Business Strategy: Lean, AI and Unit Economics Gates

A scheduling tool for field service companies grows revenue 8% a month for a year. The company is made up, and its numbers run through this whole piece. On average, the company earns back what it spent to win a customer (its CAC) in just under 12 months, which looks healthy. Underneath, two in five new customers demand a custom workflow, and those accounts take almost two years to pay back. The strategy deck never shows this. A deck describes where the company is heading, not the conditions under which it should keep spending to get there.

Treat strategy instead as a set of tools that tell you what to test, what to measure, what to stop and what to scale. Lean methods turn beliefs into checked facts. AI can turn messy signals into faster decisions and sometimes cheaper operations, but it brings its own costs. Unit economics, the money made or lost on each customer, sets hard limits that stop you from scaling a broken model. And growth hacking, done well, is building mechanisms that grow on their own. It is closer to product work than to campaign planning.

Strategy needs artifacts, gates and rituals, in that order

The instruments come in three kinds. Artifacts are what strategy is written down as. Gates are the conditions that decide whether you scale. Rituals are the recurring reviews that keep both from going stale.

Each depends on the one before it. Artifacts with no gates produce motion without a decision: the scheduler can have a detailed growth plan and still nothing that forces the question of whether custom workflows are worth what they cost. Gates with no rituals fail more slowly. Nobody checks them the week things get busy, and the miss shows up a quarter later in the cash balance.

The order matters for a practical reason. A gate needs a number to test, and the number lives in an artifact. A ritual needs a gate to apply, or it turns into a status meeting where everyone reports progress and nobody decides anything.

Six artifacts replace the strategy deck

Most strategy work fails because it produces documents that do not run the business. Replace the deck with six artifacts a team can use weekly, each short enough to read in five minutes.

Start with the Outcome Contract: a one-paragraph promise in measurable terms. For a specific customer group in a specific situation, we deliver a measurable result, so that a costly problem gets smaller, without a side effect customers fear. In B2B operations it might read: "For mid-market distributors managing 10k+ SKUs, we reduce stockout-driven lost sales by improving reorder decisions, without increasing overstocks beyond a defined threshold." The sentence names no feature, and that is deliberate.

Behind the promise sits the Assumption Register: a ranked list of what must be true, 10-20 items at most. Typical items:

Will users actually switch from what they use now?
What data and integrations do we need?
Will customers pay?
Does the value repeat, or is it a one-off?
Can we win customers through our channels at a sane cost?
What will support, compliance and computing cost per customer?

Give each item two scores: how fatal it is if wrong, and how little evidence you have. Multiply them. The top three are your work for next quarter.

For the scheduler, scored on 1-5 scales, the register might put "mid-market firms will accept a templated workflow" at fatality 5 and evidence 4 (almost none), for a product of 20. "Small crews will pay $120 a month" scores 4 on fatality but 1 on missing evidence, because hundreds already do, for a product of 4. The first assumption goes into the Proof Plan; the second does not deserve a test this quarter.

The Proof Plan takes those top three and sets, for each, the smallest credible test, the evidence you will accept, the decision rule (scale, adjust or stop) and the date you decide. "Review in Q3" names a season, and the plan needs a calendar date.

The Unit Model Sheet is one page of unit economics for each customer group. It shows the full cost to win a customer (CAC), what each customer leaves after the costs they cause (contribution margin), how many months it takes to earn back CAC (payback), whether the retention curve levels off or keeps falling, and extra revenue from upgrades, add-ons or usage. Without this page, growth is just motion.

A Loop Diagram shows on one page how growth feeds itself. In an integration loop, more connectors make the product easier to adopt, which brings more customers, who ask for more connectors. In a template loop, reusable setups make onboarding faster and bring more users, who create more templates. A strategy without a loop has to be bought again every quarter.

Last comes the Stop List. It names the customer groups you will not chase, the features you will not build, the channels you will not scale and the deals you will refuse because the numbers do not work. For the scheduler, one line would have said: "no custom workflow unless the client pays for implementation above a set threshold." Without this list, every attractive opportunity reopens the debate the other five documents were meant to settle.

Four gates must hold before acquisition spend rises

Strategy becomes credible when scaling is conditional on evidence. Each gate below has to hold before more money goes into acquisition, and they are checked in order, because a later gate cannot rescue an earlier failure.

Gate one: value is reached quickly, and "quickly" is defined per market. Minutes or hours for a self-serve product, days for a sales-led workflow, and even in a regulated enterprise a first trusted win within the first month. Long time to value raises CAC through longer sales cycles and raises churn through buyers who give up before the payoff.

Gate two: value repeats without heroic effort. If users benefit only during onboarding, audits or migrations, you have an event-based tool, and retention will decay to near zero between events. The test: do users return without reminders, does weekly usage flatten after the first 6-8 weeks instead of sliding, and can you describe the habit loop in one sentence?

Gate three: unit economics are stable by segment. Blended averages are the most comfortable form of strategic self-deception. You need to know which customers are profitable, which churn, which generate support load and which channels attract the good ones.

Gate four: growth improves efficiency over time. Activation rises as onboarding templates mature, CAC falls as referrals and partners contribute, support cost per customer drops as the product hardens, and margins hold as usage grows. If growth needs ever-increasing spend to keep the same pace, you are renting it.

Each gate fails with a typical symptom, and each symptom points to a different first check:

Gate one fails. Few trials turn into paying customers, and sales cycles stretch past 90 days. Check the median number of days from signup to the first completed job.
Gate two fails. The retention curve never levels off, and the "active" count depends on email campaigns. Check what share of weekly active users came back without a reminder.
Gate three fails. Average payback looks healthy while the support team keeps growing. Split the unit sheet by customer group before anything else.
Gate four fails. CAC creeps up 5-10% a quarter while spend grows. Check whether the share of new customers from referrals and partners is rising or flat.

One safety check applies to all four: how many months of cash you have left, reviewed every month next to the gate metrics.

The scheduler's numbers, segment by segment

Here is gate three failing in arithmetic. Segment A is small crews that use the standard flow: CAC $600, ARPU $120 a month, variable cost $45, so contribution is $75 and payback is 600 / 75 = 8 months. Segment B is mid-market firms that demand custom workflows: CAC $900, ARPU $150, but variable cost $110 because of support and configuration, so contribution is $40 and payback is 900 / 40 = 22.5 months.

Now mix them as the scheduler does: 60% A and 40% B. Average CAC is $720, average contribution is $61 and average payback is about 11.8 months. That passes the 12-month bar many SaaS boards use. It also hides a segment that will not pay back before most startups run out of cash. Worse, B often grows faster, because mid-market buyers come through sales and the sales team is paid on bookings. When B's share rises from 40% to 60%, average payback moves to about 14 months, and the company "suddenly" needs a bridge round.

The fix follows from the numbers:

turn the most common custom requests into templates;
charge for implementation above a set threshold;
price by what drives value (routes, technicians, jobs) instead of a flat fee per seat;
stop selling to firms whose workflow does not fit.

If templates cut B's variable cost from $110 to $70, contribution rises to $80 and payback falls to about 11 months. The strategy can scale because costs become predictable.

Lifetime value tells the same story from the other side. If segment A customers stay about 24 months, LTV is $75 × 24 = $1,800 and LTV:CAC is 3.0. If segment B customers stay longer, say 30 months, because switching away from a custom setup is painful, LTV is $40 × 30 = $1,200 and LTV:CAC is 1.3. The longer lifetime does not rescue the segment; it only stretches the time before the loss becomes visible. That is why payback, measured in months of runway, is the better gate for a company that is not yet profitable, and LTV:CAC the better one for a company that is.

Sketching the first draft of the model (segments, pricing, channels, cost drivers) is faster in a business model canvas builder than in a blank spreadsheet; the discipline is replacing each assumption with a measured number as you test it.

What the unit sheet is built from

A unit sheet is only as honest as the definitions behind it, and most arguments about payback turn out to be arguments about definitions. Before the sheet goes on a dashboard, write each metric down together with the grain at which it is computed.

Fully loaded CAC for a segment and month: paid media plus sales and marketing salaries, tools and agency fees for that month, divided by new paying customers in that segment that month. Leaving salaries out is the usual way a sales-led segment looks cheap. Shared costs need an allocation rule, such as logged sales time per segment, fixed for the year.
Contribution margin per account per month: recognized revenue minus the costs that grow with that account, such as payment fees, hosting or compute, support time and onboarding hours. Support time is the line that exposed segment B, so it has to come from tickets tagged by account; a company-wide average would have hidden it.
Payback: fully loaded CAC divided by the average monthly contribution of the cohort acquired in the same month, with churned accounts kept in the cohort.
Segment: assigned once, at signup, from firmographic rules such as crew size and workflow type, and never relabeled later. If accounts that churn are quietly moved to another segment, the healthy segment looks healthier than it is.

The data behind this fits in three small tables:

Accounts: one row per customer, with signup date, acquisition channel and the segment fixed at signup.
Monthly accounts: one row per account per month, with revenue, each type of variable cost and whether the account was active.
Spend: cost by channel and month, with salaries split by the rule above.

Payback, contribution and the retention curve all come from joining these three by account and month. If a dashboard number cannot be traced back to them, ask about it before it goes into a board pack.

Prove the hardest assumption first

Lean is how you find out whether the gates hold before betting the company on them. It becomes strategic when you treat it as a craft of proving the riskiest thing cheaply. One of the oldest examples dates from 1999, when Zappos founder Nick Swinmurn photographed shoes in local stores, posted them online and bought a pair at retail whenever an order came in. The test answered one question (will people buy shoes without trying them on?) before anyone built a warehouse.

Four kinds of test cover most cases:

Shadow mode. The new system runs next to the current one and you compare results, without changing how work is done.
Concierge delivery. You deliver the result by hand to see whether anyone keeps paying.
Pre-commitment. Signed pilots, letters of intent or approved budgets tied to clear milestones.
Painted door. You measure interest (requests, deposits, qualified demos) before you build.

Risk scoring for business loans shows shadow mode at work. Instead of building an AI underwriting platform, you run the model next to the current rules, step by step:

Score past deals with the model while the current rules stay in force. Nothing in live operations changes yet.
Compare default rates and approval speed with the current rules on the same deals. Remember that rejected applications never show whether they would have defaulted. The question is whether the model is better or just different.
Show loan officers what drives each score. A number they cannot question will not survive a credit committee.
Check whether decisions would have changed and whether the business result improves. If decisions do not change, the automation does not matter.

Return reduction in ecommerce follows the same logic. Before building a personalization engine, run a concierge experiment (better sizing guidance, pre-purchase Q&A, clearer product media) and measure return rate and repeat purchase. Automate only what moves the outcome. If returns fall from 30% to 26% on the concierge cohort, you know which lever the engine should pull; if they stay flat, you saved a quarter of engineering time.

AI pays only when its cost side is on the unit sheet

This kind of proof matters most for the tool people adopt on faith. AI is worth it when it changes results or costs in ways you can measure and keep. Three areas are worth testing:

Faster decisions: predicting churn to target retention offers, spotting fraud, forecasting demand for staffing.
Faster time to value: guided setup, automatic configuration, sensible defaults.
Lower variable costs: sorting and deflecting support tickets, pulling data out of documents, quality checks.

A network of clinics struggling with missed appointments makes this concrete. A model can predict who is likely not to show up, send targeted reminders, suggest safe overbooking levels and assign slots by specialty. The result shows up in how full the schedule is and in margin, which is where the board will look.

Each of these bets has a typical way to fail, and the failure usually sits right next to the win:

Faster decisions pay off when a prediction changes an action. They fail when nobody counted the cost of each prediction, the monitoring and the human review of tricky cases, and these eat the margin.
Faster time to value pays off when guided setup gives a quicker first win. It fails when the AI gives inconsistent answers and users stop trusting it. That shows up later as support load and churn, not in the demo.
Lower variable costs pay off when the AI really handles volume. They fail when regulated decisions need audit trails and explanations the system was never built to give.

Put the scheduler's support desk through that test. Say it handles 4,000 tickets a month at $6 of agent time each, or $24,000. An AI triage layer that deflects 35% of tickets saves 1,400 × $6 = $8,400 a month. If inference, retrieval and a human spot-check cost $0.90 per ticket across all 4,000, that is $3,600, leaving net savings of $4,800. Now suppose deflected customers reopen 20% of tickets, angrier than before. Those 280 reopened tickets cost more agent time than the originals, and the net saving can fall below $3,000. If a proposed AI bet cannot be written into one of the three lines above with its cost side filled in like this, it is not ready for the roadmap.

Fix the growth mechanism before buying channels

Growth hacking becomes strategy when you treat it as designing a system. Find the weakest link first. It might be a hard onboarding, no habit forming, a price that does not match the value, unreliable results, low trust or high variable costs. Then design tests that strengthen that link. Channels come later.

Take a developer platform that relies on paid ads while CAC is high and conversion is low. That is a mechanism problem. The fixes come from the product. Shorten the time to the first success with better docs and quickstarts. Ship starter templates that give a working result in minutes. Add integrations with the tools customers already use. Price by usage tiers so price follows value. When the mechanism works, channels get cheaper. If trial-to-paid conversion rises from 4% to 6%, the same ad budget brings 50% more customers, and CAC falls by a third without any change to bidding.

B2B analytics with many trials and few conversions tells the same story. Instrument where users stall, redesign onboarding around a single "first insight" milestone, add guided data import and sample data so value appears before setup is finished, then follow with habit triggers like weekly summaries and anomaly alerts. You do not market your way out of weak activation.

Measuring a loop takes one metric and one guardrail. For the template loop, the metric is the share of new accounts that start from a template created by another customer; if it rises from 10% to 25% over two quarters while time to first job falls, the loop is turning. The guardrail is support tickets per new account, because templates that half fit generate more questions than a blank setup. For the integration loop, track new customers whose first session includes a connector and the request queue for connectors you do not yet have. A loop that shows no movement after two quarters of investment is a feature, and it belongs in the portfolio review with the other bets that failed their gates.

The scheduler gets a concrete version of this. Its own customers already build workflows by hand. If the product turns the ten most common custom workflows into shareable templates, each mid-market deal that reuses one skips about two weeks of configuration. That is the same change that moved segment B's variable cost from $110 to $70 above, now doing double duty as an acquisition mechanism: sales can promise a two-day setup instead of a three-week project.

Network effects and early stage bend the gates

Markets with network effects make the case against gating. Reid Hoffman and Chris Yeh argue in "Blitzscaling" (2018) that when a market tips toward one winner, the company that waits for segment-level payback hands the market to a rival that did not wait. In a two-sided marketplace, a subsidized user on one side makes the product more valuable for the other side, so the unit sheet understates the value of each acquisition.

That argument holds only under three conditions. The network effect can be measured: each new user raises retention or conversion for existing ones. The market is close to tipping. And money to cover the losses is available on known terms. Most B2B software meets none of them. The scheduler's mid-market customer does not make the product better for small crews. In such cases, "we will fix unit economics at scale" usually means the cost of custom work grows with revenue, and payback never improves.

The second limit is early stage. Below roughly 30-50 paying customers in a segment, the numbers are too noisy to decide on. A single lost account can move payback by months. At that stage, rely on gates one and two (fast value, repeat use) and treat the unit sheet as a forecast you are testing.

Regular reviews keep all this honest:

Weekly proof review: what did we test, what changed our view, which assumption is next?
Monthly economics review: CAC by segment and channel, the payback trend, contribution margin, changes in the retention curve and support cost per customer, which is often the hidden killer.
Quarterly portfolio review: which bets earned more money and which failed their gates.

If a review ends without a decision to scale, adjust or stop at least one bet, it was a status meeting.

Pass the gates in order or do not scale at all

For the scheduler, strong growth suggests its customers reached value quickly and kept coming back, so the first two gates plausibly held. The trouble sat at the third, and the growth numbers hid it: 8-month payback in one segment and 22.5 in the other averaged to a figure that looked fine. Blended numbers keep improving for as long as acquisition outruns the damage, which is why this gate is skipped more than any other. The decision rule fits on a sticky note: if any segment's payback exceeds your remaining runway in months, freeze acquisition spend in that segment until a template, a price change or a stop-list entry brings it back under.