{"slug": "the-demo-is-the-cheapest-part-scoping-ai-work-so-it-does-not-lose-money", "title": "The demo is the cheapest part: scoping AI work so it does not lose money", "summary": "Agencies lose $1,000 to $5,000 monthly to unbilled scope creep on AI projects, with 57% reporting such losses and only 1% billing for all work, according to a 2025 Ignition survey. Research shows developers using AI tools are 19% slower despite believing they are 20% faster, and code complexity increases 41% after AI adoption. The article advises scoping ambiguity out before signing, charging for discovery, using milestone-based payments, and requiring change orders to protect margins.", "body_md": "The Ledger · Freelancing\n\n# The demo is the cheapest part: scoping AI work so it does not lose money\n\n- Agencies losing $1k-5k a month to unbilled scope creep\n- 57%, with another 30% losing more; only 1% bill for all of it (Ignition, 273 agencies, 2025)\n- Developer speed with AI tools, real vs felt\n- 19% slower, while believing they were 20% faster (METR, 2025)\n- Projects that hit scope creep\n- 52%, up from 43% five years earlier (PMI Pulse of the Profession, 2018)\n- Code complexity after a team adopts AI coding\n- +41% complexity and +30% warnings; the velocity spike fades within weeks (He et al., MSR 2026)\n- Typical digital agency net margin\n- ~13% after tax, so one cost overrun can erase a client-year of profit (Promethean Research, 2025)\n- Common deposit before work starts\n- 20-50% up front, 50/50 a frequent default (Freelancers Union convention)\n\nA vibe-coded prototype wows the client in an afternoon and hides the 80% that decides whether the project earns: eval, guardrails, integration, and maintenance. Pricing sets the number. This is the scoping and contracting discipline that protects it.\n\nThe demo took an afternoon. You wired an LLM to the client’s data, it answered three questions correctly on the call, the room lit up, and you quoted the project off that feeling. That is the moment the money starts leaking, because the demo is the cheapest part of an AI build and the most misleading part too.\n\nThe gap between how the demo feels and how the work goes is not in your head, it is measured. When the research group METR had experienced developers do real tasks with AI tools, the developers expected to go faster, and afterward [believed AI had made them about 20% faster](https://arxiv.org/abs/2507.09089). The stopwatch said they were [19% slower](https://arxiv.org/abs/2507.09089). The optimism the demo creates is the same optimism you priced the job on. Underneath the working prototype sits the part that decides whether the project earns: evaluation, edge cases, hallucination guardrails, integration with the client’s real systems, and the maintenance that begins the day you ship. The demo is the tip of the iceberg. You quoted the tip.\n\nThis is not a pricing problem. [What to charge](/ledger/pricing-ai-work/) is its own solved question with its own piece. This is the layer after the number: scoping and contracting the job so the price you set survives contact with the work. Get it wrong and the receipts are brutal. In a 2025 survey of 273 agencies, [57% said they lose between $1,000 and $5,000 every month](https://www.ignitionapp.com/news/2025-agency-pricing-cashflow-report) to work they did but never billed, another 30% lose more than that, and [just 1% manage to bill for all of it](https://www.thedrum.com/news/2025/05/22/cash-flow-crunch-us-agencies-struggle-grow-late-payments-and-scope-creep-bite). You do not have to be careless to leak money this way. You have to be slightly optimistic, once, about the scope.\n\n## The short version\n\nThe fix is not a sharper estimate, it is a scoping and contracting discipline. If you sell AI builds:\n\n**Scope the ambiguity out before you sign.**“Build an AI chatbot” hides a five-times price range. Turn it into a written spec with named integrations, an evaluation bar, and explicit out-of-scope lines.**Charge for the scoping itself.** A paid discovery phase is the cheapest insurance you can buy, and it filters out the clients who will not pay for the 80% under the waterline.**Treat fixed-bid as a risk decision, not a default.** The demo makes AI work genuinely hard to estimate, and a fixed number on a job you cannot estimate is a bet against yourself. Stage it into milestones instead.**Take a deposit and tie the rest to deliverables.** Work starts when money lands, and each payment is released by something the client can verify, not by the calendar.**Make the next request a change order, not a favor.**“While you are in there, can you also” is where the margin dies. A short clause turns it into a billable line instead of a free one.**Price the maintenance, because the demo never shows it.** The build is the down payment. The upkeep is the mortgage, and it is in the client’s name only if you wrote it that way.\n\nEverything after this is the evidence.\n\n## You did not underprice it, you under-scoped it\n\n“Build an AI chatbot” is not a project, it is a category. The same sentence covers a weekend FAQ bot answering from a help doc and a context-aware agent that reads the CRM, books actions, and has to be right in front of customers, and the gap between those two is easily five times the price. When the client said one sentence, you heard a smaller project than they did, and you quoted yours.\n\nScope creep is not the rare bad-luck outcome, it is the base rate. PMI’s 2018 survey found [52% of projects experienced scope creep](https://www.pmi.org/-/media/pmi/documents/public/pdf/learning/thought-leadership/pulse/pulse-of-the-profession-2018.pdf), up from 43% five years earlier, and that nearly half finished late. The Standish Group’s long-running software data lands in the same neighborhood, with only [about 29% of projects](https://www.standishgroup.com/) coming in on time and on budget. Treat that figure as a ballpark rather than gospel, its underlying data is proprietary and its definitions have been contested in the research literature, but the direction is not controversial: a project that lands on its original scope is the exception. A quoting process that assumes you are the exception is how you join the 57%.\n\n## The maintenance the demo never shows\n\nThe most expensive thing the demo hides is what happens to the code after you ship it. In a peer-reviewed study of 807 open-source projects that adopted the AI editor Cursor, matched against 1,380 similar projects that did not, the velocity gain was real but temporary: lines of code added jumped 281% in the first month, then faded within weeks. What did not fade was the drag. Static-analysis warnings rose about [30%, and code complexity about 41%](https://arxiv.org/abs/2511.04427) in the period after adoption. A separate vendor analysis of 470 pull requests found AI-authored ones carried [roughly 1.7 times more issues](https://www.coderabbit.ai/blog/state-of-ai-vs-human-code-generation-report) than human-only ones, though that one is best read as directional, since the company sells AI code review and inferred which code was AI-written rather than confirming it.\n\nThe point for your quote is simple. The prototype that closed the deal is the fastest, cheapest code you will write on the project, and it is also the code that costs the most to keep working. If your number covered the build and not the upkeep, you priced a down payment and signed up for the mortgage. Where AI coding genuinely pays and where it bills you back is [its own piece](/study/does-ai-coding-make-you-faster/); for quoting, just assume the maintenance line is real and put it in.\n\n## What leaking it actually costs\n\nRun the leak through a real margin and it stops looking like a rounding error. A typical creative or marketing agency clears a net margin of [about 13%](https://prometheanresearch.com/how-profitable-are-digital-agencies/) after tax, which means roughly 87 cents of every dollar billed is already spoken for. At that margin, a 25 to 30% cost overrun on one engagement does not dent the profit on that client, it erases it, and then some. The afternoon you spent unbilled “just tightening the prompts” was not free time. It was margin, and on a thin-margin services business it is the only margin you have.\n\n## Scope the ambiguity out before you sign\n\nThe defense starts before the contract, in a deliverable spec the client signs. Take the one-sentence ask and write down what done looks like: the exact integrations by name, the evaluation bar the system has to clear (and who judges it), the data you will and will not touch, and a short, explicit list of what is out of scope for this price. The out-of-scope list is the most valuable part, because it is the document you point at, calmly, when the new request arrives.\n\nThis costs you time, so charge for it. A paid discovery or scoping phase does two jobs: it pays for the work of figuring out the real project, and it tells you instantly whether this client understands that the demo was the tip. A client who balks at paying to scope the work is a client who will balk at paying for the 80% under the waterline. Better to learn that for a small fee than on a fixed bid.\n\n## Why fixed-bid is the dangerous default for AI work\n\nFixed-bid is comforting to a client and a quiet trap for you, and AI work makes it worse. A fixed price is only safe when you can estimate the job, and the whole problem here is that the demo makes AI work look estimable when it is not. Fixed versus hourly is a pricing decision the [pricing piece](/ledger/pricing-ai-work/) covers; the scoping point is narrower. When you genuinely cannot estimate the hard parts, do not bury that uncertainty inside one number you have to honor. Stage the work into milestones, price each milestone as you reach it, and let the eval phase tell you what the build phase should cost. That is not indecision, it is refusing to bet the project’s margin on the demo’s optimism.\n\n## Stage the money so scope creep becomes billable\n\nTwo clauses do most of the protecting. The first is the deposit and milestone schedule. By freelance convention, not survey law, work does not start until a deposit lands, with the [common range running 20 to 50%](https://blog.freelancersunion.org/2018/10/11/4-key-questions-freelancers-should-ask-about-payment/) of the project and a 50% up front, 50% on delivery split as a frequent default. Bigger or vaguer jobs get broken further, with each payment tied to a deliverable the client can check rather than to a date. Staged money keeps you from financing the client’s project out of your own cash flow.\n\nThe second is the change order. Every project grows, and the difference between a profitable shop and the 57% is not whether scope changed, it is whether the change got billed. A change-order clause says, in plain terms, that work outside the signed spec is quoted and approved before it happens. It does not have to be adversarial. It reframes “can you also” from a favor you owe into a small new line the client chooses to buy, which is exactly what it is.\n\nYou did not underprice the project. You under-scoped it, and the demo is why. The number you set is only ever as good as the scope that defines it and the contract that protects it, and on AI work, where the prototype lies the loudest, that protection is the whole job. Quote the iceberg, not the tip.\n\nOne email, when there's something worth sending\n\n## Get the receipts in your inbox.\n\nNo fixed schedule, no filler. You get an email when we've tested something, run the numbers, or found a tool worth your time.\n\nFree. Double opt-in, unsubscribe in one click.\n\nScoped one of these the hard way? [Compare notes in the forum ↗](https://community.okaneland.com)\n\n## Sources\n\n| Source | Link |\n|---|---|\n| METR (2025), Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity |\n|\n\n[ignitionapp.com ↗](https://www.ignitionapp.com/news/2025-agency-pricing-cashflow-report)[thedrum.com ↗](https://www.thedrum.com/news/2025/05/22/cash-flow-crunch-us-agencies-struggle-grow-late-payments-and-scope-creep-bite)[pmi.org ↗](https://www.pmi.org/-/media/pmi/documents/public/pdf/learning/thought-leadership/pulse/pulse-of-the-profession-2018.pdf)[standishgroup.com ↗](https://www.standishgroup.com/)[arxiv.org ↗](https://arxiv.org/abs/2511.04427)[coderabbit.ai ↗](https://www.coderabbit.ai/blog/state-of-ai-vs-human-code-generation-report)[blog.freelancersunion.org ↗](https://blog.freelancersunion.org/2018/10/11/4-key-questions-freelancers-should-ask-about-payment/)[prometheanresearch.com ↗](https://prometheanresearch.com/how-profitable-are-digital-agencies/)Show your work: if a number can't be shown, we don't print it.", "url": "https://wpnews.pro/news/the-demo-is-the-cheapest-part-scoping-ai-work-so-it-does-not-lose-money", "canonical_source": "https://okaneland.com/ledger/scoping-ai-projects/", "published_at": "2026-07-04 13:00:00+00:00", "updated_at": "2026-07-04 13:04:15.050896+00:00", "lang": "en", "topics": ["ai-tools", "ai-products", "ai-startups", "developer-tools", "ai-agents"], "entities": ["Ignition", "METR", "PMI", "Promethean Research", "Freelancers Union", "He et al.", "MSR"], "alternates": {"html": "https://wpnews.pro/news/the-demo-is-the-cheapest-part-scoping-ai-work-so-it-does-not-lose-money", "markdown": "https://wpnews.pro/news/the-demo-is-the-cheapest-part-scoping-ai-work-so-it-does-not-lose-money.md", "text": "https://wpnews.pro/news/the-demo-is-the-cheapest-part-scoping-ai-work-so-it-does-not-lose-money.txt", "jsonld": "https://wpnews.pro/news/the-demo-is-the-cheapest-part-scoping-ai-work-so-it-does-not-lose-money.jsonld"}}