Should I hire an agency or a one-person studio for an AI build?

Count two things: how many workstreams genuinely have to run at the same time, and what happens to you if the person building it disappears. Three or more parallel tracks, or a procurement process that requires a vendor with staff, insurance and an attestation, means hire the agency — one person cannot compress calendar time, or answer a security questionnaire written for a company with a security team. One coherent product whose hard part is judgement rather than volume points the other way: nothing is lost between the person who understood the problem and the person who writes the code. The rest is detail.

Four situations where an agency is the right answer for an AI build, and tooling does not change them. Parallel workstreams. An iOS app, a web app, a data migration and a brand system in one quarter is four calendars, not four tasks; one person works sequentially however much they automate. Continuity. Someone else can pick up the file when the lead is ill or gone, and for anything operationally load-bearing that cover is worth paying for. Procurement and compliance. Many buyers cannot sign with a sole trader: vendor questionnaires that assume a security function, SOC 2 or ISO 27001, indemnity at a set level, background-checked staff. If your legal team has those rules, arguing costs more than the premium. Specialist depth. Accessibility audits, penetration testing, a hard graphics problem, a warehouse migration at scale. Generalists plateau, and an honest one tells you where.

The advantage of a one-person studio is not price; it is that there is no translation layer. In an agency the person who sat in the discovery call is often not the person who writes the code — the requirement passes through an account manager, a producer and a ticket, each hop losing some of what you meant. In a solo engagement the person who watched a mechanic swear at a tablet in a cold bay writes the parser; a correction takes a sentence, not a change request.

Three things follow. You get the decision-maker on the phone. You pay for build hours rather than status decks. And scope stays the size of the problem, because a studio with no bench to keep busy has no reason to propose a nine-month platform where a six-week feature would do. The quiet one is coherence: one head holding the whole design is how a product ends up feeling like one product.

A one-person studio fails two ways. Bus factor: one illness, one better offer, and the project stops — and if the system is already in production, stops is not the worst word available. Capacity ceiling: a solo studio absorbs a surge for a fortnight, not a quarter, so when your timeline compresses there is nothing to add. Both are manageable in the contract rather than in conversation: source in your repository and cloud accounts from the first commit, conventional technology anyone could pick up, nothing running only on infrastructure the studio controls, and a named backup you may call.

An agency fails two other ways. Staffing: the people in the pitch are the senior people, while the people on the work are often whoever was between projects, which can mean juniors learning on your budget. That is how a utilisation model stays solvent rather than fraud; the counter is to name individuals and their weekly share in the statement of work. Margin: an organisation with people to keep busy pulls towards bigger engagements, more phases, and a change process in which every clarification is billable, so fixed-scope phases are the defence. Churn takes the rest — continuity as an org chart is not continuity in the people who know your business.

A bench of AI agents doing real jobs — research, first drafts, test writing, review passes, QA sweeps — compresses volume work considerably. It does not compress the work that decides whether a product is any good: sitting with the user, choosing what not to build, noticing that the workflow is wrong rather than the wording, and carrying responsibility for a wrong output. One person with agents is not five people. It is one person doing much less typing, with more of the week left for judgement.

The comparison between an agency and a one-person studio is therefore not day rate against blended rate, but total cost against the outcome you need, over the period you need it supported. A cheaper engagement that produces something nobody can maintain is the expensive option. Work at Max Motif is quoted per engagement rather than from a rate card; ask for a number on a specific piece of work.

Before choosing a vendor, check that a build is the right shape at all. Three cases recur. The tool already exists: an established product in your category may do most of what you described, and configuring it beats commissioning it. The problem is a process problem: if two people duplicate data entry because nobody agreed who owns the record, software will encode the confusion and run it faster. Or the scope is a fortnight, where a contracted developer on a defined task beats any studio relationship. A useful filter: can you describe the change you want in one sentence without naming a technology? If not, no vendor of any size fixes that, and both an agency and a solo studio will bill you happily while you work it out.

How many tracks must run at once — one, or several? What breaks if the work pauses for three weeks: an inconvenience, or an operation? Does procurement set a floor on vendor size, insurance or attestation? Is the hard part volume, or judgement about a domain most builders do not know? Who maintains it in two years? Any procurement floor, or several tracks at once, and the agency is the safer buy regardless of anything else. One track, judgement-heavy, in a domain you can teach one person deeply, is the case for a small studio — the case this studio is built around.

Read all of this knowing it was published by a one-person studio, and discount it accordingly. The test that survives the bias: ask both candidates what they would do if they were not hired at all. Whoever can describe a cheaper answer than themselves is worth briefing properly.

Last updated 25 July 2026.