Clients Who Will Never Vibe Code — and What They Are Actually Buying
On September 10, 2026, @levelsio described the buyer most AI agencies will actually meet:
Normies don't vibe code, they just ask something like "do my bookkeeping" or "file my tax" or "organize a movie night and send invites" or "generate a flyer for movie night" or "edit my video"
They don't ever see code, vibe code, or do anything with code, their AI chat app just does it for them
Most of the software layer has already disappeared or will completely disappear for normies
—@levelsio, 18:48:37 EDT, September 10, 2026 (X status [1]; readable full text [2]).
"Do my bookkeeping" is a deadline, not a tool request
The thread argued about whether a non-technical buyer ever sees code — a live question if you sell agent work. Read it the other way: if the buyer never inspects the software, they were never buying software.
Two items are business work with a date attached: a filing deadline, a consequence if the numbers are wrong, and an owner who carries it — the client. The rest are errands, and none of the five asks for software.
Gusto's release of September 10, 2026 reports where freed capacity went: "New hires at the smallest AI-adopting businesses were mostly care providers, cooks, and technicians".
For the wider payroll picture behind that sentence — and where the effect stops — see our breakdown of the small business AI adoption and hiring data.
The vendor that used to sell them software was paid for three things beyond it: that the output was right, that it was wired in, and that someone owned it when it broke. Those three are the offer.
Correctness: the output has to survive a reconciliation
A chat app is graded on fluency. Bookkeeping is graded against a record.
The deliverable is not "an AI bookkeeper" but a run that reconciles against the bank feed and the ledger and reports every variance it cannot explain. Scope three things: a named reconciliation step before submission; a sample the client can spot-check; exceptions escalated to a human. Price it as a verification line with hours attached, not model work.
Connection: read and write, not a chat window
Licensed software arrived wired in, and its replacement must be: bank feed, books, payroll register, CRM — read to see true state, write to change it. A chat window that produces text the owner re-types is a draft, not an automation.
Scope it system by system: read or write, whose credentials, who owns the connection when it breaks. Price integrations per system and itemise permissions separately — the part buyers underestimate. What an AI automation agency actually does follows the same order.
Governance: a named human, a sign-off, and a clock
Governance is what makes the first two sellable. Four items, small enough for a statement of work:
- Permissions. Least-privilege access per system, issued to a named account, revocable without a rebuild.
- A sign-off. Nothing is filed, submitted or sent until a named human approves it — and the scope names that person.
- The trail. Retained evidence of what ran, on which data, what was checked and who approved it, kept for a stated period.
- A response time. "Call us" is not support; a stated number of hours with a named escalation path is.
Price governance as a monthly line — the maintenance contract on the old software.
What to refuse
- Refuse scope with no reconciliation. If the output cannot be tested against a record, there is no definition of done.
- Refuse write access without a sign-off point. An agent that can file or send unprompted is a liability nobody asked for.
- Refuse any promise without a response time, and any scope where "the software handles it" with no named owner. For pricing, see AI automation ROI for small business.
Who is on the hook when "file my tax" comes back wrong — and how would the client ever know?
Nobody in that thread asked this. It decides whether the rest is a service or theatre.
Who is on the hook: whoever filed it — by default, the client. "The AI did it" is not a remedy, so the failure path must be contractual: a named signer before submission, and a clause saying who corrects an error, at whose cost, and how fast. AI agency contract tips cover that language.
How the client would know: not from the output — fluent is fluent either way. They know because reconciliation runs against the system of record on a fixed cadence, and a variance escalates instead of being silently corrected.
If my bookkeeping comes back wrong, who is liable?
Whoever filed it, unless the agreement says otherwise — so ask which named person signs off, and which party pays to correct an error. If the answer is "the model", there is no accountability layer, only tooling.
How would I detect an error before the deadline bites?
By reconciliation, not by reading. Ask which record the output is checked against, how often, and what triggers an escalation. How to vet an AI agency turns that into a question list.
What do I need in writing before an agent gets write access?
A named human sign-off before anything is filed or sent, least-privilege credentials per system, a retained log, and a published response time for corrections. A provider who will not accept all four has offered no failure path — and the buyer asking for "do my tax" holds the consequence.
Sources
- @levelsio, "Normies don't vibe code…", X status, 10 September 2026, 18:48:37 EDT — https://x.com/levelsio/status/2098181881023844392
- @levelsio, "Normies don't vibe code", first-party full text (JSON-LD datePublished 2026-09-10, dateModified 2026-09-11) — https://levels.io/normies-dont-vibe-code
- Gusto via PR Newswire, "Small businesses that adopted AI are hiring faster, new Gusto research finds," 10 September 2026, 10:00 ET — https://www.prnewswire.com/news-releases/small-businesses-that-adopted-ai-are-hiring-faster-new-gusto-research-finds-302875404.html
Quoted spans attributed to [1] were matched against the fetched text of that status; the span at [3] is the vendor's own correlation, dated 2026-09-10, cited only for where hiring went.