You can build an impressive AI agent demo in a weekend. Getting one that a business can rely on every day usually takes a few weeks. The gap is not the model; it is everything around it: a clearly defined job, access to the right data, rules for what the agent may and may not do, and a way to prove it works.
This guide walks through what goes into a production agent, how long each part takes, and what actually drives the cost.
Take the YouTube production pipeline we built. The "agent" in it is a single step: a story-planning agent that turns a schedule trigger into a structured plan for the next video. Everything else is ordinary engineering: turning that plan into render requests, waiting for clips to finish, putting them in order, combining them, uploading, publishing, and logging every run so a failed one can be picked up where it stopped. The agent step was the quickest part to get working. Making the whole run dependable, week after week, was the actual project.
Start with one job, not "an agent"
Every agent we build starts with a specific job: a defined input, a defined output, and clear rules for what the agent can and cannot do. "An AI agent for sales" is not a job. "Research each new inbound lead and draft a qualification summary for the account owner" is.
A good job definition answers:
- What triggers the agent, and what does it receive?
- What exactly should it produce, and where does that go?
- What is it never allowed to do on its own?
- When should it hand over to a person?
Here is what that looks like for a real estate team handling property inquiries on WhatsApp, the kind of workflow we built for Fursa Ghyr.
Before: "We want an AI agent that handles our WhatsApp leads."
After: "When a new WhatsApp inquiry arrives, extract what the buyer is looking for (area, property type, budget, timing) into a structured demand record. Compare it against current listings and send the best matches to the responsible broker with the conversation attached. Never confirm availability, prices, or viewings to the buyer. If the request is unclear or the buyer asks for a person, alert the team."
The second version tells you what to build, what to connect to (WhatsApp, the listings data, the CRM), what to test, and where a person stays in the loop. It also shows that most of the work is structure and routing, with judgment needed only at a few points.
What you need ready before the build
Most delays come from things the agent depends on, not from the AI itself. Before a build starts we ask for:
- A specific task or job for the agent to perform
- Examples of good and bad outputs
- Access to the relevant data sources or APIs
- The permission structure and escalation rules
- One technical contact who can support integration
- Success criteria you can measure
The build, week by week
Our AI Agent Build Sprint usually runs three to five weeks and delivers:
- The agent workflow design
- Prompt and system logic
- Tool calling setup (the actions the agent can take in your systems)
- Guardrails and permissions
- Evaluation and test cases
By the time a sprint starts, the free AI Systems Audit and scoping have already fixed the job, the systems involved, and the success criteria. A typical sprint then runs like this:
Week 1: the job, the data, and the test set. We confirm the job definition, get access to every system the agent reads from or acts in, and turn your good and bad examples into a first evaluation set. If access is going to be slow, this is the week it shows up.
Weeks 2 and 3: tools and logic. We build the actions the agent can take (look up a record, create a draft, update the CRM, send an alert), the prompt and system logic that decides between them, and structured outputs so every result has a predictable shape the next step can rely on. The agent runs against the test set throughout, so quality is measured rather than guessed.
Week 4: guardrails, failure handling, and review. We add the permission checks, the handover rules, and what happens when something fails: a timeout, a missing field, an answer the agent is not confident about. Your team reviews real outputs, and we adjust until they meet the success criteria.
Week 5, when it is needed: integration and launch. Deployment, monitoring, and the handoff documentation. Simple agents fold this into week 4.
What pushes a sprint closer to five weeks than three:
- Slow access. Waiting for API keys, credentials, or permission to connect a system.
- Data that needs work first. Information spread across PDFs, inboxes, and spreadsheets has to be cleaned up and structured before an agent can use it.
- No examples. Without good and bad outputs, the test set has to be built from scratch, and so does the agreement on what "good" means.
- Acting, not drafting. An agent that changes records or messages customers needs more guardrails and testing than one that drafts for a person to approve.
- Each extra integration or channel. Every system and every place the agent lives adds build and testing time.
What drives the cost
We do not publish fixed prices for agents, because two agents that sound alike can differ several times over in effort. Every agent sprint is priced as a fixed scope after the audit, so you know the cost before the build starts. What moves that number is mostly the agent's surroundings, not the model:
- Number of integrations. Each system the agent reads from or acts in adds build and testing time.
- Data quality. Clean, accessible data is quick. Data spread across PDFs, inboxes, and spreadsheets needs work first.
- Risk level. An agent that drafts for a person to approve is simpler than one that acts on its own.
- Volume. Model usage costs scale with how often the agent runs and how much context it reads.
- Channels. Each place the agent lives (Slack, WhatsApp, your product, email) adds its own work.
The running costs after launch
An agent is not finished at launch. Expect ongoing model usage costs, monitoring, and small improvements as real usage shows where it struggles. That is what our monthly AI Operations Retainer covers: optimization, prompt and workflow improvements, bug fixes, and small additions.
Model usage is easy to estimate once the job is defined: the number of runs a month, multiplied by how much text each run reads and writes, multiplied by the model provider's price. As an illustration, an agent that runs 2,000 times a month and reads and writes about 10,000 tokens each time uses 20 million tokens a month. At a few dollars per million tokens, that is tens of dollars, not thousands. Check current prices with your provider, because they change often and vary a lot between models.
For most business agents, model usage is the smaller part of the running cost. The larger part is keeping the agent useful: reviewing the cases it gets wrong, adjusting prompts and rules, and adapting when a connected system changes. Two things raise usage costs quickly: feeding the agent large documents on every run instead of retrieving only what it needs, and using the most capable model for steps a smaller one handles just as well.
When you should not build an agent
We do not pitch AI where automation is enough. If the task follows fixed rules every time, a plain workflow automation is cheaper, faster, and easier to trust. Agents earn their cost when the work needs judgment: reading, classifying, drafting, or deciding between options.
Our Google Reviews Automation project is a good example. A multi-location service business needed to track new reviews across its locations, see which reviews mentioned an employee or included a photo, and work out monthly bonus eligibility. Only one of those steps needs judgment: reading a review and deciding whether it names a specific employee. AI handles that step, and unclear cases are flagged for a person. Collecting the reviews, recording photos, applying the bonus rules, and producing the monthly report are fixed rules, so they run as ordinary automation. Building the whole thing as an agent would have cost more, run slower, and made the bonus numbers harder to trust.
A useful test: if you can write the rules down completely, automate them. Use AI only for the parts you cannot.
A checklist before you start
Not sure whether your use case needs an agent or a simpler automation? That is exactly what the free AI Systems Audit is for. The Custom AI Agents page shows how we build them.
