Reviewed on July 18, 2026. A label such as “high margin” tells you nothing reliable about an untested business. It can hide unpaid founder hours, revisions, sales work, refunds, software that sits idle, taxes, and the cost of checking AI output. Revenue is not take-home pay, and a cheap model call does not make the surrounding service cheap to deliver.
This guide has one purpose: help you examine a single AI-assisted service with a limited amount of money and time. It does not rank business models, forecast earnings, or recommend an investment. The result may be “stop,” and that can be the most useful result of the exercise.
Begin with one deliverable
Write the offer in one sentence: a named buyer receives a defined deliverable, under stated quality and revision limits, for a stated price. “AI consulting” is too broad. So is “content automation.” The sentence must be specific enough that a customer can accept or reject the finished work without debating what was purchased.
Do not assume demand because a sector is growing or a tool is popular. The U.S. Small Business Administration’s market-research guidance separates broad market information from direct research with prospective customers. For this test, ask people in the intended role how they handle the task now, what failure costs them, who approves a purchase, and what would make the work unacceptable. Do not lead with an AI pitch.
| Assumption | Evidence that would support it | What would disprove it |
|---|---|---|
| Buyer | A person with authority describes the task as a current problem | Interest comes only from people who cannot approve or fund it |
| Scope | The deliverable and acceptance check fit in a short written brief | Every conversation produces a substantially different job |
| Inputs | The customer has permission to provide every file and data field required | The work depends on material that cannot enter the selected tools |
| Price | An unrelated buyer voluntarily agrees to the written price and terms | Approval appears only when the work is free or the scope is unclear |
| Quality | A review method catches factual, contractual, and formatting defects | Errors cannot be found consistently before handoff |
| Effort | Observed delivery and revision time stays within the preset labor limit | Cleanup or customer communication consumes the expected saving |
| Channel | The acquisition route permits the offer and its payment arrangement | Fees, account rules, or access changes break the proposed economics |
Mark each line as an assumption until observed evidence changes its status. A compliment, wait-list signup, or model demonstration is not proof that someone will buy acceptable work at the quoted terms.
Measure one completed job from inquiry to acceptance
Start the clock at the first sales activity, not when a prompt is submitted. End it after acceptance, revisions, invoicing, and file cleanup. Keep selling, intake, tool setup, generation, research, verification, editing, customer messages, revisions, and bookkeeping as separate time categories. Failed generations and abandoned drafts belong in the record.
Give owner labor a monetary rate even if no cash leaves the account. The rate is an explicit planning assumption, not a universal wage or a tax deduction. Choose a defensible replacement cost or opportunity-cost figure, record why you chose it, and rerun the calculation at a higher and lower rate.
| Entry | How to measure it | Common omission |
|---|---|---|
| Gross service cash received | Service cash actually received during the test window before any separately logged refund, excluding tax collected for an authority | Counting an unpaid invoice or refundable deposit as final revenue |
| Transaction and channel charges | Actual marketplace, payment, currency, withdrawal, and dispute fees | Using the advertised headline fee instead of the settlement record |
| AI usage | Usage statement or invoice allocation for successful and failed attempts | Ignoring retries, retrieval, storage, media, or tool calls |
| Other delivery cash | Job-specific licenses, data, subcontracting, hosting, and support | Treating a required subscription as free because it was already owned |
| Direct human labor | All delivery, revision, correction, and replacement hours multiplied by the chosen labor rate | Pricing the model’s time while pricing the owner’s time at zero |
| Sales and administration labor | Prospecting, calls, proposals, billing, records, and policy review | Recording only billable production |
| Fixed test costs | Business registration, insurance, accounting, software, equipment, and marketing allocated to the chosen window | Leaving setup purchases outside the experiment |
| Refund and rework event | Log gross cash refunded once; cross-reference rework hours to Direct human labor, rework AI use to AI usage, and replacement cash to Other delivery cash | Treating this event flag as another subtotal and subtracting the same cost twice |
Use three plain calculations:
- Delivery contribution equals Gross service cash received minus transaction charges, Other delivery cash, AI usage, cash refunds, and priced Direct human labor. Record each cost once.
- Delivery contribution ratio equals delivery contribution divided by Gross service cash received. Do not calculate it when the denominator is zero.
- Test-window operating result equals total delivery contribution minus fixed test costs and priced sales and administration labor.
These are decision labels for this worksheet, not accounting standards. Keep the underlying entries so an accountant can map them correctly. If contribution per accepted job is positive and reasonably stable, fixed test-window costs divided by contribution per accepted job estimates how many accepted jobs recover those fixed costs only. It is not an operating break-even count because this worksheet records sales and administration labor separately. To estimate operating break-even, assign that labor consistently: subtract its variable per-job amount from per-job contribution and add only truly fixed sales and administration labor to the numerator. Split mixed costs into their fixed and variable parts, and do not count the same labor twice. This follows the single-service structure in the SBA break-even explanation, but the result becomes meaningless when job scopes or contribution amounts vary widely.
Quote from scope and observed work
A model’s price page is an input, not a service price. Buyers are also paying for intake, judgment, verification, communication, and accountability. Conversely, a buyer’s stated value does not prove you can capture a percentage of it.
Before quoting, write down the scope, acceptance method, number of revision rounds, delivery date, excluded work, payment timing, refund terms, and what happens when required inputs arrive late. Then calculate the price under three workload assumptions: expected effort, a plausible rework load, and the maximum effort allowed before you pause for a change order. If only the most optimistic row clears your minimum contribution, the quote is not ready.
For metered tools, use current invoices and usage exports. Published rates can change and may omit adjacent charges. OpenAI’s current API price table, for example, separates input, output, caching, and processing modes across models. A brochure estimate cannot replace the cost of the workflow you actually ran.
Run a bounded paid test
A free sample can reveal production defects. Only a voluntary purchase by an unrelated buyer tests whether the written offer and price can produce a transaction. Neither result establishes a repeatable market.
- Freeze the brief. Save the buyer, deliverable, input permissions, acceptance check, price, revision limit, and deadline before work begins.
- Use permitted material. Until confidentiality and vendor handling are settled, keep client secrets, personal data, and regulated records out of AI systems.
- Log the whole job. Capture every labor category, vendor charge, platform deduction, failure, and revision.
- Review without the model grading itself. Check claims against authoritative material, inspect calculations, confirm rights, and apply the customer’s acceptance criteria.
- Close the record. Note acceptance or rejection, final money received, refunds, outstanding obligations, and the customer’s reason in their own words.
Do not publish a case study, testimonial, accuracy percentage, time-saving claim, or income claim from this exercise unless the underlying evidence and customer permission support the exact statement. The FTC's advertising guidance says advertisers need evidence before an ad runs. In Operation AI Comply, the FTC says its complaint alleged that an AI service had not tested whether its output matched the human professional it claimed to replace. The agency’s business-opportunity guidance also warns readers to demand written support for earnings claims. This worksheet supplies no such earnings evidence.
Count dependencies as business risk
An offer can look viable while one external account carries the whole operation. Record what happens if a model is withdrawn, an API rate changes, a marketplace suspends the account, a payment processor holds funds, or the sole customer leaves. The answer need not be a costly backup on day one, but it should identify how work, records, and client communication can continue.
Marketplace rules belong in the cost model. Upwork’s current legal center identifies its User Agreement as version 8.1, effective January 8, 2026. It incorporates other agreements, assigns freelancers responsibility for applicable service fees, and contains a non-circumvention rule for relationships formed on the site. If you use a different channel, read that channel’s live contract instead. Do not copy an Upwork assumption into another platform.
Model terms matter just as much. OpenAI’s current Services Agreement makes the customer responsible for rights in inputs and for judging output accuracy and suitability. Its API data-controls documentation describes default abuse-monitoring retention and different application-state behavior by endpoint. Those statements apply to OpenAI’s business services, not to every AI vendor. Recheck the product, account type, endpoint, data settings, and contract that your job actually uses.
Separate tax cash from operating cost
Taxes should appear in the cash plan even when they are not placed in the delivery-contribution formula. For U.S. readers, the IRS self-employed tax center says self-employed people generally file an annual return and pay estimated taxes quarterly, and that self-employment tax may apply alongside income tax. The correct treatment depends on entity, location, income, expenses, and personal circumstances.
Keep receipts, invoices, settlement statements, contracts, mileage or equipment records when relevant, and the logic behind allocations. The IRS recordkeeping page says the chosen system should clearly show income and expenses and that supporting records must substantiate return entries. It does not say every software purchase is deductible.
Readers outside the United States need the rules of their own jurisdiction. Sales tax, VAT, invoicing, currency treatment, and registration can change a quote before income tax is considered.
Bring in professional advice at the real boundaries
Speak with a qualified tax professional before relying on expense deductions, setting estimated payments, choosing an entity, handling multi-jurisdiction sales, or paying workers. Consult a lawyer licensed for the relevant jurisdiction before using customer confidential information in AI tools, drafting unusual ownership or indemnity terms, operating in a licensed field, processing sensitive personal data, hiring or classifying workers, or promising compliance to a client.
Copyright deserves a specific check when the deliverable includes generated text, images, audio, video, or code. The U.S. Copyright Office’s 2025 copyrightability summary says protection for generative output depends on sufficient human-authored expressive elements; prompts alone do not settle the issue. Vendor language about output ownership does not guarantee copyright protection or freedom from third-party claims.
EU-facing work may also encounter AI-specific disclosure duties. The European Commission says certain Article 50 transparency obligations become applicable on August 2, 2026, including defined cases involving interactive systems, deepfakes, and AI-generated public-interest text. The Commission’s current transparency code page explains the scope at a high level. Whether a particular service is covered is a legal question, not a conclusion to draw from this article.
Set the stop rule before outreach
Choose four values in advance: a final review date, a maximum cash exposure, a maximum total labor allowance, and a minimum delivery contribution for an accepted job. Stop as soon as either cap is reached. Do not extend a cap merely because money or time has already been spent.
At the review date, continue only when an unrelated buyer voluntarily paid under the written terms, accepted the work, every cost and labor category was logged, the quality check passed, and delivery contribution met the preset minimum after pricing owner labor. If any condition fails, close the test or rewrite the offer before spending again.
One accepted job justifies another measured repetition, not hiring, debt, a long software contract, or an earnings forecast. Compare several completed records before taking on fixed commitments. A clean stop, supported by the ledger, is evidence that the framework worked.
