How do I scope and hire someone to build an AI agent for one process in my business?

By Ranjit Bhinge, founder of Blur Studio · Updated · 8 min read

Short answer: Pick one process, write down what goes in, what should come out and who checks it, and hire against that page, not against the words "AI agent". An AI agent is software that uses an AI model to decide the next step and take actions, such as reading an email, looking up an order and drafting a reply. Hire a person or studio that will price one process, hand over everything you own at the end, and run a test with real examples before it touches a customer.

Ranjit Bhinge

You have one repetitive process you want an AI agent to handle, and you want to hire the right person without paying for a vague promise.

I'm Ranjit, and I run Blur Studio: 15+ years building software, 120 apps taken live, and clients who have gone on to raise $10M+. I build AI systems and processes that help businesses do a lot more with a lot less time. I also work differently from most of what else is out there, more on that further down this page. Say hello.

Start with one process, not "an agent"

People ask for "an AI agent" the way they once asked for "an app". It sounds like one thing. It is really a process with an AI model inside it. The hiring mistakes I see come from buying the label instead of the process.

My plain definition: an agent reads something, decides what to do next, and does it, using the tools you give it. My explainer on what an AI agent is goes deeper. For hiring, the point is that the more freedom you give it, the more you must test and watch it.

Anthropic, the company that makes Claude, gives careful advice here in its guide to building agents. It says to find the simplest solution possible and only add complexity when needed, and that this "might mean not building agentic systems at all". It adds that for many applications, a single AI call with retrieval (looking up relevant documents) and examples in the prompt is usually enough, and that fixed workflows suit well-defined tasks while agents suit cases where flexibility is needed. If a good developer tells you that a fixed workflow will do the job better, that is a sign of a good developer, not a lazy one. For the difference in plain words, see AI workflow.

The one-page scope brief (fill this in first)

Before you contact anyone, answer these. A freelancer or studio can only give you a fair price once they can see the edges of the job.

  1. The process, in one sentence. "When a customer emails asking about order status, find the order and reply."
  2. Trigger. What starts it? A new email, a form, a time of day.
  3. Inputs. What does it read? Which inbox, spreadsheet, system, documents. How messy are they?
  4. Tools and systems. Which software must it connect to? Name each one.
  5. Output. A draft reply, a filled record, a message to staff? Where does it appear?
  6. Allowed actions. What may it do alone, and what must wait for a person? Sending money, deleting records and messaging customers should be in the second group at first.
  7. Who checks it. Name the person and how much time they have.
  8. Volume. How many times a day or week does this happen?
  9. Success test. "At least 9 of 10 sample cases handled correctly" is a test. "Works well" is not.
  10. Failure rule. What should it do when unsure? The right answer is almost always "hand it to a person".

If you cannot fill in questions 6, 7 or 9, you need to decide those first, before you hire. Start by choosing the process. My guide on AI implementation and the business process automation overview help with that.

Marketplace freelancer, independent developer or studio

There are three common places to hire. None is always right.

OptionGood forWatch for
Freelance marketplace (Upwork or Fiverr, sites where you buy services from independent sellers)A small, well-defined task with a clear checklistFixed packages that assume your process matches theirs; unclear ownership; no one to ask after delivery
Independent AI developerOne process, direct contact with the person who builds itOne person means a single point of failure; ask how they hand over
Studio or agencySeveral connected processes, ongoing changesLayers of staff; ask who actually writes and tests the work

My view: for one process in a small business, an independent developer or small studio who talks to you directly usually fits best. For a more complete picture of what to ask a company that sells this, see AI agent development companies and hiring AI developers. If you are not sure whether you need a builder or an adviser first, read about AI consultants.

Price versus scope: what changes the quote

Cost depends on scope, not on the word "agent". Quotes vary too much with the factors below for an average price to help. Ask for a fixed price tied to your written scope. My AI app cost guide explains what moves a quote.

ScopeWhat it usually involvesWhat pushes the price up
Drafts onlyReads input, writes a draft, a person approvesMessy inputs, many document types
Fills in recordsReads input and updates your system with a person reviewingEach extra system it connects to; poor access to your data
Acts on its ownTakes actions without approval for defined casesTesting, safety limits, logging, error handling
Several steps, several toolsChains the above across your systemsEverything above, multiplied

A worked example with made-up numbers, to show how to compare quotes. Say your process is "read order-status emails and draft a reply from the order system". Developer A quotes a fixed amount for drafts only, with a person approving every reply. Developer B quotes a higher amount for sending replies automatically. The difference is not the AI. It is the extra testing, limits and monitoring needed when no person checks. For a first project, I would take A, run it for a month, and only then decide whether B's version is worth it.

What a good quote includes

  • The scope page, repeated back to you in the builder's words.
  • A fixed price or a capped one, with what counts as a change.
  • The test plan and who runs it.
  • Running costs after launch: the AI provider's usage fees, hosting, any tool subscriptions. Ask for an estimate at your volume.
  • What happens when the AI provider changes a model or its price.

What you own at the end

This is where cheap gigs go wrong. Before you pay, get written answers to these:

  • The code and prompts. Are the instructions given to the AI, which are the heart of the agent, delivered to you in a readable form?
  • The accounts. Is everything set up in your name: the AI provider account, hosting, automation tools, the database? An agent that runs on the builder's account is rented, not owned.
  • The keys. Who holds the passwords and keys, and can you replace them? OWASP's guidance on managing secrets says to apply least privilege, so each key should open only what the agent needs.
  • The documentation. Is there a plain-English page saying how it works, what it can touch, and how to switch it off?
  • The licence. Does the contract give you ownership or a licence, and does it say anything about the builder reusing your code for others?

The contract points are worth a short call with a lawyer, as I am not one and this is not legal advice. My handoff guide on what a developer needs to take over an app is a good mirror image: it shows what a well-organised handover looks like from the other side.

Safety limits to demand

OWASP, a respected non-profit for software security, lists "excessive agency" as a top risk for AI systems. It means an AI is given too much functionality, too many permissions or too much autonomy, so a wrong or manipulated output can cause damage. Its advice includes limiting the tools an agent can call to the minimum necessary, and limiting what each tool can do. Its example is a tool that only needs to read email should not be able to delete or send. For the limits to put in place, see AI agent guardrails.

Turn that into questions for your builder:

  • What exactly can the agent read, change and send? Show me the list.
  • Which actions need a person's approval?
  • What happens if a customer's message tries to tell the agent to do something else? OWASP calls this prompt injection: input that changes the AI's behaviour in ways the builder did not intend. Ask how it is handled.
  • Is there a log of everything it did, and can I see it?
  • Is there an off switch I can use myself?

The US standards body NIST publishes an AI Risk Management Framework, a voluntary guide to managing AI risks. You do not need to read it, but a builder who has heard of it and talks in terms of risk, not hype, is a good sign.

A test plan before go-live

  1. Collect 30 real past cases from your own business, with the right answers. Include some odd and difficult ones.
  2. Run them through the agent and mark each as right, partly right or wrong.
  3. Read the wrong ones together with the builder. Fix the cause, not just the example.
  4. Run it in shadow mode for one to two weeks. It drafts, a person does the real work, and you compare.
  5. Go live with approval on. Loosen it only when the results hold up for a month.
  6. Review every month. Models, prices and your own business change.

Where a person must always check: anything involving money, legal commitments, health or safety, and the first messages to a new customer.

Red flags when hiring

  • They promise results (hours saved, revenue gained) before seeing your process.
  • They cannot show a past project in the same size of business, or will not explain how it was tested.
  • They want to run everything on their own accounts.
  • No test plan, no log, no off switch.
  • Vague scope with a fixed price, or a very specific price with a vague scope.
  • They suggest an agent where a simple rule or automation would do.

How I'd approach this

I would hire for one process, with a drafts-only first version, a written scope, everything in my own accounts and a test on 30 real cases. I would treat the second process as a separate job.

I would also ask each person who quotes: what would you leave out? A builder who can name what should stay manual, or what a plain automation can do without AI, has probably built these before. If you do not yet know which process to choose, the AI integration services guide shows how systems get connected, and my guide to building an AI agent shows what the work itself looks like.

This is the kind of project Blur Studio builds: one process at a time, scoped in writing, owned by you. You can see what a business build includes, or talk it through on a call.

Questions people ask

How do I scope an AI agent project?

Write one page covering the process in one sentence, the trigger, inputs, tools, output, allowed actions, who checks it, volume, a success test and what it should do when unsure. Hire against that page. For picking the first process, read AI implementation.

Should I hire a freelancer on a marketplace or a studio?

Marketplaces suit small, well-defined tasks. For one real business process, a direct relationship with an independent developer or small studio usually works better because you can ask questions after delivery. Check who owns what at the end in either case.

What should an AI agent project cost?

It depends on scope: how many systems it connects to, how messy the inputs are, and whether it acts without approval. Ask for a fixed price tied to a written scope, plus an estimate of running costs such as AI usage fees. My cost guide lists what moves a quote.

Who owns the AI agent after it is built?

That is decided by your contract, so get it in writing before you pay. Ask that the accounts, code, instructions to the AI and documentation are delivered in your name. Have a lawyer check the ownership and licence wording.

How do I know the agent is safe to run on my customers?

Test it on real past cases, run it in shadow mode with a person doing the real work, and start with approval switched on. OWASP advises limiting an agent's tools and permissions to the minimum necessary, so ask what it can read, change and send.

Do I need an agent, or is a simple automation enough?

If the steps are always the same, a fixed workflow is cheaper and more predictable. Anthropic's own guidance says to find the simplest solution possible, which might mean not building an agent at all. Read about business process automation first.

More in this topic: Problem solving: AI strategy and adoption (24 guides) →

Sources

Ready for the next step?

A 30 minute call. If you do not need me, I will tell you.

Ranjit Bhinge You talk to me, Ranjit. I reply within a few hours.

Not sure what you need? Tell me what is going on. I will point you the right way, even if it is not me.

Blur Studio · 30 minute call

Book a 30 minute call

You talk to me, Ranjit. If you do not need me, I will tell you.

  • We talk about your business and what is going on.
  • You leave with a clear next step, even if it is not me.
  • I reply to every message within a few hours.
Loading the calendar…