AI Profitability for Freelancers: Quebec Accepted-Work Audit

Executive Summary

AI profitability for freelancers, meaning the project economics of artificial intelligence (AI), should be measured at the level of accepted client work. The proposed audit compares retained collected revenue with paid project costs and all owner hours, including setup, briefing, production, checking, testing, revisions and administration. It reports cash contribution, effective contribution per owner hour and net time difference separately. The financial ledger excludes collected sales taxes, following Revenu Québec's separation of pretax amounts, GST/HST and QST; the worksheet is a management measure, not a tax-return method. ^{[1] [2]}

Research does not justify a universal Montreal profit multiplier. METR's early-2025 experiment found 19% longer task completion among 16 experienced developers completing 246 tasks in familiar repositories. ^{[3] [4]} Its February 2026 follow-up warns that selection and time measurement make current productivity estimates unreliable. ^[5] Its May 2026 report surveyed 349 technical workers and found median self-reported work value of 1.4 to 2 times, versus a reported speed increase of 3 times. These are subjective outcomes in a convenience sample, not verified freelance earnings. ^{[6] [7] [8]}

Tool costs require invoice-level evidence. As of October 2026, Claude Pro's support page lists US$20 monthly, says regional pricing varies, and excludes Console API usage. ^{[9] [10]} Provider prices should therefore be recorded with currency, plan, billing period and a disclosed project allocation. Bank of Canada rates are indicative; actual CAD payments provide the evidence for already-paid foreign charges. ^[11] No tax recovery, commercial permission or confidentiality protection is assumed. Accounting and contract questions require the applicable official guidance, records and professional advice.

The Hypothetical Examples show the decision clearly. A CAD$1,200 fixed-fee project moving from 12 to 8 owner hours, with paid costs rising from CAD$100 to CAD$140, improves contribution per hour from CAD$91.67 to CAD$132.50 but reduces cash contribution by CAD$40. These are author calculations from invented inputs, not Quebec observations. Freed hours generate extra income only when sold as additional paid work; unsold availability remains a separate capacity outcome. Keep a workflow when accepted quality and the chosen measured objective improve, change it when rework or allocation weakens the result, and stop it when repeated comparable work misses that objective. Preserve the underlying financial evidence rather than substituting a perceived speedup for recorded income. ^[12]

Introduction and Background

AI profitability for freelancers depends on what survives client review, what the client pays, and what the freelancer spends to deliver it. A draft completed quickly can still require extensive checking. A successful fixed-fee assignment can yield more contribution per working hour without increasing total receipts. An hourly assignment can finish earlier while bringing in less revenue. These are different business outcomes, and an audit should keep them separate.

This report proposes an accepted-deliverable audit for Montreal and Quebec independent developers, bilingual writers, designers and consultants. Its unit of observation is an agreed deliverable that reaches a documented acceptance decision. The audit is a management worksheet, not an original survey or a Quebec earnings benchmark. Revenu Québec describes accounting records as providing a running view of the business's financial position; project-level workflow records add the owner-time information needed for this decision. ^[13]

The publishing context is adjacent to this decision. 2727 Coworking supplies offices and associated services, including internet and a conference room, rather than the AI workflows being evaluated here. ^[14] Its existing rate-floor discussion addresses funding non-billable time and operating costs; this report addresses whether a particular workflow improves economics after delivery. ^[15] Its existing AI-tools guide also identifies transcript review as work that remains after generation. ^[16]

Public evidence warrants this narrower question. METR's historical developer experiment found longer completion times in a specific setting, while its newer follow-up warns that current estimates are unreliable. ^{[3] [5]} Separately, Plutio's freelancer article and Hugo's French practitioner essay link the publisher's freelance research in adoption and billing discussions. Those editorial links show different uses of research; they do not independently validate a Canadian adoption percentage or establish demand for this worksheet. ^{[17] [18]}

The practical decision is therefore: keep, change or stop the workflow after measuring comparable accepted work. Faster drafting is useful only to the extent that it improves an explicitly chosen outcome.

Methodology: Measure Accepted Work

Establish the acceptance boundary

Define the assignment before starting the timer. Record scope, quality requirements, revision allowance and acceptance evidence. The existing Quebec independent-contractor guide treats acceptance criteria, milestones, price and change control as elements defining the mandate. Use that guide as contractual context; obtain qualified advice for the actual agreement rather than assuming an audit changes its terms. ^[19]

For this worksheet, acceptance means the client has approved the agreed scope, or the documented contractual acceptance condition has been satisfied. It does not mean an AI system produced an answer, a designer exported a file, or an internal test passed. Record unresolved conditions explicitly. Keep an accepted but unpaid assignment separate from collected receipts so payment delays do not disappear.

A comparable baseline must have similar scope and a similar acceptance standard. Choose a previous assignment, a documented manual workflow, or a prospective comparison. Label estimated baselines as estimates. Familiarity with the client, reused assets and newly learned skills can change time independently of AI. METR's original experiment deliberately studied developers working in repositories they knew well, which is one reason its result cannot be transferred mechanically to a new freelance project. ^{[20] [21]}

Table 1 sets out the proposed project ledger. The fields are this report's audit design; the underlying recordkeeping discipline follows official guidance to retain income dates, amounts, sources and supporting documents. ^{[22] [23]}

The ledger links acceptance evidence to financial evidence without equating them. A technically finished project can have unpaid invoices; a paid deposit can precede accepted work. Comparing consistent observation windows makes those differences visible. The Canada Revenue Agency (CRA) calls for original documents supporting income entries. ^[12]

Log all owner hours

Use a timer or contemporaneous notes, not a remembered impression. Record all hours worked by the owner on the assignment, including contractor supervision. Keep paid contractor charges in project costs. If total team effort is also measured, record contractor hours separately and label the combined measure as total team hours. Capture the owner’s time for:

• Setup: environment, files, context gathering and access configuration.
• Briefing: prompts, examples, terminology and task decomposition.
• Production: drafting, editing, coding, layout and integration.
• Verification: source checks, tests, comparisons and technical review.
• Corrections: rejected generations and repairs to generated output.
• Client revisions: feedback review, communication and revised delivery.
• Acceptance administration: packaging, handover and approval follow-up.

These categories are proposed measurement fields, not industry-average time shares. GitHub's own guidance encourages reviewing, testing and validating generated suggestions, so omitting those activities would exclude documented parts of responsible use. ^[24]

Record overlapping agent runs separately from owner time. Waiting while doing another assignment is different from actively supervising a run. METR's follow-up identifies concurrent agent use and differences in work quality as interpretation problems; a freelancer should state the timer rule before comparing projects. ^[25]

Cash Costs and Quebec Accounting

Keep cash, allocation and tax treatment distinct

Start with collected pretax revenue, then record refunds and relevant paid costs. Maintain invoices and outstanding receivables in separate columns. This report deliberately uses a cash-management measure; it does not prescribe the accounting method used for a tax return. Revenu Québec generally describes accrual accounting for sole proprietors, with a cash-method option for commission earners. ^[2]

For registered businesses, Revenu Québec's income-journal example separates the amount before tax, Goods and Services Tax/Harmonized Sales Tax (GST/HST), Quebec Sales Tax (QST), and the tax-inclusive total. ^[1] Its expense-journal example excludes tax paid to suppliers, but it also warns that a taxpayer may need different entries. Therefore, show supplier tax separately and ask an accountant which amounts belong in the management-cost measure for the reader's registration and circumstances. No deduction or tax recoverability is assumed here. ^{[26] [27]}

Apply a disclosed allocation rule to subscriptions. A monthly charge can be divided among audited projects using recorded usage or another consistent rule. Include idle paid capacity when evaluating whether to retain the subscription for the month. For a decision about cancelling a plan, distinguish avoidable future spending from charges already incurred. Do not allocate the same charge fully to every assignment.

Provider documentation illustrates why a generic “AI subscription” field is insufficient:

• Claude Pro: its support page lists US$20 monthly, with local pricing varying by region. ^[9]
• Annual billing: Claude's pricing page lists $200 paid upfront for the annual Pro option; an allocated monthly amount differs from the payment date. ^[28]
• Separate API costs: Pro does not include Claude Console application programming interface (API) usage, and Anthropic's API documentation states payments are in United States dollars (USD). ^{[10] [29]}
• Usage-based credits: GitHub currently documents Copilot usage in AI credits; its pricing page says flex allotments can change. ^{[30] [31]}
• Unused credits: Adobe says generative credits do not roll into the following month. ^[32]
• Token meters: Google's Gemini Developer API labels paid token prices in USD per million tokens. (Source: ai.google.dev)

These are billing examples verified for this report, not recommendations or a promise of an unchanged price. The audit uses the subscriber's actual invoice, plan and billing period. Adobe's documentation also distinguishes unlimited standard image and vector generation on paid Firefly plans from its broader credit system; a count of outputs alone is consequently not a universal cost measure. ^[33]

Reconcile to CAD and preserve evidence

Use Canadian dollars (CAD) throughout the project comparison. For already-paid foreign charges, record the actual CAD amount from the payment statement and any separately identifiable conversion fee. A forecast can disclose an exchange-rate assumption, but it should not silently replace an actual charge. Bank of Canada rates express foreign-currency units in CAD and are indicative, rather than a cardholder's transaction-specific cost. ^{[34] [11]}

Include payment-processing costs where they are incurred. Stripe's Canada pricing page distinguishes successful domestic-card transactions and separately states that standard pricing has no setup or monthly fee. The relevant audit entry is still the project's actual charged fee, because payment method and account terms matter. ^{[35] [36]}

Keep receipts and vouchers. CRA explicitly requests them for business purchases, while Revenu Québec requires electronic records to remain accessible and readable. ^{[37] [38]} A management cost does not automatically establish a tax deduction: CRA distinguishes current and capital expenses and business from personal portions. Consult an accountant before making deduction or recoverability decisions. ^{[39] [40]}

Core Calculations and Worked CAD Scenarios

Calculate contribution before calling it profit

The worksheet defines cash contribution as retained collected pretax revenue minus relevant paid project costs. Retained revenue is collected revenue after refunds. Project costs include paid external inputs and the disclosed tool-cost allocation. This definition measures resources remaining before excluded annual overhead, owner taxes and owner compensation. It is not take-home profit. Its separation of pretax revenue follows Revenu Québec's income-journal structure. ^[1]

Use these proposed formulas:

• Cash contribution: retained pretax receipts minus relevant paid costs.
• Contribution margin: cash contribution divided by retained pretax receipts, where receipts are positive.
• Effective contribution per owner hour: cash contribution divided by all hours worked by the owner on the assignment, including setup, briefing, production, verification, revisions and administration, where owner hours are positive.
• Net time difference: comparable baseline owner hours minus AI-workflow owner hours.
• Incremental contribution: AI-workflow contribution minus comparable baseline contribution.
• Economic surplus: cash contribution minus a disclosed opportunity rate multiplied by owner hours.
• Cash return on investment (ROI): incremental cash contribution divided by incremental paid AI spending, only where that spending is positive.

The last formula is an explicitly defined cash ROI, not a universal productivity score. A zero denominator makes that ratio undefined. A positive time saving with negative incremental cash contribution should be shown as such, rather than hidden inside a monetized time estimate. Danish research linking adoption surveys to administrative records found null earnings and recorded-hours effects within its studied population, despite changes in work practices, reinforcing the need to distinguish the outcomes. ^[41]

Compare like-for-like deliverables

Table 2 contains Hypothetical Examples, not collected freelancer observations or Montreal averages. Every amount is CAD, excludes collected sales taxes, and assumes the stated scope was accepted and receipts collected. Calculations use the formulas above; inputs are deliberately invented to isolate contract and rework effects.

The fixed-fee example improves contribution per hour but reduces cash contribution. The hourly example loses both cash and contribution per hour despite saving time. The extra-rework example reverses the apparent drafting gain. These are arithmetic outcomes of the stated assumptions, not estimates of how often any result occurs. They show why a rate-floor calculator and a workflow audit answer different questions: the publisher's rate guide at 2727coworking.com addresses funding non-billable time and operating costs.

In the fixed-fee Hypothetical Example, an assumed opportunity rate of CAD$60 per hour produces baseline economic surplus of CAD$380 and AI economic surplus of CAD$580, an increase of CAD$200. That is a valuation of owner time, not additional receipts. Someone prioritizing cash this month can reasonably interpret the same assignment differently from someone prioritizing free time.

A sensitivity check should vary rework hours, subscription allocation and receipts independently. Change one assumption at a time and retain the original result. A large gain that disappears after a small correction-time increase is fragile. NetLift's rework calculator similarly makes correction minutes and hourly labour cost explicit, but its labour-cost estimate does not establish a freelancer's collected revenue or sold capacity. (Source: netlift.app) (Source: netlift.app)

Analysis of Key Segments

Developers: accepted functionality includes checking

For development work, define the acceptance package: functioning requirements, integration, tests, documentation and handover. Record time spent reading unfamiliar code and investigating generated changes. GitHub recommends review, testing and validation of generated suggestions; those activities belong inside the measured workflow. ^[24]

METR's original experiment involved comparable submitted pull-request quality and a familiar mature-repository setting. Its result concerns the time required to reach that standard, not simply typing speed. ^[42] The experiment's 19% longer completion time is evidence that checking and context can matter, while the accompanying 16-developer, 246-task scope limits extrapolation. ^{[3] [4]}

A developer's baseline should therefore match repository familiarity and acceptance burden. Separate a scaffold that a client accepts as a prototype from production work requiring integration and review. If an agent tackles several issues concurrently, record human attention and accepted output for the whole batch. A stopwatch attached to one run can miss attention spent elsewhere. METR's later design update explicitly raises concurrent use as a measurement difficulty. ^[25]

Bilingual writers: verification belongs in delivery time

For English and French writing, specify audience, terminology, source accuracy, register and cross-language consistency. Canada's Translation Bureau identifies professional revision of AI translations as a way to limit errors. ^[43] The Office québécois de la langue française (OQLF) also publishes AI vocabulary intended to support workplace francisation; terminology review can use an authoritative reference rather than an unchecked generated glossary. ^[44]

Do not convert an experimental writing speedup directly into client profit. MIT's account of the published writing study describes tasks that did not require precise factual accuracy or company and customer context. Those exclusions matter when the paid mandate requires source verification and client-specific bilingual copy. ^[45]

Older translation research is relevant only as a caution about measurement. A pilot involving authentic English-to-French jobs found lexical differences between post-edited machine and human translations. It studied a particular translation setting, not contemporary Quebec copywriting or a current large language model. An audit should check the agreed language standard instead of assuming linguistic equivalence from fluent output. ^[46]

Designers and consultants: define quality before scoring speed

Design acceptance may include brand consistency, originality, technical file requirements and suitability for the intended channel. One advertisement-design experiment found no significant differences in visual appeal, brand alignment or usefulness despite evaluator-rated creativity differences. It does not establish a commercial acceptance or income advantage. ^[47] A separate visual-ideation experiment found greater fixation on an initial example, illustrating that idea quality and generation volume need separate measurements. ^[48]

Consultants should distinguish producing an analysis from verifying its recommendation. The published Jagged Frontier experiment found benefits on tasks within the tested system's capabilities, but lower correctness on its deliberately difficult task outside that frontier. This is a task-selection limitation, not a current benchmark for every consultant. ^[49]

Across these segments, retain:

• Quality evidence: the agreed checklist and client decision.
• Verification evidence: tests, reviewed sources or documented comparisons.
• Revision evidence: what changed and why.
• Financial evidence: receipts, costs and credits.
• Boundary evidence: work excluded from the mandate or the measurement.

Freed Capacity and the Keep, Change or Stop Decision

Record what happens after time is freed

Saved time can become leisure, reduced fatigue, unpaid sales work, learning or another paid assignment. Each can be valuable, but only the last produces additional project receipts. METR's survey explicitly distinguishes speed from work value and warns that doing extra tasks can inflate perceived value. ^[50] The audit adds a further distinction: value is not necessarily collected income.

Use q, the fraction of freed hours actually sold, only for a clearly labelled capacity scenario. Let r be measured contribution per hour on the additional work. Proposed additional contribution equals q × freed hours × r, when freed hours are positive. Keep that model beside the actual ledger, and never add revenue already included in collected receipts a second time.

In a Hypothetical Example extending the fixed-fee case, selling half the four freed hours at an assumed contribution rate of CAD$100 per hour adds CAD$200. Combined contribution becomes CAD$1,260, compared with the CAD$1,100 baseline, for an increase of CAD$160. With no additional sale, the original cash change remains negative CAD$40. These figures are author calculations from hypothetical inputs, not a demand forecast.

An actual additional project needs its own acceptance, receipt and cost records. If prospecting consumes the freed time without a sale, record sales activity and zero additional collected contribution. Do not impute the freelancer's advertised rate to unsold hours. Official recordkeeping guidance requires income amounts and sources, which provides a practical boundary between observed money and a scenario. ^[22]

Apply a decision rule suited to the objective

Before examining the result, state whether the objective is more monthly cash, better contribution per owner hour, less time working, or higher accepted quality. Then use the same objective across comparable assignments.

• Keep: accepted quality is maintained and the chosen measured outcome improves.
• Change: gains depend on excessive checking, fragile assumptions or inconsistent scope.
• Stop: repeated comparable observations miss the chosen objective after paid costs.
• Investigate: acceptance, baseline or receipt evidence is incomplete.
• Retest: the model, tool configuration, client requirements or workflow changes.

Treat these as management rules proposed by this report. A small self-selected project log cannot establish a causal effect. METR's follow-up shows how participant and task selection can change measured results even in a planned study. ^[51]

Retain unsuccessful attempts and unfinished assignments in the log. Compare whole observation periods as well as accepted projects, because acceptance-only analysis can hide paid subscriptions and effort spent on work never delivered. Revenu Québec's broader accounting guidance stresses careful records across the fiscal period. ^[52]

Data Analysis and Evidence

Table 3 separates study populations and outcomes. None supplies a directly measured Quebec solo-freelancer profit margin after subscriptions, verification and sold-capacity constraints.

The table's differences are substantive. A randomized time estimate, a self-reported value multiplier, an employee output rate and a platform earnings comparison do not measure the same thing. Combining their percentages into a single expected ROI would mix incompatible denominators. The accepted-work ledger makes the decision's denominator explicit.

Version discipline matters too. This report uses the published customer-support paper's 5,172-agent, 15% figures and the published consulting paper's 32% average quality improvement. ^{[58] [49]} Earlier working-paper summaries should not be blended with those versions. These are historical task results, even where the publication date is recent.

Upwork also reports that generative AI and creative-production contract starts rose 90% year over year while earnings per contract declined 13%. ^[62] Activity volume and earnings per assignment can therefore move differently in the platform data. Neither metric deducts an individual freelancer's verification time or subscriptions. The researchers' own CEPR account of earlier platform changes similarly distinguishes monthly jobs from monthly earnings. Those market-level outcomes should be kept separate from within-project delivery efficiency. ^[63]

Canadian business evidence needs a population check. Statistics Canada's Canadian Survey on Business Conditions (CSBC) methodology for the second quarter of 2025 describes a cross-sectional, voluntary survey targeting establishments with employees. That population differs from a solo freelancer without employees. ^{[64] [65]} Its AI analysis concerns employee tasks, employment and operational practices, rather than accepted-deliverable contribution. ^[66]

The conservative finding is a measurement requirement: public research provides plausible mechanisms and bounded results, while the freelancer's own comparable records determine the keep-or-stop decision. The reviewed evidence does not provide a transferable Montreal profit percentage.

Implications and Future Directions

A useful audit should remain reproducible after the first appealing result. Store the acceptance criteria, raw time entries, invoice references, allocation rule and calculation version together. Keep estimated baseline hours distinct from observed hours. Revenu Québec requires readable electronic records, and CRA generally specifies a minimum six-year retention period for business records, measured from the end of the last tax year to which they relate. ^{[38] [67]} Use professional advice for the precise records and retention obligations applying to the business.

Re-run the comparison when a material input changes:

• Tool changes: model version, billing meter or subscription tier.
• Client changes: acceptance standard, feedback process or required language.
• Scope changes: complexity, integration burden or revision allowance.
• Commercial changes: fixed fee, hourly terms or collection timing.
• Capacity changes: demonstrated demand for additional accepted work.
• Cost changes: actual foreign charges, processing fees or idle subscriptions.

A billing change need not alter quality, and a quality change need not alter subscription cost. Reconcile them separately. Anthropic documents an asynchronous Batch API discount of 50% on input and output tokens, but that discount says nothing about the human time or turnaround suitability of a particular assignment. ^[68] GitHub's changing flex allotments provide another reason to retain the actual plan snapshot used in an audit. ^[31]

Review results at the project and observation-period levels. Annual operating costs still need funding even if the project worksheet shows strong contribution per hour. The publisher's rate-floor guide at 2727coworking.com addresses this broader pricing requirement. Its physical-workspace offerings, including hot and dedicated desks documented in its virtual tour, are an overhead category to evaluate separately, not evidence that an AI workflow earns more. ^[69]

Commercial permissions, client confidentiality and intellectual-property terms remain contract questions. Verify the applicable provider terms and client agreement for the actual work; consult a lawyer where needed. This report does not assert a commercial licence or confidentiality guarantee. Likewise, have an accountant assess tax treatment rather than treating a positive management result as proof of deductibility. CRA's guidance differentiates current, capital and business-only expenses. ^{[39] [40]}

Future shared datasets would be more useful if they published accepted scope, tool generation, human review time, receipts and allocation rules together. Until then, a transparent small log is decision evidence for that freelancer, with acknowledged selection limits, rather than a population benchmark.

Conclusion

AI profitability for freelancers is a question about accepted work and retained money. A useful audit begins with comparable scope, records all hours worked by the owner, and reconciles paid costs to the actual billing period and currency. It then reports cash contribution, contribution per owner hour and net time difference separately. Opportunity value and prospective capacity sales can help a decision, provided they remain clearly labelled scenarios.

The worked examples show why there is no automatic conversion from drafting speed to profit. Fixed-fee work can improve contribution per hour while reducing project cash contribution. Hourly work can release time while reducing receipts. Additional checking can erase an apparent production gain. Extra capacity changes cash outcomes only when it becomes additional accepted, paid work.

The evidence table supplies context for those mechanisms, not a Montreal forecast. Its experiments, surveys, administrative records and platform data address different populations and outcomes. Their percentages should not substitute for the reader's ledger.

The practical rule is to choose an objective, measure comparable completed work, retain unsuccessful attempts, and repeat the audit when inputs change. Keep a workflow that reliably improves the chosen outcome at the required quality; change one whose benefit depends on avoidable rework or unstable assumptions; stop one that repeatedly fails the test. Where evidence is incomplete, record that uncertainty and gather the missing acceptance, time or financial record before drawing a confident conclusion.