Quick answer (updated August 2026): For a small business, buy. In MIT Project NANDA’s preliminary 2025 enterprise research, purchased or externally partnered AI tools reached deployment about twice as often as internal builds, roughly 67% versus 33% (self-reported). So: turn on the AI already inside software you pay for, buy where a proven category exists, assemble automations and custom GPTs for odd jobs, and build only when the workflow is truly proprietary and worth multiples of an illustrative year-one build cost of about $5,200.
Why every guide answers this question for an enterprise
Search “build vs buy AI” and every ranking page is written for a company with a data team: Scale AI’s enterprise guide (whose answer is to buy Scale), EY’s six-question framework, consultancy essays about proprietary data moats. In that world, “build” means machine-learning engineers and a proof-of-concept budget.
You run a 15-person company. For you, “build” means paying a freelancer or agency to write custom software that you then own, maintain, and debug. That is a different question with a different answer, and this page answers it for the owner-operated business between $500k and $20M in revenue. One disclosure before the framework: Hourback sells no software and takes no commissions, so nothing below pays us either way. If you want the broader lay of the land first, start with our owner’s field guide to AI for small business.
The evidence: buying wins, about two to one
The best available data comes from MIT Project NANDA’s preliminary 2025 report, The GenAI Divide (52 interviews, 153 leader surveys, and 300+ public AI initiatives, all enterprise-focused). In that sample, external partnerships with AI vendors reached deployment about 67% of the time, against about 33% for internally built tools, and employee usage rates were nearly double for external tools. The report is candid about its limits: the figures are self-reported, and “the correlation between external partnerships and success does not necessarily prove causation.”
Read the caveat, then read the setting. These were enterprises with engineering departments, and their builds still lost two to one. A 15-person company has less spare technical capacity than any firm in that sample, so the inference, and it is an inference, is that the gap runs wider for you, not narrower. The wider failure pattern is its own subject; we cover it in why small business AI projects fail.
The real bottleneck is integration, not access
A second dataset explains why the buy side wins. The Federal Reserve Banks’ 2025 Small Business Credit Survey (6,525 employer firms) found 46% of small businesses already use AI, but just 7% of those AI users have fully integrated it. Getting access to AI is now trivial; getting it wired into how the work actually flows is the part that fails, and it fails in every column of the build-vs-buy ledger.
That is the test to apply to any option in this article. A bought tool comes with the integration already designed: the receptionist answers your existing phone number, the scribe joins your existing calls. A built tool makes you the integrator, which is precisely the job the adoption-to-integration gap says small businesses do not have capacity for. When two options look similar on price, pick the one where someone else owns the wiring.
The decision at a glance
| The workflow | The answer |
|---|---|
| A feature your current software vendor already shipped | Turn it on; you are paying for it now |
| Phones, scheduling, intake notes, email triage | Buy: a proven category exists |
| Repeat questions answered from your own documents | Assemble: custom GPT or Claude project, hours of setup |
| Multi-step administrative work between two systems you already use | Assemble: Zapier or Make automation |
| Genuinely proprietary workflow with large, measured hours at stake | Run the threshold math below before building |
The rest of the article shows the work behind each row.
Step one: check the software you already pay for
Before buying anything new, inventory what you have. Your field-service platform, practice-management system, accounting software, and CRM have almost certainly shipped AI features since you last read their release notes: drafted replies, call summaries, invoice matching, scheduling suggestions. These features cost nothing extra on many plans, live inside a system your team already opens every day, and work on data that is already there, which removes the integration step where most AI projects die.
The check takes an hour. List every system you pay for, search each vendor’s site for “AI” plus the product name, and note what is included in your tier versus gated behind an upgrade. Turn on only what maps to a pain you can name. An unused AI feature you already own is harmless; a new subscription bought to duplicate it is pure waste.
Buy when a proven category exists
For the most common small-business pains, the market has already produced mature, competitive tool categories: AI phone receptionists, meeting and medical scribes, email triage tools, and scheduling assistants. Pricing in these categories is public and modest. AI receptionists, for example, run from $49 to $599 per month across published price pages (Rosie, Goodcall, Smith.ai, Slang.ai, as of July 2026), against human answering services (Smith.ai, Ruby) whose published tiers run $250 to $2,100 per month.
Buying here is not settling. A category leader has survived thousands of other businesses’ edge cases, ships updates weekly, and answers support tickets; a custom build gives you none of that at many times the price. The buying skill is sequencing, not selection: pick the pain with the fastest payback first. Our guide to what to automate first ranks the common candidates.
Assemble: the real middle path
Enterprise guides skip the option that fits small businesses best. Between buying a product and commissioning software sits assembly: wiring proven pieces together with no code to own. Two moves cover most cases. A Zapier or Make automation passes data between systems you already use, so a signed proposal creates the project, the folder, and the kickoff email by itself. A custom GPT or Claude project loaded with your own documents answers the questions your team answers by hand a hundred times a month: pricing rules, policy lookups, proposal boilerplate.
Assembly takes hours, not months, and costs almost nothing beyond subscriptions you may already carry. It handles exactly the odd-shaped workflows that make owners assume they need custom software. One gate applies: keep patient, client-confidential, and sensitive financial data out of consumer AI accounts. We walk through setup, use cases, and that data gate in our guide to custom GPTs and Claude for business.
Build only when both tests pass
Custom development earns its place in exactly one situation, and it requires passing two tests at once.
The proprietary test. No category exists because the workflow is genuinely particular to how your business makes money: a pricing engine built on twenty years of your own job history, a quoting method competitors cannot copy. “Our process is unique” usually fails this test on inspection; most unique processes are standard processes with local vocabulary. Scheduling, invoicing, follow-up, and intake all feel unique from inside the business, and none of them are.
The numbers test. The measured value at stake must be worth a multiple of the build’s full cost, maintenance included. The failure data argues for a hard threshold rather than optimism: Gartner predicted in 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by end of 2025, citing, among other factors, escalating costs and unclear business value, and S&P Global’s 451 Research found the share of companies abandoning most of their AI initiatives jumped from 17% to 42% in a year.
Pass one test but not the other, and the answer is assemble or buy.
The threshold math, done honestly
Here is the comparison most build-vs-buy articles never print. The figures are illustrations, not client results, with build costs rounded up on purpose. Labor uses Upwork’s published marketplace medians: $50 per hour for AI engineers and $100 per hour for machine-learning engineers.
| Year-one cost | Buy ($40–$200/mo tool) | Build (modest custom tool) |
|---|---|---|
| Up-front | $0 | 80 hrs at $50–$100/hr = $4,000–$8,000 |
| Ongoing | $480–$2,400/yr subscription | Maintenance at 2–4 hrs/mo = $1,200–$4,800/yr |
| Year-one total | $480–$2,400 | $5,200–$12,800 |
| Deployment rate in MIT NANDA’s sample | ~67% (external tools) | ~33% (internal builds) |
The last row is the one owners skip. A build that costs $5,200 and reaches deployment a third of the time needs the workflow to be worth far more than $5,200 to justify the bet. A working rule: do not commission custom software unless the workflow’s measured annual value is at least three times the build’s year-one total, and a proven tool covering 80% of the job does not exist at $200 a month.
Knowing your own numbers is the hard part, and it is what an AI opportunity assessment produces: a scored map of your workflows with the hours and math attached, for a flat $999. If we can’t save your business at least ten hours a week, your assessment is free.
Bottom line
For a small business the default is buy: purchased or externally partnered AI tools reached deployment about twice as often as internal builds in MIT NANDA’s 2025 research (~67% vs ~33%, self-reported). Order of operations: turn on the AI inside software you already pay for ($0), buy where a proven category exists ($40–$599/mo published pricing), assemble odd jobs with automations and custom GPTs (hours, near-zero cost), and build only when the workflow is genuinely proprietary and worth a multiple of an illustrative year-one build cost that starts around $5,200. The two-to-one odds are not a rule of nature, but they are the best evidence anyone has, and they point the same direction as the price tags.
Sources
- MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (preliminary report), 2025: report PDF
- Federal Reserve Banks, 2026 Report on Employer Firms (2025 Small Business Credit Survey), 2026: fedsmallbusiness.org
- Gartner, “Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025,” 2024: gartner.com
- S&P Global Market Intelligence, “Generative AI shows rapid growth but yields mixed results,” 2025: spglobal.com
- Upwork, AI engineer cost page and machine learning engineer cost page, fetched 2026: upwork.com, upwork.com
- Rosie, published pricing, as of July 2026: heyrosie.com
- Goodcall, published pricing, as of July 2026: goodcall.com
- Smith.ai, AI receptionist and human answering pricing, as of July 2026: smith.ai, smith.ai
- Slang.ai, published pricing, as of July 2026: slang.ai
- Ruby, published pricing, as of July 2026: ruby.com