Guide · AI & automation

Hiring an AI agent developer: what to ask, what it costs, and when you don't need an agent

Before you hire AI agent developers, check whether you need an agent at all. 2026 AI agent development cost ranges, 10 questions to ask, and the red flags.

By Akash Bansal, founding engineerUpdated 4 November 20267 min read

Do I need an AI agent, or just automation?

Usually you need a workflow, not an agent. Anthropic's engineering team draws the line clearly. Workflows are "systems where LLMs and tools are orchestrated through predefined code paths". Agents are systems where LLMs "dynamically direct their own processes and tool usage". Their advice is to find "the simplest solution possible", which "might mean not building agentic systems at all" (Anthropic, Building effective agents).

I agree from experience. Of the AI systems I've shipped, most are workflows:

• Aviate reads résumés from Gmail, scores them and surfaces the best ones. Fixed steps, one judgment call per résumé. A workflow.
• Brahmo grades long-form exam answers: Gemini OCR, then semantic matching, then Claude scoring on a weighted rubric, with a human expert override. A pipeline of fixed stages. A workflow. The exam prep app scenario shows that pipeline inside a product.
• NextMantra runs live voice interviews where the next question depends on what the candidate just said. That conversational loop is the most agent-like thing I've built, and even there the conversation serves a fixed goal: scoring the candidate against the role's rubric.

Use this table before you write a job post.

The cost of getting this wrong is real. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs, unclear business value or weak risk controls. It also warns about "agent washing", where existing chatbots and RPA are rebranded as agents (Gartner, June 2025).

How much does AI agent development cost in 2026?

Published ranges run from about $5,000 to well over $300,000, and nearly all of them come from firms selling the work. These are the ranges vendors publish (all vendor sources):

For a reference point from a solo builder: my AI feature or agent work starts at $8,000, and an MVP built around AI starts at $15,000, both as fixed quotes after scoping. That buys one focused system done properly, not an autonomous digital employee. To get a rough build range for your own AI parts before you talk to anyone, use the app cost calculator.

Running cost is the number most quotes leave out, and it scales with usage. A rough way to estimate it yourself: tokens per task × price per token × tasks per month. A résumé scored by a strong model at current Claude Sonnet 5.5 list prices ($2 input, $10 output per million tokens) costs about a cent, by my rough estimate of 3,000 input and 400 output tokens. A live voice conversation is billed per minute across audio, speech recognition, the LLM and speech synthesis, so it costs far more per task. Ask every candidate developer for this calculation for your own use case. The website chatbot cost breakdown runs the same arithmetic per conversation, at current model prices.

What should I ask an AI agent developer before hiring?

Ask questions that separate people who have shipped AI to real users from people who have built demos. Good answers are specific, and they include the word "evaluation".

1. "Does this need to be an agent?" A good developer will try to make it simpler. Be wary of anyone who answers "yes" before asking about your process.
2. "Show me something you've shipped that real users depend on." A live product beats a GitHub demo. Ask what broke after launch.
3. "How will we know it's working?" Look for an evaluation set: real examples with known right answers, run before every change.
4. "What happens when the model is wrong?" Good answers include confidence thresholds, a human review path, and logs of every decision.
5. "What will each task cost to run?" They should estimate tokens per task and show the arithmetic.
6. "Which model provider, and how hard is it to switch?" Prompts and evaluations should be portable. Lock-in should be a choice you make, not an accident.
7. "Where do my data and keys live?" In your accounts, from day one.
8. "How do you stop prompt injection and data leaks?" Untrusted input such as emails, documents and web pages must never be able to trigger privileged actions.
9. "What do I get at handover?" A repo, a design doc, the evaluation set, a runbook, and the ability to change a prompt without calling them.
10. "Fixed price or hourly, and what's in scope?" Either can work. Vague scope with hourly billing is how AI projects drift.

What are the red flags when hiring AI developers?

• Demo-only portfolio. A chatbot over a PDF is a weekend project. Ask for production systems.
• No mention of evaluation. Without it, every prompt change is a guess.
• "The AI will figure it out." Autonomy is a cost, not a feature. Every unconstrained step is a new failure mode.
• Keys in their account. If they hold the model API billing and the cloud account, they hold your product.
• Only a build price. A quote without a per-task running cost hides the number that scales.
• Framework first. Choosing a heavy agent framework before understanding the task usually adds layers you will later debug.

For the general version of this, beyond AI, read how to vet a developer as a non-technical founder.

Freelancer, agency, platform or founding engineer: who should build it?

The comparison of these routes beyond AI is in technical cofounder vs founding engineer vs agency.

When don't you need an AI agent developer at all?

You don't need one if the task has no judgment in it, if a feature in a tool you already pay for does the job, or if you haven't written down what a right answer looks like. That last case is the most common. If you can't write ten examples of correct and incorrect outputs, no developer can build a reliable system yet. Spend a week writing those examples. It is the cheapest AI work you will ever do, and it becomes the evaluation set any good developer will ask for.

Also hold off if your volume is tiny. Twenty documents a week may be faster to handle by hand than to automate, test and maintain.

How do I brief an AI agent developer?

Send a one-page brief before any call. It filters out people who only build demos, and it makes quotes comparable. Include five things:

1. The task today. Who does it, how often, how long it takes, and what it costs you when it's late or wrong.
2. Ten examples. Real inputs with the output you'd want, including at least three hard or ambiguous cases.
3. Where the data lives. Gmail, a CRM, PDFs, a database, and who controls access.
4. What the system may do on its own. Draft only, or also send, update or delete? Every action it can take alone needs a reason.
5. Your volume and budget. Tasks per month, and a build range. That lets a developer give you a running cost alongside the build price.

A developer who reads this and comes back with "this is a workflow, here's a smaller version" is usually the one to hire.

What does a good AI agent engagement look like?

1. Scoping (week 1). The process mapped step by step, the evaluation examples collected, a design doc, and a fixed quote with running costs per task.
2. Thin slice (week 2). One end-to-end path working on real data, measured against the evaluation set.
3. Hardening (weeks 3–4). Edge cases, retries, rate limits, human review, logging and alerts.
4. Handover. Everything in your accounts, documented, and changeable without the original developer.

Two worked examples, both composite scenarios, show the shape: automating résumé screening from Gmail (a workflow) and building an AI voice interviewer (closer to an agent). A third, AI automation for a dental practice, shows how to split the work between tools you buy and the gap you build.

If you've written down what "right" looks like and need one engineer to build, evaluate and ship the system, this is the work I do, from a fixed quote with weekly demos, and you own the code, the keys and the evaluation set. Here's how I work. If the honest answer is "you need a Zapier flow", I'll say so on the first call.