You watched the demo. The agent answered questions, routed tasks, called an API. It looked done. Then you tried to ship it and discovered that the demo was maybe 20% of the work.
This article is about the unglamorous timeline of shipping an AI agent to production: what actually happens between "it works in my notebook" and "it works reliably for real users." No hype. Just the phases, the surprises, and what determines whether you get there in six weeks or six months.
Who This Is and Is Not For
This is for a founder or a technical lead who already knows they need an AI agent. You have a use case: customer support routing, document processing, internal knowledge retrieval, something specific. You have a budget to build it properly. You are not here to be convinced that AI is useful.
This is not for you if you want a prototype to show investors and call it done. A prototype is a different project with a different price and a different scope. If you need that, I can help with that too, but read this first so you know the difference.
Phase 1: The Scope You Think You Have vs. the Scope You Actually Have
Weeks 1 to 2
Every agent project starts with a deceptively simple brief. "We want the agent to answer customer questions from our knowledge base." Fine. But then:
What happens when the knowledge base has no good answer? Who owns the knowledge base and how often does it change? Does the agent need to escalate to a human, and if so, through what channel? What counts as a wrong answer, and what is the consequence of one? Which languages does your customer base actually use?
These are not edge cases. They are the product. Skipping this phase produces an agent that works in demos and fails on Tuesday morning.
I spend the first two weeks doing nothing except mapping these questions with the team that will own the system. The output is a one page spec: inputs, outputs, failure modes, and the single metric that will tell us if the thing works.
Phase 2: The Stack Decision (and Why It Matters More Than the Model)
Week 2 to 3
GPT 4o, Claude, Gemini, a fine tuned open model. Everyone asks about the model first. The model is usually not the bottleneck.
The real decisions:
| Decision | Why It Costs You Later If You Guess |
|---|---|
| Retrieval strategy | Bad chunking means the agent confidently answers with wrong context |
| Tool call design | Poorly scoped tools produce loops and hallucinated parameters |
| State management | Stateless agents forget context; over stateful agents are slow and expensive |
| Observability | No logging means no debugging when it breaks at 2am |
| Fallback behavior | What does the agent do when the LLM call fails or times out? |
I use a combination of LangChain or direct API calls depending on the complexity, Next.js for any frontend surface, and structured logging from day one. The stack is boring on purpose. Boring stacks are debuggable stacks.
Phase 3: The Build That Actually Takes Time
Weeks 3 to 7
This is where most estimates fall apart. The core agent logic takes one to two weeks. The remaining two to four weeks go to:
Prompt iteration. You will write the system prompt three times before it stops producing embarrassing outputs on realistic inputs. Budget for this.
Retrieval tuning. If the agent uses RAG, the first retrieval pipeline will have a recall problem, a precision problem, or both. You find this by running real queries, not synthetic ones.
Tool reliability. Every external tool call is a contract. APIs change, rate limits hit, auth tokens expire. Each one needs a retry strategy and a graceful failure path.
Evaluation harness. Before you can call the agent good, you need a set of test cases that represent real usage. Building this is not glamorous and it is not optional.
I am a builder, not a consultant. The code, the evals, the infra: I own all of it. If you have an internal team that wants to learn while we build, I can run the project so they are onboarded by the time we ship.
If this timeline matches the project you are sitting on right now, it is worth a conversation before you start. Work with me at ledianshera.com and we can scope it in one call.
Phase 4: Production Is a Different Environment Than Your Laptop
Weeks 6 to 9
The agent works. Now you deploy it. This is where "almost done" becomes two more weeks.
Production brings:
Latency. The agent that responded in 1.2 seconds locally now takes 3.8 because you are hitting a real API under load, running retrieval against a real vector store, and routing through a real auth layer. Cost at scale. Token counts that looked reasonable in testing become expensive at production volume. You need to optimize prompts, cache where possible, and monitor spend. User behavior you did not anticipate. Real users ask things your test set never covered. You need a way to capture those inputs and retrain or adjust without a full redeploy cycle. Security surface. Prompt injection is real. Input sanitization and output validation are not optional.
I set up monitoring before go live, not after the first incident. That means structured logs, cost dashboards, and a simple alert for failure rate spikes.
What Determines Whether You Hit 7 Weeks or 16 Weeks
The honest answer:
Your side of the project. The biggest source of delay is not the engineering. It is access to the right stakeholder to make a product decision, or waiting for someone to share the data the agent needs to be good at its job. Teams that assign one clear owner to this project move twice as fast.
Scope creep. "Can we also make it handle X?" is how a 7 week project becomes a 16 week project. New capabilities go on a roadmap. They do not get added mid build.
Data quality. If your knowledge base is a folder of PDFs from 2019 with no structure, the first three weeks include cleaning that data before building anything.
The Objection Worth Addressing Directly
You might be thinking: "This sounds expensive for something that might not work."
Fair. Here is what I can tell you honestly. I have shipped AI systems for real businesses through AlbTech Solutions. Some of them took longer than the first estimate because the problem turned out to be harder than the brief suggested. When that happened, I said so early, not at the end.
I do not do fixed scope contracts for agent work because agent work is iterative by nature. I do milestone based agreements where you see working software at the end of each phase and decide whether to continue. You are never locked in past the current milestone.
If the agent does not perform to the agreed spec at the end of the evaluation phase, that is a conversation we have then, not a surprise invoice you receive later.
For founders who want to move faster on the initial product, my team also offers structured startup MVP development if the agent is part of a larger product you need to validate quickly.
What Happens When You Reach Out
Send me a message at ledianshera.com. Tell me what the agent needs to do and what you have tried so far. Within one business day I will reply with either a few clarifying questions or a proposed 30 minute call.
On that call we do one thing: figure out whether your project is in the right phase for what I do. If it is, I send a milestone breakdown within 48 hours. If it is not, I tell you that and point you somewhere better.
No pitch deck. No proposal theater. One call, one honest answer.