Last updated: September 2026
Roughly $629,370. That was the year-three revenue an AI model projected for a coffee-shop concept in a published stress test by LivePlan — a number that looked rigorous, carried three years of monthly detail, and quietly rested on assumptions nobody had verified. That is the exact failure mode this guide exists to prevent.
Here is the tension in 2026: 67% of small business owners and marketers already use AI for content and research work, and 79% say it has improved their output quality. Feasibility studies are the document type that benefits most from that speed — and the document type that AI can damage the most, because a feasibility study's entire value is that its numbers survive scrutiny.
So this guide covers how to write a feasibility study with AI in a way that holds up: the five tests every study must pass, a step-by-step workflow, a prompt pack, a verification protocol for AI-generated numbers, a go/no-go scorecard, and the honest list of times you shouldn't run a study at all.
Which document do you need right now?
Most people searching for feasibility guidance actually need a different document — or need two in sequence. Start here:
| Document | Core question | When you need it | Typical length | What it decides |
|---|---|---|---|---|
| Feasibility study | Should this project exist at all? | Before committing money or time | 15–30 pages | Go or no-go |
| Business plan | How do we execute and fund it? | After feasibility says go | 10–50 pages | Funding and operations |
| Pitch deck | Why should you bet on us in 10 slides? | In front of investors | 10–15 slides | The next meeting |
| Lean canvas | What's the sharpest version of the hypothesis? | Idea stage, pre-research | 1 page | What to test first |
| Financial model | Do the numbers connect and reconcile? | Alongside any of the above | Spreadsheet | Cash reality |
A feasibility study answers one question — should this project exist — by testing its market, technical, financial, operational, and legal assumptions with numbers you can defend under questioning. AI accelerates the research and the first drafts; you own the assumptions and the final go/no-go call. If the answer is yes, the study then graduates into a business plan written with AI, which is a different document with a different audience.
What a feasibility study actually tests
Asana's project management guide and the Corporate Finance Institute converge on the same five lenses. Here is each one mapped to the concrete job AI can do for it — and the job it can't.
1. Market feasibility. Is there real demand, at a price that works? AI can compile competitor scans, draft survey questions, and structure a price-sensitivity analysis in minutes. It cannot tell you whether the 40 people who clicked "yes" on a survey will actually pay. Send the survey yourself.
2. Technical feasibility. Can you actually build or deliver this with the tools, skills, and vendors available? AI is genuinely strong here: it will generate vendor checklists, equipment lists, and cost tables you can correct rather than research from scratch.
3. Financial feasibility. Startup costs, operating costs, revenue scenarios, break-even, ROI. AI drafts the structure and the scenario math. Every input must come from quotes, listings, and benchmarks you can cite. This is the section where hallucinated numbers cause the most damage.
4. Operational feasibility. Do you have the team, the processes, and the capacity? A useful trick: have AI draft the operating procedures the project implies, then check whether anyone on your team could realistically run them. That draft doubles as the start of your SOPs if the project goes ahead.
5. Legal and regulatory feasibility. Licenses, permits, zoning, insurance, tax. This is the one lane where AI gets read-only access: it can produce a checklist of questions to ask a professional, and nothing more. It must never produce the answers.
A full study runs the five tests in order and kills the project fast if any one of them fails. That ordering is the entire efficiency of the format: preliminary analysis first, expensive research only if the cheap screen passes. CFI's process starts with exactly that preliminary screen before any money is spent on surveys or specialists.
Where AI earns its keep — and where it quietly lies
For the drafting layer, the productivity numbers are real: 68% of businesses report increased content-marketing ROI from AI assistance, and 71% are very satisfied with AI writing tools, per Semrush's aggregated statistics. A feasibility study's narrative sections — the market description, the competitive scan, the report formatting — benefit from the same speed advantage.
Two failure modes are specific to financial AI work, and both matter for a document whose purpose is due diligence:
- Invented precision. LivePlan's FAQ on AI forecasting asks why AI-generated startup forecasts often look more accurate than they are. The answer: models produce clean monthly progression and confident figures, which reads as rigor. The cleanliness is a formatting property of language models, not evidence of research.
- Stale or transplanted pricing. Ask a model for equipment costs and you may get a price from a market that isn't yours, at a date you can't verify. Every cost line needs a source URL and a date checked by a human.
The working rule: AI is your research assistant and first-draft writer. It is never your source. A number without a citation attached is not part of a feasibility study; it's part of a wish list.
How to write a feasibility study with AI, step by step
1. Run the preliminary screen (kill-fast test). One page: the concept, the customer, the price, the one thing that would kill it. Prompt the model to argue the project fails — hard. If the strongest argument against survives your rebuttal, stop here. You just saved months; a simple study takes a few weeks and a complex one several months, per Asana, so dying early is the highest-ROI outcome a study can produce.
2. Test the market. Define the target customer precisely, then gather demand evidence: search volume, competitor pricing, published reports, and ideally 30+ survey responses or ten customer interviews. Use AI to design the survey and summarize transcripts; keep the raw collection human.
3. Test technical and operational delivery. Build the capex table (equipment, buildout, licenses) and the opex table (staffing, rent, subscriptions) with AI drafting the structure and you filling every cell from quotes. SBA guidance applies here: plan for at least one year of monthly operating expenses, five if you can.
4. Build the financial case. Projected income statement, opening-day balance sheet, and cash flow under three scenarios: base, downside (−30% revenue), and cost shock (+20%). Compute break-even with the SBA formula — Fixed Costs ÷ (Price − Variable Costs) — and add a ~10% buffer for the costs nobody predicts.
5. Run the legal screen as a checklist. Have AI list every license, permit, registration, and insurance type your sector and jurisdiction typically require. Take that list to a local professional and confirm it. The AI output is the question set, never the answer.
6. Review vulnerability. Re-read the whole study asking one question: which single assumption, if wrong, flips the conclusion? Mark it. LivePlan's coffee-shop test found two such assumptions — customers per day and average ticket — that individually moved break-even by quarters.
7. Score it and decide. Run the go/no-go scorecard below. The decision rule: any lens scoring below 3, or a weighted total below 3.4, means no-go or redesign. Write the decision down with the date and the deciding evidence, whether the answer is yes or no.
8. Hand off what survived. The verified cost tables, the break-even math, and the market evidence graduate directly into your business plan's financial and market sections. Nothing gets re-researched; that's the study compounding into the plan.
The prompt pack
Market sizing with forced citations:
Estimate the reachable market for [product/service] in [city/region]. Build it
bottom-up: target customers × reachable share × annual spend. For every figure,
name the source, and if you cannot cite a verifiable publication, write
SOURCE NEEDED instead of a number. Then rate your overall confidence 1–5 and
explain what most lowers it.Demand survey design:
Design a 10-question market survey for [concept] targeting [customer type].
Rules: every question must test willingness to pay, price tolerance, or current
solution dissatisfaction — not generic interest. Include two attention-check
phrasings and a final question asking for a 15-minute follow-up interview.
Do not include any question whose answer would just flatter the concept.Cost tables with capex/opex split:
Build two tables for [business type] in [location]. Table 1 — startup costs
(capex): equipment, buildout, deposits, licenses, initial inventory, with a
"source to verify" column. Table 2 — monthly operating costs (opex): staffing,
rent, utilities, software, marketing, insurance. Leave every cell you cannot
source as TBD. Do not fill gaps with typical-sounding numbers.Break-even with the formula embedded:
Using my verified cost tables below, compute monthly break-even units with
Fixed Costs ÷ (Price − Variable Costs). Then recompute with a 10% buffer added
to fixed costs, and again with price 15% lower. Show which input break-even is
most sensitive to. Cost tables: [paste].Red team — argue it fails:
You are a due-diligence analyst whose job is to kill weak projects. Read this
feasibility draft and attack it: the 5 weakest assumptions, the 3 numbers most
likely to be wrong, and the single scenario in which this business fails
quietly rather than loudly. Rank your attack points by severity. Draft: [paste].You can run this pack in ArWriter with its article writer, which keeps your project context, currency, and verified figures attached across prompts so the finished study reads as one document. Drafting work starts at $4.99/month on the Plus tier — a rounding error against the cost of a wrong go decision.
Numbers that survive due diligence
A study "survives" when every load-bearing number traces to a source a stranger can check. Build this audit table before anyone else reads the document:
| Assumption | Value in study | Source | Confidence | Verifier |
|---|---|---|---|---|
| Daily customers (base) | 140 | Foot-traffic count, competitor benchmark | Medium | You, on-site |
| Average ticket | $8.75 | Competitor menu pricing, 3 venues | High | You |
| Buildout cost | $180,000 | Two contractor quotes | High | Quotes on file |
| Rent | $4,200/mo | Listing + agent confirmation | High | Contract |
| Break-even month | Month 14 (stress) | SBA formula + 10% buffer | — | Recompute monthly |
The LivePlan stress test behind the 140-vs-95 logic: the model projected 140 customers/day at an $8.75 ticket, and flipping to 95 at $7.25 — both entirely plausible real-world values — changed the business from viable to not. A +20% buildout overrun did further damage. The studies that survive due diligence are the ones that already contain their own worst case.
When you should skip the study
Asana's guide includes a useful honesty list, and it deserves repeating: don't run a study when the decision is already made (you're just decorating it), when the project is small enough that a one-page canvas and a weekend of research would answer it, or when the window is so short that a months-long study guarantees failure by delay. Studies exist to prevent expensive mistakes, not to manufacture confidence for decisions already made.
The $629,370 forecast: a cautionary case
Daniel Okafor had run two restaurants in Rotterdam and wanted his own place — a 40-seat bistro in the Kralingen district. He let an AI model build his year-three revenue projection first and work backward, the inversion of how due diligence actually works. The model produced a beautiful document: $629,370 in year-three revenue, smooth monthly growth, a coherent staffing ladder.
Then his accountant asked where the daily covers figure came from. It didn't come from anywhere — the model had generated it. Daniel re-ran the projection with real inputs gathered over three weeks: seated-competition counts on Tuesday and Saturday nights, average ticket from three comparable menus, and a rental quote in writing. Base case landed near 140 covers a day at a $8.75-equivalent ticket; his downside case was 95 at $7.25. Under the downside, break-even moved past month 14 and his working-capital requirement nearly doubled.
He went ahead — but smaller. He signed for a location with 20% lower buildout cost than the AI projection assumed, kept the stress case in the document he showed his bank, and set the kill criterion in writing: below 60% of base covers for three consecutive months, he renegotiates or exits. The study's job was never to say no. It was to make sure he decided with his own numbers.
The go/no-go scorecard
| Lens | Weight | Score (1–5) | What a 1 looks like | What a 5 looks like |
|---|---|---|---|---|
| Market | 30% | — | Survey interest, no spending evidence | Repeat purchases or signed LOIs |
| Financial | 30% | — | Break-even only in base case | Break-even survives −30% revenue |
| Technical | 15% | — | Delivery path undefined | Quotes and vendors confirmed |
| Operational | 15% | — | No team for core functions | Team mapped, gaps costed |
| Legal | 10% | — | Unknown requirements | Checklist confirmed by professional |
Decision rules: any lens at 2 or below is a hard stop regardless of total. A weighted total of 3.4 or lower means redesign, not rationalization. And the scorecard is dated — a no-go today at 2.9 is a document you can honestly revisit if the market shifts.
Questions people ask about feasibility studies
What are the 5 major components of a feasibility study?
Technical feasibility (can it be built), market feasibility (will people pay), financial feasibility (do the numbers work), operational feasibility (can your team run it), and legal feasibility (is it permitted). A complete study tests all five — skipping one is how projects fail in exactly the dimension nobody examined.
How long does a feasibility study take to complete?
A simple study takes a few weeks; complex ones run several months, according to Asana's guidance. The variable is almost always primary research — surveys, interviews, quotes — not the writing. AI compresses the drafting phase dramatically but cannot compress the time it takes to gather evidence you can cite.
Who should conduct a feasibility study?
The founding team, with specialists for finance and legal review. Outside consultants add distance and format polish, but the judgment calls — is this demand real, can we deliver — land best with the people who will live with the consequences. AI now covers the research-assistant role consultants used to bill for.
What happens if a feasibility study shows a project isn't viable?
You celebrate, quietly, because the study just returned your entire investment by preventing it. A documented no-go should specify what would need to change — price, cost structure, market size — to revisit the decision. Many viable businesses are the second or third design of a project whose first version failed its own study.
Why do AI-generated startup forecasts often look more accurate than they are?
Because language models produce clean, internally consistent numbers — smooth monthly progressions, tidy margins — and visual consistency reads as analytical rigor. The precision is a property of the output format, not the underlying research. Verification protocols exist precisely to separate those two things.
What's the difference between a financial forecast and a financial prediction?
A forecast is a conditional projection: given these stated assumptions, here is the expected outcome. A prediction asserts what will happen. Feasibility studies should contain forecasts with visible assumptions, because reviewers — and lenders specifically — evaluate the assumptions as much as the arithmetic.
How is this different from a business plan?
The study decides whether the project should exist; the plan describes how to execute and fund it once the answer is yes. The study's verified cost tables and market evidence flow directly into the plan's financials. One is a gate, the other is a road map — and conflating them is why weak projects get plans instead of autopsies.
What to do next
If you're holding a concept, run the preliminary screen tonight — it's one prompt and one honest hour. If the screen passes, work the steps in order and don't let the model write a single uncited number. And when the scorecard says go, your next document is already mapped: turn the surviving numbers into a business plan written with AI, then a pitch deck built with AI for the rooms it opens.
ArWriter and its article writer handle the drafting layer of this entire workflow from $4.99/month, with longer-form auto-writing tools on higher tiers. Later, when the project is live, the same research becomes a lead magnet created with AI, an ebook written with AI for depth, a white paper written with AI when the buyers are corporate, a webinar script written with AI to present it live, and a nurture email sequence built with AI to work the audience it builds.
