# AI agent development services for companies that need agents in production

AI agent development is building software in which a language model plans steps, calls your systems and completes a task inside rules you set. Goldenroz builds custom AI agents and agentic workflows for funded startups and enterprises. Projects start at $25,000, fixed price per phase after a scoping call. A working prototype of a focused agent typically takes one to two weeks, and production roughly four to eight weeks depending on integrations.

**From $25,000, fixed price per phase**

Updated: 2026-11-13. By Shai Goldenberg.

## Who this is for

- Operations and support leaders with a high-volume process a person can explain step by step: order intake, ticket triage, quoting, reconciliation.
- Product teams adding an agent to an existing SaaS product or mobile app.
- CTOs who have a working demo and now need retries, cost limits, permissions and logs before it touches customers.
- Companies whose data sits in an ERP, CRM or SQL database and who want an agent to use it without giving the model open access.

## What is included

- **Scope and success measure:** One written page that names the task, the systems the agent touches, what it may never do, and the number that decides whether it worked.
- **Agent design:** A split between what the model decides and what plain code decides. Anything that can be a rule stays a rule.
- **Tools and system access:** Connections to your CRM, ERP, database and messaging channels through MCP servers or direct APIs. See [AI integration](https://goldenroz.com/ai-integration/).
- **Durable execution:** Agent runs that survive a crash or a timeout: leases so two workers never take the same job, retries, and a scheduled sweep that picks up stalled runs.
- **Model choice and fallback:** Claude, OpenAI or another model chosen per task, with a fallback chain and a fixed reply when every model call fails.
- **Permissions and human approval:** Per-user access, read-only by default, and an approval step before any action that costs money or cannot be undone.
- **Cost and audit logging:** Token and cost tracking per run, and a log of every tool call, input and result.
- **Handover:** The code lives in your repository from the first commit, with a runbook and an evaluation set you can rerun after every change.

## How the work runs

1. **Scoping call (Before the project):** We map the process, the systems involved and the cost of a wrong action. You get a fixed price for the first phase.
2. **Prototype (About 1 to 2 weeks):** A working agent on one narrow task, run against real examples in a test environment. You see where it fails before you pay for production.
3. **Production build (Roughly 4 to 8 weeks):** Integrations, permissions, fallbacks, approval steps, logging and the evaluation set. The range depends on how many systems the agent connects to.
4. **Launch with limits (At launch):** The agent starts on a share of the volume, or with every action approved by a person. Limits widen as the logs show it holds.

## Agentic AI development: what changes when the model takes actions

A chatbot returns text. An agent reads a request, decides which steps to take, calls tools and checks the result before it continues. Agentic AI development is the engineering around that loop: which tools exist, what each one is allowed to do, and what happens when a step fails halfway.

The model is the smallest part of the work. Most of the code in a production agent handles timeouts, duplicate requests, partial failures, rate limits and cost.

We keep agentic workflows narrow. One agent does one job with a short list of tools. When a process needs several jobs, we chain small agents with plain code between them, so each step can be tested and logged on its own. If the task is a conversation with customers, see [AI chatbot development](https://goldenroz.com/ai-chatbot-development/).

## Security and human approval are part of the build

An agent with database access is a new user in your system, and it reads text written by strangers. A customer message or a document can contain instructions aimed at the model. We treat every input as untrusted and limit what the agent can do no matter what the text says.

In practice that means read-only access by default, per-user permissions enforced in the database, approved queries in place of model-written SQL, masking of personal data before it reaches the model, and an audit log of every call. Actions that spend money, send messages to customers or change records go through an approval step until the logs justify removing it.

For systems that already run in production, a [security audit](https://goldenroz.com/security-audit/) checks what an agent or an anonymous user can reach today.

## Model choice: Claude, OpenAI and others

We choose the model per task. Claude is our default for agents that use tools and follow long instructions. OpenAI models and others are used where they fit the task, the cost target or your existing contracts.

The agent is built so the model can be swapped. A fallback chain moves to a second model when the first one times out. Model usage is billed by the provider to your own account, so you see the real running cost.

## What backs this up

- **A production ordering bot:** Shai built a WhatsApp ordering bot on Claude that runs in production with a model fallback chain, timeouts, retries and deterministic fallback replies.
- **Durable agent loops:** His production agent loops use leases, a scheduled sweep, retries, and token and cost tracking per run.
- **An on-prem MCP gateway:** He owns a gateway product of about 27,000 lines with 13 connector types, built in about two weeks, that lets Claude or ChatGPT answer from ERP, CRM, SQL and Microsoft 365 systems read-only.
- **A named engineer:** [Shai](https://goldenroz.com/about/shai/), our CTO, scopes and builds the agent. The work is not passed to a bench of contractors.
- **Published pricing:** The starting price is on this page and on the [pricing page](https://goldenroz.com/pricing/). You own the code, prompts and data.

## What it costs

Every phase has a fixed price agreed before work starts. Model usage is paid to the model provider from your own account.

- Focused agent or agentic workflow: From $25,000 (Fixed price per phase)
- Larger AI platforms: $100,000 and up (Several agents, many integrations or a full product)
- Scoping call: $0 (Ends with a written scope and a price)
- Model usage: Paid to the provider (Billed to your own account)

## Who it is not for

- Budgets under $25,000.
- Teams that have not yet decided where AI pays off. Start with [AI consulting](https://goldenroz.com/ai-consulting/).
- Training a foundation model or classic machine learning research. We build on existing models.
- Agents that move money or delete data with no human approval. We do not ship that.

## Questions buyers ask

### How much does it cost to build an AI agent?

At Goldenroz, AI agent projects start at $25,000 and are priced per phase after a scoping call. Larger platforms with several agents and many integrations run $100,000 and up. Model usage is a separate running cost paid to the model provider.

### How long does it take to build an AI agent?

A working prototype of a focused agent typically takes one to two weeks. Production takes roughly four to eight weeks, and the main driver is the number of systems the agent connects to. These are typical ranges, scoped per project.

### What is the difference between an AI agent and a chatbot?

A chatbot answers in text. An agent takes actions: it looks up records, creates orders, updates tickets or runs a multi-step task, and checks the result. Conversation-first projects are covered under [AI chatbot development](https://goldenroz.com/ai-chatbot-development/).

### Which models do you use: OpenAI, Claude or others?

We choose per task. Claude is our default for tool-using agents, and we use OpenAI models and others where they fit the task or your contracts. Every agent is built with a fallback chain so a second model takes over when the first fails.

### How does the agent connect to our CRM, ERP or database?

Through MCP servers or direct API integrations. MCP, the Model Context Protocol, is an open standard for giving a model controlled access to tools and data. Details are on the [AI integration](https://goldenroz.com/ai-integration/) page.

### How do you keep an AI agent secure?

Access is read-only by default and scoped per user. The agent runs approved queries instead of writing its own SQL, personal data is masked before it reaches the model, and every call is logged. Actions that cannot be undone require human approval.

### Who owns the code and the data?

You do. The code, prompts, evaluation cases and logs live in your repository and your accounts from the first day. If we stop working together, you keep everything.

### What happens after launch?

The agent ships with cost tracking, logs and an evaluation set you can rerun when a model or a prompt changes. New tools and tasks are scoped as further fixed-price phases. There is no required retainer.

---
Page: https://goldenroz.com/ai-development/
Book a free one-hour call: https://calendly.com/daniarozin/daniel-rozin-consultation-1-hour-clone
