# Custom AI agent development for work that crosses systems

> A custom AI agent is a language model connected to your business tools with limited permissions, so it can finish a multi-step task such as qualifying a lead or resolving a refund request. We deliver one production agent with its tool connectors, approval steps, test set and run logs, on your accounts. Typical build: 6 to 12 weeks, $15,000 to $90,000, model usage at cost.

- Page: https://methodcrew.com/services/custom-ai-agents
- Last updated: 2026-10-28
- Publisher: [The Method Crew](https://methodcrew.com/)
- Service: Custom AI agents and copilots
- Category: AI systems

We build agents that read a request, look things up in your CRM, helpdesk or inbox, and take the next step, with a person approving anything that matters.

## Key facts

- Price: $15,000 to $90,000 per project. One production agent with 2 to 3 integrations: $15,000 to $40,000. Multi-tool or multi-agent systems: $40,000 to $90,000. Model usage billed at cost, no markup.
- Timeline: 6 to 12 weeks
- Typical tools: OpenAI API, Anthropic API, Google Gemini API, Model Context Protocol, OpenAI Agents SDK, LangGraph, Langfuse, n8n, Slack, Microsoft Teams, HubSpot, Salesforce, Zendesk, PostgreSQL
- Change requests: $175 an hour, quoted in writing first
- Model and tool usage: billed to your own accounts at cost, no markup
- Ownership: code, workflows, prompts, accounts and data belong to you

## Who it is for

- Sales teams of 5 to 50 reps who research and log every lead by hand
- Support teams on Zendesk, Intercom or Freshdesk whose tickets need lookups in other systems
- Operations and finance teams moving the same data between 3 or more tools
- Companies of 10 to 500 staff whose off-the-shelf AI tool stops at its vendor's edge

## Signals you need it

- **One task touches 3 or more systems.** Someone reads an email, checks the CRM, opens billing, then writes a reply. Each step is simple. Together they take a morning.
- **Your rule-based automations break on exceptions.** Every new email format needs another branch in Zapier or n8n. An agent reads the variation and picks the right path.
- **Your vendor's AI stops at its own walls.** The helpdesk bot cannot see refunds in Stripe. A custom agent works across the tools you use.
- **You can describe a good result.** If 2 of your people would handle a case the same way, an agent can follow that pattern. If they disagree, fix the process first.

## What we hand over

- **Agent job description.** A 2-page spec: the goal, every tool the agent may call, the actions it may never take, and when it must hand a case to a person. You approve it before we write code.
- **Tool connectors with scoped access.** API or Model Context Protocol connections to your CRM, helpdesk, inbox and databases, each on a service account limited to what the spec allows.
- **Evaluation set.** 50 to 200 real cases from your history, each with the outcome your team would accept. It runs before every prompt change, model upgrade or new tool.
- **Approval and handoff flow.** Slack or Microsoft Teams messages where a person approves, edits or rejects a proposed action, plus a handoff note whenever the agent stops.
- **Run log and cost report.** Every step, tool call and model cost recorded in Langfuse or your own database, searchable by customer, date and outcome.
- **Runbook and supervisor training.** How to pause the agent, change its instructions, read the logs and add a test case, taught to the people who supervise it.

## How the engagement runs

1. **Map** (1 to 2 weeks): We sit with the people who do the work, pull 50 or more real cases from your systems, write the agent job description and agree the pass bar the pilot must meet.
2. **Prove** (2 to 3 weeks): The agent runs in shadow mode on live cases. It reads, looks things up and drafts the action, but changes nothing. Your team compares its drafts with what they did. If it misses the bar, we stop there.
3. **Build** (3 to 7 weeks): We switch on write access behind approvals and add error handling, spend limits, logging and the evaluation suite, with a demo every week. Approvals relax only where the logs justify it.
4. **Run** (Monthly, optional): We review logs weekly, turn failures into test cases and retest when a model version changes. You get a monthly report on volume, approvals, failures and cost, or we hand everything over.

Every engagement follows [The Method](https://methodcrew.com/approach): Map, Prove, Build, Run.

## What is a custom AI agent, and how is it different from a chatbot?

A custom AI agent is a language model that can call tools: search the CRM, read a ticket, update a record, send a message. It works through a task in several steps and picks each step from what it finds. A chatbot answers questions. An agent finishes work.

That sets the risk. A chatbot that gets something wrong writes a bad sentence. An agent that gets something wrong can refund the wrong order, so most of the build is about tools and limits.

A copilot sits in between: an assistant inside a tool your team already uses, such as a CRM sidebar, that drafts while a person decides. Many agents should ship first as copilots and earn autonomy from their logs.

| | Chatbot | Workflow automation | Copilot | AI agent |
|---|---|---|---|---|
| What it does | Answers questions | Runs fixed steps on a trigger | Drafts inside a tool | Takes several steps across tools |
| Who decides | The user | The rules | A person, every time | The agent, within limits you set |
| Copes with new variations | Partly | No | Yes | Yes |
| Running cost | Low | Lowest | Low to medium | Medium, per task |

If fixed rules do the job, we build [business process automation](https://methodcrew.com/services/business-process-automation) instead, because it is cheaper to run and easier to audit. The longer comparison is in [AI agents vs automation vs chatbots](https://methodcrew.com/insights/ai-agents-vs-automation-vs-chatbots).

## Can an AI agent work inside the tools we already use?

Yes, if the tool has an API, and most business software does. We connect agents to HubSpot, Salesforce, Zendesk, Intercom, Gmail, Outlook, Slack, Microsoft Teams, Stripe and internal databases. Where a vendor publishes a [Model Context Protocol](https://methodcrew.com/glossary/model-context-protocol) server, we use it. Otherwise we write the connector on the models' standard function calling interface.

Each connection runs on its own service account, scoped to the job description. A support agent might read orders, write internal notes on tickets and send nothing to a customer without approval. It never gets admin rights because they happen to be convenient.

People meet the agent where they already work. A sales rep gets a Slack message with the lead summary and a button to accept the suggested owner. A support rep sees a drafted reply as an internal note in the helpdesk. Nobody learns a new app.

There are 2 limits worth knowing early. Tools without an API, such as some older desktop accounting software, need a workaround: a scheduled export, a shared inbox or a person in the loop. And an agent is only as current as the data it reads. If your CRM runs a month behind, so will the agent. Map flags both before you sign anything.

## How do you keep an AI agent from taking the wrong action?

With several layers, each built on the assumption that the one before it can fail. The model will sometimes misread a case, so the design limits what a misreading can cost.

1. The agent can call only the tools in its job description, with the narrowest access that does the job. Read access is generous. Write access is earned.
2. Anything that moves money, changes a contract or messages a customer waits for a person's approval, until the logs justify relaxing that rule for one specific action.
3. Hard limits live in code, outside the prompt: refund ceilings, daily action caps, a monthly spend limit. The model cannot talk its way past them.
4. Everything the agent reads is treated as data. Emails and attachments can carry hidden instructions, and OWASP ranks [prompt injection](https://methodcrew.com/glossary/prompt-injection) first in its [Top 10 for LLM applications (2025)](https://genai.owasp.org/llmrisk/llm01-prompt-injection/). We keep instructions apart from content and test against planted text.
5. The evaluation set runs before every change, and every live action is logged with its inputs, so a bad decision can be traced in minutes.

A supervisor can pause the agent from Slack or Teams with one command. Where the stakes stay high, such as payments or legal commitments, a person stays in the loop permanently, and the spec says so.

## How much does it cost to build a custom AI agent?

Most single agents cost $15,000 to $40,000 to build: one production agent with 2 to 3 tool integrations, approval steps, an evaluation set and handover. Systems with many tools, several cooperating agents, or write access to finance or contract records cost $40,000 to $90,000. A narrower agent that follows a fixed sequence of steps, such as the [inbound lead qualification agent](https://methodcrew.com/blueprints/inbound-lead-qualification-agent), is built as a workflow with model steps and costs $12,000 to $25,000 over 4 to 6 weeks.

The price moves with 5 things:

- Tools the agent writes to. Reading is cheap to build. Each write path needs its own approvals, limits and tests.
- Input mess. Clean form submissions are easier than free-text emails with attachments in 3 languages.
- Evaluation set size. A support agent covering 40 ticket types needs far more test cases than a lead router covering 4.
- Approval design. One approver in Slack is simple. Tiers by amount, region or customer take longer.
- Access. If credentials or a security review take 3 weeks, the calendar stretches even when the hours do not.

Running costs are billed at cost, with no markup: model usage from OpenAI, Anthropic or Google, plus hosting. We measure cost per task during Prove, so the monthly estimate in the proposal comes from your own data. Monitoring and tuning after launch is optional, at $1,500 to $5,000 a month.

## When is a custom agent the wrong choice?

When something cheaper does the job. We would rather say so on the first call than after the pilot.

Buy before you build if the work lives inside one vendor. If your whole support process runs in Zendesk or Intercom, their own AI agents deserve a trial first, and we can help you judge one against your real tickets. The same goes for the agents built into HubSpot and Salesforce. A custom agent is worth its cost when the task crosses vendors, depends on your own business rules, or when you want to own the prompts, logs and tests outright.

Use plain automation if the steps never change. A form that routes on a dropdown does not need a language model.

Wait if the process is not settled. An agent copies a process. If your team handles the same case 3 different ways, the agent will too, only faster.

If the job is answering questions from documents with no action at the end, a [RAG knowledge system](https://methodcrew.com/services/rag-knowledge-systems) fits better and costs less to run.

An agent build does not include training or fine-tuning your own model (unless the evaluation set proves prompts and retrieval cannot reach the bar), consumer-facing apps, CRM data cleanup, which we scope separately, or software licenses, which you buy in your own name.

## What changes for your team

- Reps start the day with inbound leads already researched, scored and logged in the CRM.
- Tickets that need a lookup arrive with the account history pulled in and a reply drafted.
- Every agent action sits in a log you can search by customer, date or tool.
- Approvals happen in Slack or Teams in one click.
- You can switch model vendors without a rebuild, because the tools, prompts and tests are yours.

## Tools we work with

OpenAI API, Anthropic API, Google Gemini API, Model Context Protocol, OpenAI Agents SDK, LangGraph, Langfuse, n8n, Slack, Microsoft Teams, HubSpot, Salesforce, Zendesk, PostgreSQL

## What it costs

$15,000 to $90,000 per project. One production agent with 2 to 3 integrations: $15,000 to $40,000. Multi-tool or multi-agent systems: $40,000 to $90,000. Model usage billed at cost, no markup.

- Change requests: $175 an hour, quoted before work starts
- Model and tool usage: billed at cost, no markup

All 20 price bands: https://methodcrew.com/pricing

## Blueprints that use this service

- [Support knowledge assistant](https://methodcrew.com/blueprints/support-knowledge-assistant): built in 6 to 12 weeks, $20,000 to $60,000 per project.
- [Contract review assistant](https://methodcrew.com/blueprints/contract-review-assistant): built in 8 to 12 weeks, $20,000 to $40,000 per project.

## Common questions

### How long does it take to build a custom AI agent?

Most agents take 6 to 12 weeks from scoping call to production: 1 to 2 weeks of Map, a 2 to 3 week shadow-mode pilot and a 3 to 7 week build. A single agent with mostly read access lands near the short end. Agents that write to finance systems, or several agents working together, land near the long end because approvals and testing take longer.

### Which AI models do you use?

We test the task on current models from OpenAI, Anthropic and Google during Prove and pick the one that passes your evaluation set at the lowest cost per task. Often a smaller, cheaper model handles routing and a larger one handles the hard cases. Because the tools and the test set are separate from the model, you can switch vendors later without rebuilding the agent.

### Do we own the agent and its code?

Yes. The code, prompts, evaluation set, logs and connector settings are yours, and everything runs on accounts in your name: your model API keys, your cloud, your Slack or Teams workspace. If we part ways, you keep a working system and a runbook that explains how to operate, change and pause it.

### Will our data be used to train AI models?

OpenAI and Anthropic both state that they do not train on data sent through their business APIs by default. We confirm the current terms, data retention settings and hosting region for your account during Map, and we keep sensitive fields out of prompts when the task does not need them. If your sector needs a specific contract term, we flag it before the build starts.

### What happens when the agent is unsure?

It stops and hands the case to a named person, with a note explaining what it found and why it stopped. The job description lists the stop conditions in plain language: missing data, conflicting records, amounts above a limit, a customer asking for a human, or anything outside its scope. Expect frequent handoffs in the first weeks, falling as the test set grows.

### Can the agent learn from our corrections?

Not on its own, and that is deliberate. Models do not retrain on your edits in the background. We log every correction your team makes, review them each week, turn the useful ones into new test cases and update the agent's instructions. Changes go live only after the full evaluation set passes, so a fix for one case cannot break another without anyone noticing.

## Related services

- [RAG and knowledge systems](https://methodcrew.com/services/rag-knowledge-systems): Your documents turned into an assistant that cites its sources. $15,000 to $75,000 per project.
- [AI automation](https://methodcrew.com/services/ai-automation): Workflows where a language model reads and sorts, and your existing systems do the rest. $4,000 to $35,000 per project.
- [AI ops monitoring and improvement](https://methodcrew.com/services/ai-ops-monitoring): Quality checks, cost control and model updates for live AI systems. $1,500 to $5,000 per month.

## Further reading

### Guides

- [What AI consulting costs in 2026: rates, packages and price drivers](https://methodcrew.com/insights/ai-consulting-cost)
- [AI agents vs workflow automation vs chatbots: how to pick](https://methodcrew.com/insights/ai-agents-vs-automation-vs-chatbots)
- [How to build a company knowledge assistant that cites its sources](https://methodcrew.com/insights/build-rag-knowledge-assistant)

### Terms explained

- [Agentic AI](https://methodcrew.com/glossary/agentic-ai): Agentic AI is the umbrella term for systems where language models plan steps and take actions through tools, with less step-by-step direction from people.
- [AI agent](https://methodcrew.com/glossary/ai-agent): An AI agent is a language model that chooses and calls tools in a loop, reading each result, until it finishes a task or hands off to a person.
- [AI copilot](https://methodcrew.com/glossary/ai-copilot): An AI copilot is an assistant built into software your team already uses that drafts and suggests work, while a person makes every final decision.
- [AI guardrails](https://methodcrew.com/glossary/ai-guardrails): AI guardrails are checks placed around a language model that block, rewrite or escalate inputs and outputs that break your rules on safety, accuracy, privacy or scope.
- [AI hallucination](https://methodcrew.com/glossary/ai-hallucination): An AI hallucination is output from a language model that is fluent and confident but false, such as an invented citation, policy, figure or product detail.
- [Context window](https://methodcrew.com/glossary/context-window): A context window is the maximum amount of text, measured in tokens, that a language model can read and write in one request, counting instructions, documents and its reply.
- [Fine-tuning](https://methodcrew.com/glossary/fine-tuning): Fine-tuning is further training of an existing language model on your own examples, so it learns a specific format or task without long instructions every time.
- [Function calling (tool use)](https://methodcrew.com/glossary/function-calling): Function calling lets a language model ask your software to run a named function, such as a CRM lookup, by returning the function name and arguments as JSON.

## Next step

Request a 30-minute scoping call: https://methodcrew.com/contact. Or email contact@methodcrew.com.