test.
Stage 4: Put one controlled use into service
After the workflow, information, system, and review are approved, use the process on a limited set of matters or users. Keep the manual path available. Schedule a decision date. Continue only if the workflow produces a net gain without weakening confidentiality, accuracy, supervision, client service, or records.
Stop the pilot when the conditions change.
- The tool, account plan, terms, retention, or settings differ from what the firm approved.
- Client information appears in a system or test that was not approved for it.
- The output bypasses the named reviewer or reaches a client automatically.
- Material errors cannot be detected consistently or require more review than the old process.
- Users expand the task beyond the written scope.
- No one can explain where the final record belongs or who owns the next step.
Design Lawyer Review Before the First Test
“A lawyer will review it” is not a control. It is a hope wearing business casual. State who reviews, what that person checks, which source controls, what requires correction, and what cannot be delegated.
The review should match the consequence of error. A draft internal agenda does not require the same review as a client update, a calculation, a deadline, a legal research memo, or a court filing. The tool may assist with the work. The responsible lawyer still owns professional judgment and the final result.
A practical review checklist
- Source: Did the reviewer compare the result with the actual document, record, rule, authority, or approved model?
- Completeness: What material fact, issue, document, instruction, or exception is missing?
- Accuracy: Are names, dates, amounts, quotations, calculations, citations, and assignments correct?
- Scope: Did the output do only the task the firm approved?
- Judgment: Did the system make or imply a decision that belongs to a lawyer?
- Confidentiality: Was only approved information used in the approved system and account?
- Communication: Is the tone accurate, clear, and suitable for the intended reader?
- Record: Were the final work, material corrections, approval, and source record saved where required?
For the Florida ethics and risk gates that apply around every workflow, use the companion pillar Use AI Without Betting the Firm.
Measure the Whole Process
The easiest way to overstate AI value is to time the generation step and ignore everything around it. A draft that appears in 40 seconds may require 18 minutes of source checking, six minutes of repair, and a second lawyer who no longer trusts the process.
Count current and pilot work from the first required input-gathering step through the reviewed, filed result. For both, include gathering inputs, preparing instructions, running the tool, checking the output, correcting it, handling exceptions, and saving the work. Record one-time setup and training separately; allocate them across expected volume when judging the pilot’s overall return.
Net hands-on time saved per matter = current-process median − pilot-process median, using the same start and finish.
Track quality separately. A faster process that creates more errors is not efficient. It is merely quick.
Five measures that matter
| Measure | Definition | Decision question |
|---|---|---|
| Total hands-on time | Hands-on minutes from the first required input-gathering step through the reviewed, filed result. | Is the process actually faster after review and correction? |
| Cycle time | Elapsed time from trigger to completed handoff. | Does the work move sooner, or merely sit in a new queue? |
| Material corrections | Substantive fixes, omissions, wrong facts, or wrong assignments. | Is quality improving, stable, or deteriorating? |
| Rework and exceptions | Time spent repairing failures or shifting to the manual path. | Are rare cases consuming the claimed savings? |
| Client or user impact | Follow-up questions, delay, satisfaction, adoption, or abandonment. | Did the workflow improve the experience of the people who use it? |
Decide: continue, revise, or stop
Set the decision rule before the pilot. For example: continue if median total hands-on time falls by at least 20 percent, no high-impact error occurs, material corrections remain below an agreed level, and users follow the review step. Revise if the problem is fixable. Stop if the task is a poor fit, review consumes the savings, or risk cannot be controlled.
A stopped pilot is not a failed program. It is a cheap answer. Buying more licenses before testing would have been the expensive answer.
Use What the Firm Already Pays For (After Checking It)
Many firms already have AI features inside email, word processing, meetings, file storage, practice management, legal research, PDF, transcription, and billing software. Start with an inventory of what the firm owns and what lawyers and staff actually use. Those are often different lists.
Examples may include Microsoft 365, Google Workspace, legal research platforms, practice-management systems, and general-purpose assistants. They are examples, not endorsements. The useful question is not whether a brand offers AI. It is whether the exact product, plan, configuration, data protections, permissions, and workflow fit the firm’s use.
Seven questions before using an existing feature
- Is the feature included in the firm’s licensed plan, and who has access?
- Is it enabled automatically or controlled by an administrator?
- What information can it reach, and what information may users add?
- What do the current terms say about retention, access, training, deletion, and ownership?
- Can permissions, sharing, logs, or connectors expose more information than the task requires?
- Can the firm limit the feature to approved users, data, and workflows?
- Who will monitor changes to the feature, terms, settings, or linked services?
No new software is a useful preference, not a commandment. If the current stack cannot perform the task safely or reliably, do not force it. The right answer may be a process fix, a different feature, a specialized product, or no automation at all.
Florida Guardrails Still Apply
Florida Bar Ethics Opinion 24-1 permits lawyers to use generative AI while keeping existing professional duties in place. For a workflow pilot, the practical pressure points are confidentiality, understanding the tool well enough to use it competently, supervising lawyers and staff, verifying output, charging proper fees and costs, and communicating accurately.
A pilot does not suspend those duties. Fictional or public test material helps the firm learn without exposing client information. Before client work enters the process, the firm should approve the exact system and use, analyze confidentiality and any need for disclosure or informed consent, train the users, and define source-based review.
Efficiency also affects billing. If a workflow reduces time on an hourly matter, the bill should reflect time actually spent. Measure saved time for the firm, but do not resurrect it on the invoice. Ghost hours are still ghosts.
Primary Florida guidance: Florida Bar Ethics Opinion 24-1. The Florida Bar Guide to Getting Started with AI also recommends beginning with administrative or other low-risk tasks, protecting client information, and reviewing AI output.
One sensible starting point for a controlled first workflow.
Get the Playbook and occasional Small Law Firm AI Workflow Brief emails about useful workflow changes, Florida guardrails, and tool updates for firms under 25 lawyers. Unsubscribe anytime.
- Florida rules and AI risk controls
- AI-readiness checklist
- Five ready-to-run workflow pilot cards
- AI tool-approval scorecard
- Daily verification checklist
- Sample client disclosure language
- Incident-response protocol
- 90-day roadmap, one-page pilot brief, partner agenda, and primary sources
A 30-Day Plan That Produces a Decision
Week 1: Find and map the friction
- Interview three people who see the process from different angles.
- Choose one task that passes the frequency, pain, repeatability, reviewability, recoverability, and measurement tests.
- Map the current trigger, input, steps, handoffs, output, review, exceptions, and record.
- Collect five to ten recent examples and establish the baseline.
Week 2: Design and test safely
- Complete the one-page pilot canvas.
- Confirm the exact tool, plan, settings, permitted information, and users.
- Build ordinary, difficult, ambiguous, and deliberately flawed test examples.
- Draft the instructions, expected format, review checklist, and hard stops.
Week 3: Run in shadow mode
- Run the new method beside the existing process without external use.
- Record production, review, corrections, rework, exceptions, and user observations.
- Revise the workflow, not merely the prompt, when failures reveal a process problem.
- Train the small group that will perform and review the controlled pilot.
Week 4: Use, measure, and decide
- Put one approved use into service with a limited group or matter set.
- Keep the manual path available and record every exception.
- Compare the full process with the baseline.
- Decide continue, revise, or stop; document the reason and next review date.
The month should end with evidence and a decision. It should not end with a larger committee, a folder of prompts, or twelve people saying they are “still exploring.”
Make the Workflow Easy to Use
Adoption is not a personality test. If the new process requires five tabs, a 900-word prompt, three naming conventions, and a ritual sacrifice to the permissions menu, people will return to the old process. They are not resisting change. They are choosing the path that works.
The minimum adoption package
- One-page instructions written in the order the work occurs.
- One approved starting point inside the software people already open.
- One example of a good input and reviewed output.
- One review checklist matched to the task.
- One manual path for exceptions and system outages.
- One named owner who answers questions and records changes.
- One scheduled review after enough real use to judge the process.
Train with the actual workflow and representative material. Do not spend an hour touring every feature. The team needs to know what starts the process, what it may use, what the output should look like, what must be checked, and when to stop.
Frequently Asked Questions
Do we need to buy new software before building a workflow?
Usually not. Begin by inventorying the features already inside the firm’s email, document, meeting, storage, practice-management, legal-research, PDF, transcription, and billing systems. Evaluate the exact product and plan. Buy something new only when the existing stack cannot perform the task safely, reliably, or simply enough.
What is the difference between a prompt and a workflow?
A prompt gives instructions to an AI system. A workflow defines the full path from trigger and input through steps, review, exceptions, handoff, record, and measurement. A prompt may support one step. It does not own the process.
What is the best first workflow for a small law firm?
Choose a frequent, painful, repeatable, reviewable, recoverable, and measurable task. Intake issue lists, meeting notes to tasks, client-update drafts, email chronologies, document-request checklists, and weekly matter reports are often easier to test than legal strategy or final advice.
May the pilot use client information?
Begin with fictional, public, or otherwise approved information. Before client information enters the process, approve the exact system, account, configuration, data, task, users, review, and record. Analyze Florida confidentiality duties and any need for client disclosure or informed consent.
How many people should participate in the first pilot?
Use the smallest group that represents the work: often one process owner, one user, and one responsible reviewer. A firm-wide launch produces more variation before the workflow is stable and makes failures harder to diagnose.
How long should a first pilot run?
Long enough to test representative examples and normal exceptions, but short enough to force a decision. Four weeks is often sufficient for a frequent task. A lower-volume workflow may require more time or may be a poor first choice.
How should we calculate time saved?
Compare median hands-on time for current and pilot cases over the same boundary: from gathering required inputs through the reviewed, filed result. Include production, review, correction, exceptions, and record handling. Record one-time setup and training separately or allocate them across expected use. Generation speed alone is not the measure.
What if the workflow saves no time?
Look for another valid benefit: fewer omissions, faster cycle time, more consistent client communication, better records, or less interruption. If the process produces no meaningful improvement, or creates added risk, stop. The pilot supplied an answer.
May staff run the workflow?
Staff may perform approved steps with training and supervision. Define permitted tools, information, tasks, escalation points, and records. A lawyer must retain responsibility for professional judgment and review work that requires it.
When should the firm seek outside help?
Get help when no one owns the process, the data path is unclear, partners disagree about risk, the workflow crosses several systems, the team cannot define review, or prior attempts produced licenses without adoption. Outside help should leave the firm with a working process, written instructions, measures, and trained people.
Pick the recurring workflow that steals part of every Tuesday.
We will discuss the current process, the software the firm already owns, the information involved, where lawyer review belongs, and whether AI is actually the right answer.
No sales presentation. No obligation. No parade of product logos.
About Greg Yates
Greg Yates is the founder of Actionable AI and a former Big Law corporate equity partner at Steptoe (NYC Managing Partner) and Seyfarth Shaw. He helps Florida law firms with fewer than 25 lawyers improve recurring work using AI capabilities in software they may already own. He is independent, does not sell software, and will say when AI is not the answer.
Learn more about Greg, browse the Florida AI Guides, or review Actionable AI services.
Primary Authorities and Further Guidance
- Florida Bar Ethics Opinion 24-1: Generative AI, confidentiality, oversight, fees, and advertising.
- Rules Regulating The Florida Bar: Current professional-conduct rules and updates.
- The Florida Bar Guide to Getting Started with AI: Current Florida Bar educational guidance, examples, and cautions; product information changes.
- ABA Formal Opinion 512: Additional guidance on competence, confidentiality, communication, supervision, candor, and fees.
- NIST AI Risk Management Framework: Voluntary framework for governing, mapping, measuring, and managing AI risk.