An AI agent for a business is a language model wrapped in five things: a defined job, tools it may call, data it may read, rules for when it may act alone, and an evaluation set that proves it works. In our experience a focused agent takes 4 to 8 weeks to a proof of concept and 10 to 16 weeks to production with monitoring. The model is the smallest line item. Tools, permissions and evaluation are where the time goes.
Most guides on how to build an AI agent start with a framework and a code sample. That is the easy part. The hard part is deciding what the agent is allowed to do, how you will know it is doing it well, and what happens on the day it is wrong. This guide covers the architecture, the three ways to build, what the work costs in weeks, and the failure modes we design for in agents that are live today.
The examples come from three agents Techparser built and runs: Evoriqa, a multi-tenant customer-support agent live across eight channels; an AI reviewer for financial-statement packages at a CPA firm; and AI PDF Chat, a retrieval assistant that cites the page it answered from.
What an AI agent is, and what it is not
IBM’s definition is a useful baseline: an AI agent is a system that can use external tools and data sources to execute tasks with minimal human intervention, usually with a large language model as its engine. The word doing the work is “execute”. A chatbot answers. An agent reads a ticket, looks up the order, drafts the refund and, if allowed, sends it. The difference is tools and permissions, not a smarter model.
That also tells you when not to build one. If the task is a fixed sequence with no judgement in it, a workflow automation in n8n, Make or Zapier is cheaper and more predictable. If the task needs one answer from one document, retrieval with a plain model call is enough. Agents earn their complexity when the input varies, the steps depend on what was found, and a wrong action has a cost you can bound.
The five components every business agent needs
- A job description. One sentence a person could be hired against: “Answer support questions from this tenant’s knowledge base and draft replies for approval.” Agents fail most often because the job was “help customers”.
- Tools. The functions the agent may call: search the knowledge base, look up an order, create a ticket, send an email. Each tool is code you write and test, with its own permissions.
- Data. What the agent may read, and for whom. In Evoriqa every retrieval is filtered by tenant inside PostgreSQL with pgvector, so one business’s content can never appear in another’s answer.
- An autonomy policy. Which actions run unattended, which need approval, and which are forbidden. Evoriqa ships with a shadow mode where the agent drafts and a person approves until the numbers justify switching it on.
- An evaluation set. Fifty to two hundred real cases with the answer you expect. Every prompt, model or tool change is scored against it before it reaches users.
Three ways to build: scratch, framework or platform
IBM’s guide lists the same three routes we see in practice: code the loop yourself in Python or JavaScript, use an agent framework such as LangChain, LangGraph, AutoGen or crewAI, or buy a platform that hosts, deploys and monitors the agent for you. None is wrong. They trade control for speed in different places.
| Approach | Time to first demo | Control over tools and data | Where it hurts later | Fits when |
|---|---|---|---|---|
| Custom loop (model API + your code) | 1 to 2 weeks | Total: every tool, prompt and permission is yours | You own retries, tracing and evaluation tooling | Regulated data, custom tools, cost control matters |
| Agent framework (LangChain, LangGraph, crewAI, AutoGen) | 2 to 5 days | High, inside the framework’s abstractions | Framework upgrades and hidden prompt behaviour | Many tools, multi-step planning, team knows Python or TypeScript |
| Hosted agent platform or no-code builder | Hours | Limited to the platform’s connectors | Per-seat or per-run pricing, data leaves your accounts | Single workflow, standard SaaS tools, no engineers in-house |
Our default for anything that touches customer data is a custom loop or a thin framework on top of a model we can swap. AI PDF Chat uses LangChain with Qdrant for vectors and Gemini embeddings, and the model behind it is a configuration change. The financial-statement reviewer uses no framework at all: a FastAPI service calls Claude, and every number the model sees was already checked by deterministic code.
What it costs: the drivers in weeks
We do not publish fixed prices because the model call is rarely the cost. The table below is the estimating baseline we use for a two-engineer squad; multiply weeks by the day rate you are quoted and the comparison across vendors becomes honest.
| Driver | Low | Typical | High | What pushes it up |
|---|---|---|---|---|
| Tools and integrations | 1 to 2 read-only tools: 1 week | 3 to 5 tools incl. one write action: 2 to 4 weeks | CRM, inbox, billing and documents: 5 to 8 weeks | Each write action needs an approval path and an audit log |
| Data and retrieval | One clean document set: 0.5 to 1 week | Per-customer isolation, background ingestion: 2 to 3 weeks | Many sources, permissions per user: 4 to 6 weeks | Isolation, citations, re-indexing on change |
| Evaluation set and quality bar | 50 cases, one reviewer: 1 week | 150 cases, scored weekly: 2 weeks | Regulated accuracy, deterministic checks: 3 to 5 weeks | Every prompt change must be re-scored |
| Autonomy and safety | Draft-only: 0.5 week | Shadow mode, approval gates: 1 to 2 weeks | Unattended actions with rollback: 3 to 4 weeks | Irreversible actions need gates, logs and limits |
| Monitoring and cost control | Logs: 0.5 week | Dashboards for volume, accuracy, cost per run: 1 to 2 weeks | Metered credits, spend caps, alerts: 2 to 3 weeks | Model cost is per run; a busy tenant can become a loss |
Add the rows and a single-workflow agent lands at 4 to 8 weeks. An agent connected to a CRM, inbox and documents, with a supervised rollout, lands at 10 to 16 weeks. Those are the ranges on our AI development service page because they are the ranges we deliver against.
The failure modes to design for before launch
- The model does arithmetic. It will be confidently wrong. In the financial-statement reviewer, code refoots every total and balances the balance sheet; the model only judges disclosures and wording.
- The model call fails silently. Rate limits, empty replies and timeouts all happen. The reviewer gives the model at most two attempts and reports the failure; AI PDF Chat was redesigned around an embedding provider that returned empty results instead of an error.
- The agent acts when it should have asked. Put every irreversible action, sending email, changing billing, closing a ticket, behind an approval gate until the evaluation numbers hold.
- Cost grows with usage, not revenue. Meter every conversation and cap spend per customer. Evoriqa runs a single credit balance per tenant so a busy or abusive workspace cannot become a loss.
- Data crosses a boundary. Filter retrieval by tenant or user inside the query, not in the application after the fact.
A build plan that fits in a quarter
- Week 1: write the job description, list the tools, decide the autonomy policy, and collect 50 real cases with expected outcomes.
- Weeks 2 to 4: build the tools and retrieval, run the agent against the cases, and fix the top three failure types.
- Weeks 5 to 6: shadow mode with real traffic, people approving, accuracy and cost per run on a dashboard.
- Weeks 7 to 10: switch on unattended actions one at a time, each with a gate and a rollback, and keep re-scoring the evaluation set after every change.
Frequently asked questions
- Is it free to build an AI agent?
- The software can be. Frameworks such as LangChain and LangGraph are open source and several models have free tiers. The cost is engineering time, which is dominated by tools, permissions and evaluation, plus model usage once real traffic arrives. A hobby agent can cost nothing; a business agent with write access to your systems is a 4 to 16 week project.
- Can you build an AI agent with ChatGPT?
- You can prototype one with ChatGPT or any assistant that supports tools, and that is a good way to test whether the job is well defined. A production agent needs its own code path: tools you control, retrieval filtered by customer, approval gates, logging and an evaluation set. The model behind it can be OpenAI, Claude, Gemini or an open model behind the same interface.
- What is the 10/20/70 rule for AI?
- A rule popularised by BCG: roughly 10 percent of the effort in an AI project is algorithms, 20 percent is technology and data, and 70 percent is people and process. Our cost table shows the same shape. The model call is a small line; tools, data isolation, autonomy policy and evaluation are the project.
- How do I build an AI agent from scratch in Python?
- Write a loop: send the user request and tool definitions to the model, execute whichever tool it calls, append the result, and repeat until it returns a final answer. Wrap that loop with a step limit, a timeout, logging of every call, and a check that the chosen tool is allowed. That is a working agent in a few hundred lines; the rest of the work is the five components above.
- What are some AI agent examples in business?
- A support agent that answers from a company’s own knowledge and drafts replies for approval, a document reviewer that checks a financial package against the firm’s rules, a research assistant that answers across uploaded PDFs with citations, and an operations agent that reads an inbox and files tickets. The first three are live Techparser builds described in the case studies linked below.
About the author
Zoraiz Ejaz
Co-founder, Techparser
Zoraiz Ejaz is a co-founder of Techparser and leads its engineering and product practice. He has spent close to a decade designing, building and scaling mobile, web and AI products for startups and enterprise teams across health, fintech, payments, social and education, from first architecture and release pipelines through to launch and years of production support. He writes about how to scope, cost and ship software that lasts.
