An experienced investor can usually tell within the first five minutes of looking at a financial projection whether it was built from real assumptions or reverse-engineered from a target the founder wanted to hit. The tell isn't sophistication — plenty of unconvincing projections have beautiful formatting and elaborate scenario tabs. The tell is whether the underlying assumptions connect to anything real: actual historical data, a defensible growth rate, unit economics that hold together at scale. AI can help build a projection model fast, but speed isn't the problem most founders have — the problem is that a fast, well-formatted model built on ungrounded assumptions is just a more polished way of failing the same scrutiny.

What investors are actually testing for

Before building anything, it's worth being precise about what a projection is being scrutinized for, because it's rarely "is this exactly what will happen" — every experienced investor knows projections are wrong. What's actually being tested is whether the founder understands their own business's economics well enough to have built a coherent model, and whether the assumptions inside it are internally consistent and roughly defensible against what's already known about the market and the company's own historical performance. A model that's "wrong" in its exact numbers but built on a sound, explainable method passes this test; a model that's suspiciously exact but falls apart the moment someone asks "why did you assume 15% month-over-month growth starting in month four" fails it immediately.

Step 1: anchor every top-line assumption to real historical data

The single highest-leverage thing you can do is refuse to let any growth assumption exist without a historical anchor. If you have any operating history at all — even six or nine months — feed that actual data to AI and have it calculate your real historical growth rate, customer acquisition cost trend, and retention curve, rather than starting from an assumed target growth rate and working backward. A worked example: Ferro Analytics, an early-stage B2B software company, initially projected 20% month-over-month growth for the next 24 months because that number appeared in a competitor's public pitch deck. Once AI was given their actual signed-customer history — nine months, growth that had actually averaged 11% month-over-month with real deceleration in months 7 through 9 as the founder's personal network of early customers ran out — the honest anchor point was 11%, decelerating further as the model moved past the founder's direct network into paid acquisition, a transition that predictably slows growth for almost every early-stage company.

This is uncomfortable, because 11% decelerating looks much less exciting than 20% flat, but it is the number that survives a follow-up question, and the 20% number was not.

Step 2: build unit economics before you build the top line

Investors scrutinize unit economics — customer acquisition cost, gross margin per customer, payback period, lifetime value — more closely than the top-line revenue number, because unit economics reveal whether growth is actually value-creating or just expensive. Have AI calculate your real customer acquisition cost from actual marketing and sales spend divided by actual customers acquired in the same period (not divided by leads, not divided by trials — actual paying customers), and your real gross margin per customer from actual delivery costs, including the support and infrastructure costs that founders often forget to allocate.

For Ferro Analytics, this step surfaced a problem the founder hadn't noticed: fully loaded customer acquisition cost was $4,200, but average first-year revenue per customer was only $3,600 — meaning the company was losing money on every new customer in year one, recovering it only in year two if the customer renewed. This is a survivable, common early-stage pattern if retention is strong and it's disclosed honestly with a clear payback timeline; it is a model-destroying discovery if an investor finds it themselves in the underlying numbers after the founder's narrative claimed healthy unit economics.

Important

Never let a projection model imply unit economics are healthier than your real numbers show, even by omission. An investor who finds a bad number you disclosed and explained treats it as evidence of founder competence. An investor who finds the same bad number themselves, uncovered rather than disclosed, treats it as evidence the rest of the model can't be trusted either — and reasonably so.

Step 3: build the expense side with the same rigor as revenue

Most founder projections put far more care into modeling revenue growth than expenses, which is backward from an investor's perspective, because expense assumptions are usually more knowable and their errors compound just as fast. Feed AI your actual historical spending by category — payroll, infrastructure, marketing, office and administrative — and have it project each category forward based on real historical growth rates and known future commitments (a planned hire, a facilities change, a software contract renewal), rather than a single blended "expenses grow with revenue" assumption that hides the fact that payroll, your largest cost, doesn't scale continuously — it jumps in discrete steps every time you hire.

Step 4: build three scenarios, and make the assumptions behind each one explicit

A single-scenario projection reads as either naive or manipulative to an experienced investor — naive if the founder genuinely believes point estimates are meaningful this far out, manipulative if the single scenario shown is suspiciously the best-looking one. Build a base case grounded in the historical anchors from steps 1 and 2, a downside case that stress-tests what happens if growth decelerates faster or churn runs higher than the base case assumes, and an upside case tied to a specific, named catalyst (a new sales channel opening, a product launch) rather than just "growth is faster." AI is genuinely useful here for running the same set of formulas across three sets of assumptions quickly and consistently, which is tedious to do by hand and easy to get inconsistent across scenarios if you don't.

Step 5: stress-test the model against the questions you know are coming

Before sending a projection anywhere, have AI act as an adversarial reviewer: feed it the full model and ask it to identify every assumption that isn't directly traceable to a historical number or a named, specific driver, and flag anything that looks like a round number chosen for convenience (exactly 15%, exactly $50 CAC, exactly 5% churn) rather than a calculated figure. Round, convenient numbers are one of the fastest tells that a model was built backward from a target rather than forward from real data, and they're also one of the easiest things to fix once flagged.

What survives scrutiny and what doesn't

A model survives scrutiny when every assumption has a one-sentence, honest answer to "where did this number come from" — a historical average, a named contract, a documented industry benchmark cited specifically rather than vaguely. A model fails when the honest answer to that question is "it felt like a reasonable number" or "we needed the model to show profitability by month eighteen." Investors have seen thousands of models; they are extremely good at detecting the second kind of answer even when it's dressed up in confident language, because the tell isn't in the number itself — it's in how quickly the founder's explanation for it falls apart under one follow-up question.

The build order matters as much as the numbers

Notice that this entire process runs in a specific order: historical anchors first, unit economics second, expenses third, scenarios fourth, adversarial review last. Founders who build in the opposite order — start with a target valuation, work backward to a revenue number that supports it, then invent growth assumptions to hit that revenue number — produce models that look identical on the surface but collapse under exactly the questions described above, because nothing in them actually traces back to something real.

What to do when you genuinely have no historical data yet

Pre-revenue companies face a real version of this problem: there's no historical growth rate to anchor to, because there's no history. The honest approach here isn't to invent one — it's to substitute the closest available proxy and label it clearly as such. Have AI pull comparable public or disclosed growth benchmarks for companies at a similar stage in a similar market, and use those, explicitly cited, as the base-case assumption rather than a number invented to look achievable. Pair this with a bottom-up sales capacity model instead of a top-down growth-rate model wherever possible — if your plan depends on closing new customers, project revenue from a specific, named pipeline (how many sales conversations are realistically possible per month, at what historically observed close rate, at what average contract value) rather than from an abstract percentage. A bottom-up model built from operational assumptions an investor can independently sanity-check — how many salespeople, how many calls each can realistically make, a close rate benchmarked against the specific industry — survives scrutiny in a way an unanchored top-down growth percentage never does, precisely because every input is checkable against something outside the model itself.

It's also worth explicitly separating what's a forecast from what's a target inside the deck itself. A slide that says "we are modeling 11% month-over-month growth based on our first nine months of data, and we believe disciplined execution against the plan below could push this toward 15%" tells an investor exactly which number is grounded and which is aspirational, and that transparency reads as founder maturity rather than as a weaker pitch. Blurring the two together, so the aspirational number quietly becomes the modeled number, is precisely the pattern experienced investors have seen fail before and are actively screening for.

The AI-Powered Business Operating System devotes a full section to exactly this build order, with a worked template for connecting historical data to forward projections without letting a target number quietly drive the assumptions. The book's Financial Projection Template Suite gives you the structural spreadsheet to build from, and the Revenue and Expense Tracking resource covers the unglamorous but essential discipline — keeping clean historical data in the first place — without which none of the steps above are possible to do honestly.