Automate the workflow that eats the most staff hours in a week, fails in a way you can count, runs through tools you already pay for, and can be checked by a person before anything irreversible happens. Anything on your list that misses one of those four tests waits its turn. The method is a week of notebook work before any software is opened: watch the repeating tasks happen, time them and log their failures on the same terms you will measure afterwards, score each against four questions, write one sentence saying what done looks like, then choose the tool. Firms that begin with the tool are the firms whose automation ends up sitting unused.
Exhibit 1
The five moves run in a fixed order and the software is chosen last
Watch the work happen instead of asking people what they do
Ask a competent team what we should automate and you get the task everybody dislikes. That answer is sincere and it is usually the wrong one, because dislike and cost are different things. The job three people hate may take a few minutes each. The job nobody complains about may run every time a form arrives, pass through four systems, and have a retype in the middle.
Toyota's production system has a name for that discipline, genchi genbutsu, and its glossary explains it as go and see: go to the place where the work happens, confirm the facts yourself, and find the root cause before you solve anything. Lean practice names the place where work happens the gemba, and a gemba walk means watching the work and asking questions before you decide on an action. The Lean Enterprise Institute puts the purpose plainly: grasp the current situation by looking at the work, not from a report. In a service business the gemba is a shared inbox, the job tracker, the quoting screen, and the desk where two people do the work.
RAND gives the reason to insist on the walk. James Ryseff, Brandon de Bruhl and Sydne J. Newberry reported in 2024, from interviews with 65 experienced data scientists and engineers, that the most common reason an AI project fails is that the business side misunderstands or miscommunicates the problem to be solved (the report). Their interviews ran from August to December 2023 and the results were written for Department of Defense readers, so read the finding as a diagnosis of a pattern. You buy a tool against a description of the work, and somebody wrote that description from memory. Automation aimed downstream of your real constraint only makes the queue in front of it tidier, so finding the bottleneck that is actually holding the business back comes before choosing the first automation. The same week of watching serves both questions.
Write the week down in a fixed set of columns
One notebook, one week, six columns:
- the task, in the words the staff already use for it
- who does it, and how many times it came up in the week
- roughly how long one run takes, from first click to finished
- what went wrong, including the small catch someone made before sending
- which systems the information passes through on its way
- whether anyone checks the output before it leaves the building
Two rules for filling them in. Time the task as it happens, since a week of real timing beats a month of estimates. And log the catches, because the half-second pause where someone notices the wrong address is your failure rate, and it never reaches a report.
Measure the task before you touch it
The UK government's service manual tells teams to establish a baseline of how the service performs now and to judge any later change against that baseline, and it puts the measurement work at the start: work out your measurements as you build the service, not afterwards. For a firm your size that means timing the task as it runs today, on the same measure you will use after the tool goes in.
An after number with no before number is a story. It cannot be checked next quarter, and it drifts upward in the retelling. A 2025 working paper from MIT's Project NANDA reported that most of the organisations it studied saw no measurable profit and loss return from generative AI, and its authors label their own findings preliminary and interview based. Define success for your first automation in advance, in one number you can check later.
Exhibit 2
A workflow measured before and after on the same terms is the only one that can prove a change happened
One panel has a before column filled in on the same measure as its after column, so the gap between them is a finding you can act on and repeat. The other has an after column and a memory of how bad things used to be.
Score the list on four questions
Score every task you wrote down on four things: staff hours per week, the cost of one failure, how many systems the work crosses, and whether a person can approve the output before it becomes real. The first two are counts from your notebook. The last two decide what kind of automation is allowed.
For the first question, use Google's definition. Their site reliability engineers define toil as work that is "manual, repetitive, automatable, tactical, devoid of enduring value, and that scales linearly as the service grows". The same chapter adds that no single task needs all six properties, and that the more of them a task has, the more likely it is worth automating. One of the six tests is judgment: "If human judgment is essential for the task, there's a good chance it's not toil." Google also separates toil from overhead, since meetings, goal paperwork and hiring admin are overhead, tied to running the organisation instead of running the service.
Check the deletion test before the automation test. Peter Drucker wrote in Harvard Business Review in May 1963 that there is surely nothing quite so useless as doing with great efficiency what should not be done at all. The report nobody opens, and the reconciliation kept alive because one client asked for it years ago. Deleting one of those is the cheapest automation available and it needs no vendor.
RAND's second warning sits next to it: some tasks "could have quickly been captured by a few simple if-then rules". A fixed rule written on a card often beats a model, and a model aimed at a problem that simple is a maintenance bill. Their advice on selection is to pick a problem worth staying with, since "leaders should be prepared to commit each product team to solving a specific problem for at least a year." Your version of that commitment is picking a workflow that will still exist next year, which rules out anything built around one client's current habit.
Exhibit 3
A repeating task reaches automate first only when a person can approve the output
What to leave for later
Leave for later anything that sends, changes a record, or spends money with nobody approving it. Human in the loop means a person stays inside the automated process, operating it, supervising it, or making the decision at set points, and IBM's definition states the cost on the same page. Keeping a person in the loop takes time and attention, which is why it belongs on the decisions that matter. Workflows in this group suit a draft-and-approve design, where the machine prepares and a person releases.
Write the approval down. NIST's AI Risk Management Framework, published in January 2023, asks organisations to have written policies defining roles and responsibilities for human-AI configurations and for oversight of AI systems, and mapping a system includes defining, assessing and documenting the processes for human oversight. NIST treats oversight as a range, with configurations running from fully autonomous to fully manual, and says some systems specifically require human oversight while others do not. The practical version for a small firm is to name who signs off on each automated action before the tool goes live.
Anything that touches protected health information waits for the paperwork. If your firm handles protected health information for a clinic, practice or health plan, HHS treats the vendor you hand that information to as a business associate, which includes a third party AI chatbot that works with patient data, and permits the disclosure only when the covered entity has satisfactory assurances in a contract, the business associate agreement, that the information will be safeguarded. HIPAA regulates who may touch protected health information and on what written terms, so the question to ask of any tool is whether patient data reaches it. What HIPAA requires of a practice using ChatGPT is a separate conversation with a deadline attached.
A workflow only you understand waits too. It cannot be specified, and a tool built on an unspecified workflow produces output only that one person can judge. If you are the only reviewer, the automation adds a review queue to your week, and what the research on context switching says about the owner who does everything explains why the interruptions cost more than the minutes.
The FTC told businesses in 2023 that they cannot escape responsibility for an AI product by blaming a third party developer, and cannot claim the technology is a black box they did not know how to test. The Consumer Financial Protection Bureau's position on algorithmic decisions is that the law is the same whichever tool you use, since a creditor using artificial intelligence or machine learning must still give the specific reasons for an adverse action. For a small firm the rule is simple: the tool makes the suggestion, you keep the decision and the liability.
The three workflows that usually win first
An enquiry form that waits for a human to call back. The cost of a missed lead is a lost job, the failure is countable in the CRM, the hours are real, and a person can approve the reply before it goes out.
A document retyped from one system into another. This is the purest toil on any service-firm list: manual, repetitive, tactical, and it scales linearly as you add clients. It crosses two systems, and it happens every time a job moves from quote to invoice.
The same customer question answered by hand every week. It is short, high frequency, low judgment, and the answer already exists in writing somewhere, which makes it the clearest of the everyday tasks AI can take over. The draft needs a person to read it, and the review belongs on the answers that carry a price.
Exhibit 4
A parallel run on real work is what turns a demo into a workflow you can trust
Run it beside the old way until the staff abandon the old way
Run the new version on real work while the old version still runs, and compare the two outputs on the measure in your baseline. A parallel run catches the client whose address format breaks the import and the question phrased in the one way your prompt cannot handle. Keep the old path open until the outputs agree on that measure, then close it and say out loud that you closed it.
You will know the pick was right when the staff stop doing the old version without being told. A tool people forget to open is a bad pick, and no amount of training fixes it, because the workflow it replaced was not the expensive one. Check the measure again a quarter later, on the same terms as the baseline, and let that number decide the next pick.
Score the list before you build anything
- Did you watch this task happen, this week, at the desk where it happens?
- Do you have a before number on the measure you will use after?
- Would the work exist if you deleted it, and who would notice?
- How many systems does the information cross, and do you control all of them?
- Can one named person approve the output in seconds, and is that written down?
- Can you say what done looks like in one sentence, with a number in it?
A workflow that passes all six is the one to build first, and everything else keeps its place in the notebook. Our AI enablement work covers the part of this from the scoring conversation through the parallel run and the sign-off list.