AI automation agency, in-house hire or a vetted contract engineer: which fits a small business?

What are the four roles that build AI agents and automations?

Four different people get called "the AI person", and they are not interchangeable. The no-code configurer builds in Zapier, Make or n8n: a trigger, a few filters, an LLM step that drafts or classifies, and a write back to your CRM or inbox. The automation engineer does the same job in code where the platforms run out: webhooks, custom API clients, a self-hosted n8n instance, retries and logs. The AI agent engineer builds LLM applications proper: tool calling, retrieval over your documents, evaluation sets that prove the agent behaves, guardrails and observability. The AI automation agency packages any of the above as a fixed-price or retainer project delivered by its own staff.

Rates differ by almost a factor of ten across these roles and regions. SecondTalent's cost guide puts US freelance AI agent developers at $95-142 per hour for junior, $119-185 for mid-level and $155-235 for senior, agency or dev-shop rates at $180-295, $230-380 and $295-490 for the same levels, and South Asian freelancers at $26-41 per hour, according to SecondTalent, updated 24 September 2026. Arc states that AI developers hired through its platform typically charge $60-100+ per hour, according to Arc.dev, 2026, and its rate guide puts AI and ML engineering at $110-190 per hour, per Arc.dev freelance developer cost guide, 2026. Hyparz publishes indicative hourly bands for AI engineers that track the rates Arc.dev publishes; the quote depends on specialism, seniority and engagement length.

What does an AI agent developer actually do?

An AI agent developer builds software in which a language model decides, within limits you set, what to do next: which tool to call, which record to read, whether to answer or escalate. A workflow fixes every step in advance; an agent does not. The job is mostly engineering around the model: defining the tools the agent may call, writing the prompts and schemas that keep outputs structured, building retrieval over your help centre or contracts, and writing evaluations so you see the failure rate before customers do. In a small-business engagement the responsibilities look like this.

• Scoping: turning "handle our inbound leads" into a written spec with inputs, allowed actions, hard limits and a definition of a correct outcome.
• Tools and integration: exposing your CRM, calendar, helpdesk or ERP to the model through typed functions or an MCP server, with the narrowest credentials that still work.
• Retrieval: indexing your documents so the agent answers from your material and cites where each answer came from.
• Evaluation: a test set of real cases, run after every change, reporting accuracy, escalation rate and cost per run.
• Guardrails and approvals: which actions run automatically, which need a human click, and how the agent fails safely when a tool errors.
• Deployment and handover: hosting, logging, token-spend tracking and a runbook so the system runs without the builder.

Agent developer or automation specialist?

If every step of the process can be drawn as a flowchart with no judgement calls, you need an automation specialist, not an agent developer, and you will pay less for a more predictable system. If the process needs reading, judgement or choosing between actions, you need an agent developer, and evaluation will be a third or more of the effort. The full responsibility list is in AI agent developer job description.

Agency: what you get, what it costs, where it goes wrong

An AI automation agency sells a result. You describe the process, they run discovery, quote a fixed price or a retainer, build on their stack and hand over a working workflow, often with a support plan. For an owner with no technical staff and no appetite to manage a contractor, that is genuinely attractive: one conversation, one invoice, one accountable party.

On price, DevCom states that small businesses typically spend $7,000-15,000 per workflow depending on integrations, model complexity and data preparation, according to DevCom, 2 March 2026. For hourly work, SecondTalent reports agency rates of $180-295 per hour for junior, $230-380 for mid-level and $295-490 for senior staff, per SecondTalent, September 2026, roughly double the freelance bands on the same page. The difference pays for the agency's sales, project management and margin.

Where agencies go wrong is rarely the build and almost always ownership. Ask early where the workflow will live. If the answer is the agency's n8n cloud workspace, the agency's Make organisation or the agency's OpenAI key, then the prompts, run history and connections belong to a company you can stop paying at any moment, and leaving means rebuilding. Retainers are the second trap: a monthly fee that buys availability rather than defined work, with an incentive to keep the system slightly dependent on them. The third is lock-in through a proprietary framework or a reseller margin on tokens, so the usage bill arrives through the agency and you never see what the model actually costs.

None of this makes agencies the wrong choice; it makes the contract the important part. Insist on these points in writing before the deposit.

• Every workspace, account and API key is created under your business email and billing, with the agency added as a collaborator you can remove.
• Prompts, workflow exports, evaluation sets and documentation are deliverables handed over at each milestone, not extras.
• Model usage is billed to your own provider account at list price, not re-invoiced through the agency.
• The fixed price states what "done" means in measurable terms: resolution rate on a test set, or records processed without manual correction.

In-house hire: when it pays and what it really costs

An in-house hire makes sense when agents are no longer a project but a product. The signals are concrete: agents touch revenue every day, the business changes them weekly as the process evolves, several workflows share the same data and tooling, and someone has to answer when a run fails at nine on a Monday. At that point the knowledge needs to live in a person on your payroll, with the context to make small changes quickly and the incentive to keep the system boring.

The cost is higher than the salary line. SecondTalent prices a mid-level in-house AI agent developer at the equivalent of $130-205 per hour and cites a published US employee median of $189,500 a year for the role, according to SecondTalent, September 2026. On top of base pay you carry benefits, equipment, recruiting and ramp time, and an engineer joining a business with no AI tooling spends the first weeks building evaluation and deployment basics before any workflow ships. The sourced picture for permanent roles is in cost to hire AI engineers.

The other cost is concentration: one engineer is a single point of failure, and one resignation leaves nobody who can read the code. For a sense of how far along most small businesses are, the Goldman Sachs 10,000 Small Businesses Voices survey of 1,256 owners found 76 percent currently use AI but only 14 percent say it is fully embedded in core operations, according to Goldman Sachs 10,000 Small Businesses Voices, 17 March 2026. If you are in the 76 percent but not the 14 percent, a full-time hire is usually premature; start on contract and convert once the work is clearly continuous, as set out in contract vs full-time AI engineers.

Vetted contract engineer: the middle path

The option agency pages leave out is contracting an individual engineer who has already been screened, works hourly inside your tools and accounts, and leaves you with everything when the scope is done. It combines the speed of freelance with the screening and continuity that open marketplaces lack, and it keeps ownership where an agency build often loses it: your n8n instance, your model provider account, your CRM credentials, your repository.

You scope by the week rather than by the project: week one, connect the CRM and ship the first workflow in shadow mode; week two, add the LLM step and the evaluation set; week three, turn on live actions with approvals. You see hours as they are logged and can stop, pause or extend without renegotiating a fixed-price contract. Because the engineer builds in your accounts, the question of who owns the prompts never arises.

On Hyparz this is the standard engagement. Engineers are pre-vetted, and about 5% of AI-role applicants pass vetting. You are matched typically within days, every engagement starts with a one-week risk-free trial, and you keep any work delivered during it. A dedicated account manager handles scope changes and re-matching, automated timesheets show you the hours, and source-code protection is part of the agreement. Specialisms cover workflow automation (n8n, Zapier, Make, Airflow), LLM applications, RAG and AI agents, so the same route works whether you need an n8n expert for a week or an AI agent developer for a quarter.

The honest limits: someone in the business still has to answer questions, test the first runs and approve what goes live, roughly an hour or two a week, and there is no fixed-price guarantee, so the scope document and the weekly check-in are your controls. For most owners, two hours a week is a fair price for keeping the keys.

How do timelines compare?

Published guides agree on the shape even where they disagree on the numbers: vendor platforms and no-code builds take weeks, custom agents take months. RaftLabs puts platform deployments at four to six weeks, a custom MVP at eight to twelve weeks and production-grade hardening at a further twelve to sixteen weeks, according to RaftLabs, 16 August 2026.

The builder moves that clock less than your data does. In one documented case from a HubSpot partner agency, a family-run manufacturer's customer agent went on to resolve 1,601 of 2,732 conversations in four months, but only after months of inbox and help-desk consolidation first, according to Hypha, 9 October 2026. If your help centre is thin or your CRM has three fields for the same thing, that cleanup is the project, whoever you hire. A realistic first workflow with a contract engineer or no-code specialist looks like this.

• Week 1: written scope, credentials issued with minimum permissions, a test set of 30 to 50 real cases from your inbox or CRM.
• Weeks 2-3: the workflow runs in shadow mode, producing drafts a person reviews; the engineer fixes what the reviews surface.
• Week 4: live actions for the cases the test set shows are safe, approvals on the rest, runbook and handover.
• Agency path: add discovery, proposal and contracting at the front and a handover milestone at the end. In-house path: add the recruiting cycle, which for AI roles routinely runs to months.

Can I hire someone to integrate agents with my existing software (CRM, ERP, helpdesk)?

Yes, and for a small business this is most of the work. The agent itself is a few hundred lines; the integration decides whether it is useful. It means read access to the right objects (contacts, deals, tickets, orders), field mapping so the agent writes where your staff do, authentication (OAuth apps, scoped API keys, token refresh), rate limits, and a choice between webhooks pushing events to the agent or the agent polling on a schedule. In 2026 the common pattern is to expose each system as a set of typed tools, increasingly through an MCP server, so the same CRM connector serves the next agent without a rebuild.

Costs vary with the systems involved. Adevs puts first-year AI agent integration for small businesses at $12,000 to $180,000, according to Adevs, 2026, and the width of that range is the point: a HubSpot or Zendesk instance with a clean API sits at the bottom, a legacy ERP with a nightly CSV export sits at the top. Cost drivers are detailed in AI agent development cost and the service in AI integration services. Before hiring for integration work, ask these questions; the answers tell you which of the four roles you need.

• Does each system have a documented API or a native connector in Zapier, Make or n8n? If yes to all, a configurer may be enough.
• Will the agent write to the system or only read? Writes need idempotency, approvals and an audit trail, which is engineering work.
• Is there a sandbox, or will the first run touch live customer records?
• Who owns the OAuth app or API key, and can it be scoped to exactly the objects the agent needs?
• What happens when a system is down mid-run, who is told, and what does the monthly maintenance budget cover?

How do you vet whoever you hire?

The method is the same for an agency, a freelancer or a candidate; only the pressure you can apply differs. You want evidence of a system that ran in production, a clear account of how they measure whether it works, and a look at how they handle your data and credentials. A screening platform does part of this (on Hyparz about 5% of AI-role applicants pass vetting), but run the short version yourself, because the platform cannot know your process. Detailed question sets are in AI agent developer interview questions and the full process, including a paid trial task, in how to vet AI engineers.

• Ask for one real workflow or agent they built, and what broke. A credible builder has a story about a tool that failed or a prompt that drifted, and what they changed. No story means no production experience.
• Ask how they measure it. You want "evaluation set", "resolution rate on held-out cases" or "cost per run", not "it works well".
• Ask where the credentials live. The right answer names your accounts, scoped permissions and a secrets store, never a shared document.
• Ask what they would not automate. Anyone who says "everything" has not run an agent on real customers.
• Run a paid one-week trial on your data with a defined output: one workflow in shadow mode, evaluation numbers and a short write-up of trade-offs. Hyparz makes that week risk-free and you keep the work delivered.
• Check the handover artefacts in the proposal: workflow export, prompts, test set, runbook, monitoring. If they are not listed, they are not coming. For agencies, ask to speak to the person who will actually do the work.

Which path fits your situation? A decision table

Find your row. Where two rows apply, the one with the stricter ownership or data requirement wins.

Red flags with any of the three

Most failed small-business agent projects were visible in the first conversation. One general warning first: marketed resolution rates and measured ones differ. A 60-day test of Intercom's Fin agent across four small businesses averaged 38 percent resolution against "up to 50 percent" in the vendor's marketing, according to Builts.ai, 2026, so treat any builder's headline number as a claim to test on your own cases.

• Agency: the workflow will live in their workspace; the proposal lists no handover artefacts; model usage is billed through them; the retainer has no defined scope; a fixed price is quoted before anyone has seen your data or a sample of real cases.
• In-house: the candidate talks models but not evaluation and has only demos with no users; or you are hiring because "we need an AI person" rather than for a specific, continuous workload, and the job description could be three different roles.
• Freelancer or contractor: asks for admin credentials to everything; cannot explain how to scope permissions; has no written scope and no weekly check-in; no screening behind them and no replacement if they disappear mid-build, which is the gap a vetted platform with an account manager closes.
• Any of the three: promises the agent will "replace" a role on day one; no budget line for maintenance, which one engineering estimate puts at 15-25 percent of build cost per year, according to Pykero on dev.to, 2026; no plan for what happens when the model is wrong.