Phase 1 Optimize: A Practical Way to Pick Your First AI Project (“The Staircase Methodology”)
This is the how-to article in the series. The first piece explained why small and mid-sized businesses keep getting stuck on AI; the second showed the Staircase Methodology running end to end on a home-inspection business. This article lays out Phase 1 Optimize, step by step, with the case study numbers used as a worked example along the way.
Optimize is about using AI to save time and cost on work you already do. It is usually the right place to start: optimization wins are faster, cheaper, and safer than growth bets, and they teach you what AI is good at.
Before getting into the steps, one practical note. The methodology comes with two printable worksheets that structure this entire sequence for you: one covers the goal, brain dump, and scoring (Steps 1 to 4), and the other covers the four checks and the written commitment (Steps 5 to 6). You do not need anything more than two pages to run Phase 1. Both worksheets are included at the end of this article, and the full versions live in the methodology paper on SSRN. It might help to glance at them now, then follow along as each step is explained below.
Should You Start With Phase 1 at All?
For most SMBs, yes. But it might make sense to skip ahead to Phase 2, Grow, if any of these clearly apply: you have already shipped two or three optimization projects and know the tooling; your business is pre-product-market-fit or in revenue decline, so optimizing a process you might shut down is wasted motion; your realistic optimization wins are small while your growth wins are far larger; or you are an operator with strong technical fluency looking for leverage rather than learning. If none of these strongly fit, Phase 1 is the place to begin.
Two principles sit underneath every step: strategy first, technology second (start with what the business needs, not with what AI can do), and focus on what matters most (you cannot bolt AI onto an already-full plate, so every project means something else has to give).
Steps 1 to 4: Find the Right Candidates
Step 1: Set a Goal
Pick one specific, dated goal and finish the sentence. The methodology offers three starters, depending on where the pain is. For example: “In 3 months, I want to spend ___ fewer hours per week on ___”.
The blanks matter. A vague goal like “more money” or “grow the business” filters nothing. Specificity is what makes the rest of the exercise useful. In the case study, the owner chose “In 3 months, I want to spend 30 fewer minutes per client,” which is concrete enough to test every later decision against.
Step 2: Brain Dump
On a single page, list every repetitive process, recurring task, and time-sink in the business. Do not filter yet; aim for 10 to 20 items. Good candidates happen more than once a week, are things you or your team complain about, or simply feel like “robot work.”
Step 3: Connect the List to the Goal
Now look at the list and circle every item that would directly help reach the goal. Cross out the rest. You can also cross out things you do not want to automate, like personal client interactions you value, or work that is obviously not an AI’s job. The rigorous feasibility check comes later, so here you just trust your gut on the obvious cases.
This is where the goal earns its keep. If many items get crossed out, that is fine; you have narrowed your focus. If nothing connects to the goal, that is a real signal to rewrite the goal or dig deeper into the brain dump.
Step 4: Score With Bleed
Score each surviving item with a simple formula: Bleed = Pain × Frequency. Rate each on a 1-to-5 scale. For Pain, 1 is mildly annoying, 3 costs real money or time, and 5 is existential or cannot scale. For Frequency, 1 is monthly, 3 is weekly, and 5 is daily. Multiply, and the highest scores rise to the top.
A note on the math: Bleed is a ranking aid, not a measurement. The number from 1 to 25 is not a unit of anything real; it is just a fast, honest way to sort a long list, and you commit to actual numbers later, only for the top items. The logic is that pain alone is not enough to justify automation. A once-a-year headache is not worth the effort, but frequency turns pain into a wound that will not heal, and those are the ones worth fixing.
In the case study, this produced report creation 20, inspections 15, scheduling 8, and invoicing 6, which matched the owner’s instinct that reports were the real drain. By the end of Step 4, you should have your top three to five candidates.
Step 5: Run the Top Candidates Through Four Checks
Each top candidate now has to clear four checks, in order. They are often answered in a single conversation, with an expert, a knowledgeable contact, or AI itself, but they are logically separate and each needs an explicit answer.
Check 1: Can AI Actually Do This?
Not every problem is an AI problem. AI handles some things well: text drafting and generation, summarizing, pulling structured information out of documents and images, data analysis, pattern recognition, FAQ responses, basic forecasting, translation, etc. It handles other things poorly or not at all: physical tasks, regulated decisions without human review, deep relationship judgment, and anything requiring true accountability or context it does not have.
Sometimes the honest answer is “partially”: AI drafts, a human approves and ships. That still counts as a yes, with a checkpoint built in. If the answer is no, cross the item out. And keep in mind that capability moves fast; something AI could not do well last month may be solved today.
Check 2: How Much Effort?
Estimate effort for each remaining item using four brackets. Small is roughly 1 to 3 days, using an existing AI tool more or less as-is. Medium is about a week, with light setup and easy integrations. Large is about a month, a custom build with more complex integrations. Extra-large is more than a month, which it might be a good idea to skip on a first attempt unless you are completely sure.
Do not know how to estimate? That is normal, and the skills gap is the most cited barrier for SMBs. Three options usually work: ask someone in your network who has done similar work, ask AI itself to scope the task and its likely complexity and failure modes, or hire help. Even a rough answer beats not starting.
Then filter by how many AI projects you have shipped. If you have shipped none, narrow down to Small items only, and prefer the one with the highest Bleed; the goal is not the biggest opportunity but to ship one thing and learn. Medium and Large become accessible once you have one Small project behind you. If you have shipped one or more, or you have help, then all three sizes are fair game, but bigger projects need a higher bar of pain: keep a Small item if its Bleed is at least 6, a Medium if at least 12, a Large if at least 20.
Among items that pass, the best choice is the one that beats its threshold by the widest margin. If fewer than two candidates remain, loop back to Step 4 and pull in items further down the list, so that if your top pick later fails, you have a real alternative. In the case study, reports came in at Medium with a Bleed of 20, clearing the threshold of 12 by the widest margin, which is why it became the lead candidate with scheduling held as backup.
Check 3: Return on Investment, as Payback
Express ROI as a payback period: Payback in months = Implementation Cost ÷ (Hours Saved per Month × Hourly Cost). Under 6 months is good to go. Between 6 and 12 months is still a yes, just plan for it. Over 12 months means picking a different item.
Do not know your hourly cost? A potential starting point is $50 an hour for solo operators and $100 for skilled team members; use whatever fits and refine later. Payback is just an easier way to express ROI than the classic formula, and easier to compare across projects and defend to a partner or lender. In the case study, a custom reporting cost about $5,000 to build, saved about 20 hours a month at an assumed $75 an hour, and paid back in about 3.3 months.
Check 4: Reversibility and Risk
This is the check that most often prevents real damage. Score Risk = Impact × Stickiness, each on a 1-to-5 scale where 1 is safest. Impact asks: if this produces a bad result, how bad is the consequence? Stickiness asks how stuck you are with that result, and it has two dimensions. Technical Stickiness is how easily you can turn the tool off or restore the old workflow. Reputational or legal Stickiness is how easily you can take back the output once it has reached customers, regulators, or the public.
Stickiness is the maximum of those two, not the average, and that choice is deliberate: you are only as unstuck as your worst dimension. A tool that is trivial to switch off but whose output has already reached customers is still sticky, because the damage is done. Thresholds: below 4 is low, proceed; 4 to 9 is medium, build in a rollback plan; 10 or above means STOP, add a human checkpoint to lower the risk and re-score, or pick another project.
This is exactly where the case study’s report project was almost crossed out. Impact was 4, Technical Stickiness 1, but Reputational Stickiness 4, because a sent report cannot be unsent. Risk came to 16, above the STOP line. Adding one checkpoint, the inspector reading every AI-generated section before it goes out, dropped Reputational Stickiness to 2 and Risk to 8, which cleared the project on safer terms.
Step 6: Commit With a Metric, Then Ship
Before building, commit three things in writing. First, one specific, measurable success metric tied to your Step 1 goal, like “reduce the time to draft a client report from 60 minutes to 15 within 4 weeks.” Vague metrics like “save time” lead to projects that quietly run forever. Second, a check-in date, 2 to 8 weeks out, when you evaluate against that metric and decide to expand, adjust, or kill the project. A project without a kill condition runs forever whether or not it works. Third, an honest answer to what this project replaces: you are already at full capacity, so name what you will stop, postpone, or delegate to make room. If the answer is “I will just squeeze it in,” the project will quietly die.
Then ship it. Do one project, measure it against your metric, and only then come back for the next one. If you have a team, be transparent about what is changing and where the freed-up time is going, and start with your most curious team member rather than the whole group at once.
What Comes Next
The final article in the series covers Phase 2 Grow: the same staircase, but aimed at capturing revenue rather than saving time. Most of the steps are identical, with a different scoring formula and a sharper focus on risk, since growth projects are usually customer-facing by definition.
The full “The Staircase Methodology for AI Adoption in US Small and Mid-Sized Businesses”, with complete references, is on SSRN for the deeper version. This article is the practical, faster-to-apply companion to it.
Next article “Phase 2 Grow: Using AI to Capture Revenue, Not Just Save Time (“The Staircase Methodology”)”.
Previous (second) article “The Staircase Methodology at Work: How One Small Business Picked Its First AI Project (Case Study)”.
First article “Why Most Small Businesses Get Stuck on AI, and How the New Staircase Methodology Helps”.
Note: the views expressed in this article are my own and do not represent the official positions of any past, present, or future employers, clients or stakeholders.
