Most firms that fail at AI do not fail at the build. They fail at the first decision, the one made weeks before a single line of code gets written: which workflow, exactly, are we going to automate first? Pick the wrong workflow and it will not matter how well the module is built. Pick the right workflow and even a middling build will change how the firm feels by the end of the quarter.
After a decade of shipping software inside professional-services firms, I have five questions I walk every managing partner through before we scope anything. If a candidate workflow clears all five, it is almost always a winner. If it fails two or more, we keep looking. There is no prize for being first to build the wrong thing.
1. Does the same person do it more than once a week?
This is the frequency test, and it is the single most reliable filter I have. High-frequency workflows compound value fast. Low-frequency workflows, no matter how painful, produce a tool that nobody remembers how to use by the time it is needed again.
A paralegal who turns depositions into outlines every week is a candidate. A once-a-quarter compliance filing is not, even though the filing is more painful in the moment. Build the weekly workflow. The quarterly one can wait until the weekly one has paid for itself ten times over.
2. Does the workflow have a clear input and a clear output?
The workflows that automate cleanly have crisp boundaries. Something arrives. Something leaves. The gap between the two is where the work happens, and that gap is what a module can compress.
Intake forms become drafted matters. Trust-account statements become reconciled ledgers. Deposition transcripts become issue-tagged outlines. Client emails become drafted replies routed through a partner queue. Every one of those is a good first candidate because the shape of the work is legible to both the partner and the machine.
Compare that to “improve associate training” or “make our knowledge management better.” Those are real problems. They are not workflows. A first AI build needs edges.
3. Is there a person inside the firm who actually knows how it works?
Not the person in charge of the workflow. The person who does it. The paralegal who has run intake for seven years. The senior associate who knows which clauses the partner will actually sign off on. The bookkeeper who reconciles the trust account at the end of every month.
If that person exists, is named, and can give the builder eight to ten hours over two weeks, the module will reflect how your firm actually works. If that person does not exist, or is unwilling to sit with the builder, the module will reflect something generic, and the firm will rightly reject it.
This is the failure mode I see most often at firms that have been burned by consultancies. The consultancy scheduled twelve-partner requirements meetings and nobody ever talked to the person who actually does the work. Do not repeat that mistake with your AI build.
4. Can you measure the improvement in a single number?
Pick a number before we start. Hours per matter. Turnaround time in days. Review cycles per engagement. First-pass acceptance rate. Any one number that the person running the workflow would recognize as a real measure of how well their day is going.
Two things happen when you name the number up front. First, you protect yourself against scope creep, because every suggested feature now has to answer the question of whether it moves the number. Second, you give the firm a real story to tell itself six months later. “Workpaper turnaround went from nine days to three” is a different sentence than “we invested in AI this year.”
5. Would the reclaimed hours actually get used?
This is the question partners skip, and it is the one that determines whether the module pays back six times or sixty.
If your senior associates are billing 1,400 hours a year and you free up 200 of them, do those hours turn into more client work, more business development, more time training juniors? Or do they disappear into the overhead of the day? The answer depends on the firm, not the software. Firms that have already decided what they would do with reclaimed capacity get a compounding return. Firms that have not, get a nicer tool and a flat P&L.
I am not asking partners to run a full business case. I am asking them to finish one sentence out loud before we build: “If this workflow was eighty percent faster, we would use the time to ____.” If the sentence finishes cleanly, we have a real engagement. If it does not, we keep looking.
What to do with the checklist
Put it in front of your partner group. Make a list of five to eight candidate workflows. Score each one against the five questions. Most firms find that one workflow obviously beats the others, and that the top candidate is almost never the one the partners first proposed.
That is a good sign. It means the checklist is doing its job. The workflow that wins is usually the one that a senior non-partner inside the firm has been quietly complaining about for years, sitting one rung below the workflows the leadership team spends its Monday meetings on. Those are the ones that ship cleanly and pay back fast.
The rest is execution, and execution on a narrow, well-chosen workflow is a solved problem in 2026. What is not solved, and is not going to be solved by your vendor’s sales team, is the work of picking the right first workflow. That part is still yours.
If you want a second set of eyes on the shortlist, that is exactly what the thirty-minute bottleneck audit is for. Bring your candidates. I will tell you which one I would build first and why, and if none of them clear the checklist, I will tell you that too.
