Every vendor sells agents now
Sound familiar?
Your board wants an AI agent strategy, three vendors have pitched "autonomous agents", and nobody can say which of your workflows one would actually improve, or what it would cost to run.
An AI agent is software that works towards a goal rather than following fixed steps. Given a task, it decides what to do next, uses tools such as your CRM, database, email or internal APIs, checks the result, and keeps going until the task is done or it needs a person. That is different from a chatbot, which only answers, and from a traditional automation, which follows the same rules every time.
Agents are worth building when a workflow needs judgement at several steps but follows a recognisable pattern. They are a poor fit when the rules are fixed (a normal automation is cheaper and more reliable) or when every mistake is expensive and hard to undo.
Good first AI agent projects
IfSupport tickets need looking up orders, accounts or policies before replying
ThenA support agent that gathers the context and drafts the reply for a person to approve.
IfInbound documents such as invoices, forms or contracts need reading and routing
ThenAn intake agent that extracts the data, checks it against your systems, and routes exceptions.
IfSales or operations staff spend hours collecting data from several tools
ThenA research agent that pulls the information together into one brief.
IfThe process follows fixed rules every time
ThenSkip the agent. A workflow automation is cheaper and more predictable.
The best first projects share three traits: the task happens often, a person can check the output quickly, and a mistake is caught before it reaches a customer. Start with the agent drafting and a person approving, then reduce the checks as the results earn trust.
How a production agent is built
Define the job and the limits
Write down the exact task, the tools the agent may use, and the actions it must never take without approval.
Connect the tools safely
Give the agent narrow, permissioned access to the systems it needs, and log every action it takes.
Build guardrails
Validate inputs and outputs, cap spending and retries, and route anything uncertain to a person.
Test on real cases
Run the agent on 50 to 100 real past cases, including the awkward ones, and measure how often it gets them right.
Launch with a person in the loop
Start with approval on every action, then relax it where the results justify it.
Monitor quality and cost
Track accuracy, escalations and model costs per task, and re-test when you change the model or prompts.
The model is the easy part. Integrations, guardrails and evaluation are where most of the build time goes, and they decide whether the agent can be trusted. Our guide to choosing an AI model covers the model side, and integrating LLMs into web applications covers the engineering.
Planning an AI agent for your business?
Get a free consultationWhat an AI agent costs
| Scope | Typical range | What drives the cost |
|---|---|---|
| AI proof of concept | $10,000 to $30,000 | One workflow, test data, no production hardening |
| Single-task AI agent | $15,000 to $80,000 | Integrations, guardrails, and evaluation |
| Multi-agent or enterprise system | $100,000 to $400,000+ | Orchestration, compliance, many connected systems |
| Running costs after launch | 15 to 25% of build cost per year | Model usage, hosting, monitoring, and maintenance |
Model usage is the running cost that surprises people, because an agent may call a model many times per task. Our AI API pricing guide shows how to estimate it, and the custom software cost guide puts these figures in context.
US compliance points to settle early
An agent that reads and acts on business data inherits every obligation attached to that data. Settle these before the build, not after:
- Health data: if the agent touches protected health information, HIPAA applies, and every vendor in the chain, including the model provider, needs a business associate agreement
- Customer data: state privacy laws such as California's CCPA give consumers rights over their data, including knowing what is collected and asking for it to be deleted
- Security reviews: many US buyers expect suppliers to show a SOC 2 report or answer a security questionnaire before go-live
- Model providers: check whether your data is retained or used for training, and choose enterprise terms that rule it out
- Audit trail: keep a log of what the agent saw, decided and did, so any action can be explained later
What to ask an AI agent development company
- Show us an AI feature you have shipped to production, not a demo
- How do you test an agent before launch, and how do you measure accuracy?
- What can the agent do without approval, and how is that enforced?
- How do you keep model costs per task under control?
- How will our data be handled, stored and protected?
- How many of our business hours overlap with your team for reviews?
A good partner will talk you out of an agent where a simpler automation would do. For help scoping or building one, see our AI integration services.
Key takeaway
Pick a frequent task that a person can check quickly, build integrations, guardrails and evaluation properly, launch with a person in the loop, and settle data and compliance questions before you write the first prompt.
Frequently Asked Questions
Related service
AI integration
Practical AI inside your product or workflow: document processing, summaries, search, assistants, and automation.
Written by
Tricolens
AI & Engineering
