Agent, workflow, or chatbot?
I'm Emil Hjorth, and I run em8, a founder-led AI studio in Mora, Sweden. Most "AI agent development" calls I take start with the wrong noun, so here are the three things people mean:
| You are buying | Who decides the next step | Good for | Weak when |
|---|---|---|---|
| AI workflow | You, in advance. The model fills in one step (classify, extract, draft, score) | Repeatable jobs with a known path: lead routing, inbox triage, CRM writeback | The path really is different every time |
| AI agent | The model, inside the limits you set: which tools, how many steps, when to stop | Research, multi-tool prep, cases where the next step depends on what it just found | Nobody can say what "done" looks like, or one wrong action is expensive |
| Chatbot | The person typing | Answering questions from a knowledge base | You wanted work done, not a conversation |
If your job fits the first row, you don't need agent development. You need AI workflow automation services. That's cheaper to build, easier to test, and boring in the right way.
When you should not build an agent
I turn down agent projects when one of these is true:
- No one can write down "done". If the person who does the job today can't describe a good result, a model can't either.
- The tools don't exist yet. An agent needs an API, a database, or a CRM it's allowed to read and write. When the data sits in someone's head or a PDF nobody keeps updated, that's step one, not the agent.
- One wrong action is expensive and nobody reviews it. Sending money, deleting records, emailing a customer. Those get a human approval step or they don't ship.
- Volume is tiny. If the job happens five times a month, a checklist beats a system.
Saying no here is the cheapest advice you'll get. The AI automation agency page covers how we judge fit in general.
How em8 builds an AI agent
Week 1 is a diagnostic in your systems, not a kickoff deck.
- Shadow the job. Pull 20 real cases from the CRM, the inbox, or wherever the work lives. Write down the exceptions, not the happy path.
- Define tools and limits. List what the agent may read, what it may write, and what it must never touch. Set a step limit and a stop condition. Decide where a person approves.
- Build on your stack. The model runs on your API key. Orchestration runs on your n8n or the runtime you already own. Results go back into your system of record. If the keys live in our tenant, that's a vendor, not a build.
- Evaluate before launch. Run the agent against real past cases and score the output against what your best person would have done. No eval means no production.
- Log everything. Every tool call, every prompt, every decision gets logged so you can see why it did what it did and fix it at 3 a.m.
We build mostly on Claude. Emil holds the Claude Certified Architect credential, and em8 is Certified in the Claude Partner Network. The Claude automation guide covers the API side if you want to build yourself.
What AI agent development costs
Most clients start with a diagnostic ($2,500–$6,000). There's no separate rate card for agents. After that, a first agent in production typically costs from $8,000, and a retainer is $4,000–$12,000 a month.
| Engagement | Band | What you get |
|---|---|---|
| Diagnostic | $2,500–$6,000 | Week 1 in your systems: 20 real cases, tools and limits mapped, agent or workflow decided, kill list |
| First production path | from $8,000 | One agent live: tools, guardrails, eval set, logs, docs, your keys |
| Retainer | $4,000–$12,000 / month | New tools, prompt and model updates, breakage, coverage |
Model usage is a separate line you pay straight to the provider on your own key, and it depends on volume, so I won't guess a number before the diagnostic. If the scope is bigger than one production path, read how much automation costs before anyone quotes you.
Case 01: the agent we can cite
An $18M B2B company's reps spent about half their day on prospect research and hand-written cadences. We built a system with Claude on the client's n8n that does that prep. A rep ticks one box in Salesforce, the system researches the prospect, writes a personalised cadence, and puts it in SalesLoft.
- Build time: 3 weeks, live in production
- Per cadence: about 60 seconds of rep effort
- Quality: 9.7/10, LLM-judged over 20 runs
- Capacity: 175 hours a week of rep time redirected from prep to selling. That figure is a model (7 reps × 10 prospects a day × 30 minutes), not payroll removed.
Speed and quality come from the client's production logs. The full walkthrough is on Case 01.
Studio, freelancer, or in-house?
- Freelance AI agent developer: good when you have one well-defined agent and someone internal who owns it afterwards. Weak when the job crosses CRM, data, and ops.
- In-house hire: right when agents become a product line. It's slow when you need one in production this quarter.
- Studio (em8): a founder-led team that shows up in week 1 inside your CRM and n8n, ships one production path, and leaves you the keys. em8 has built and run 7 companies and worked with 50+ clients since early 2024.
If you only need the n8n side, hire an n8n expert is the cheaper route.
Red flags when you hire for AI agent development
- A demo on their data, never yours.
- No eval set, so "it works" means "it worked once".
- Agents that run on the vendor's API keys and the vendor's tenant.
- "Autonomous" with no step limit, no stop condition, and no human approval on risky actions.
- Pricing per agent without a scope of tools and cases.
Frequently asked questions
How much does it cost to build an AI agent?
Most clients start with a diagnostic ($2,500–$6,000). After that, a first agent in production typically costs from $8,000. Model usage is billed by the provider on your own API key and depends on volume.
How long does AI agent development take?
Week 1 is the diagnostic. A first production path typically follows within a few weeks. Case 01 went live in 3 weeks.
What is the difference between an AI agent and AI workflow automation?
In a workflow you fix the steps in advance and the model fills in one of them. In an agent the model chooses the next step and the tool within limits you set. Most business jobs need a workflow, not an agent.
Which model do you build agents on?
Mostly Claude. em8 is Certified in the Claude Partner Network. We pick the model per job, and the key stays in your name.
Who owns the agent?
You do. It runs on your API keys and your infrastructure, the prompts and workflows are documented, and you can stop working with us without losing the system.
Can em8 build an AI agent for us?
Yes, if the job has a clear "done", the tools exist, and a person can review risky actions. Book a 30-minute discovery call. If a workflow fits better, we'll point you to AI workflow automation services.
Book the call
Discovery is 30 minutes, free, and not a pitch. Bring one job your team hates doing. You'll leave knowing if it's an agent, a workflow, or a checklist. Book a discovery call.