AI agent for a business process
It takes the routine steps itself, works to the rules, and hands a person only the exceptions.
- Buildfrom $2,860
- Timeline4–6 weeks
- Supportfrom $570/mo
Who this suits: Processes done by hand every day to the same rules, where the rules are written down nowhere except in somebody’s head.
What is in the first version
- one whole process
- explicit rules
- exception queue
- decision log
The figures above are the package this sits on: AI agent or process automation. Anything past its edges the calculator adds before the work starts, not after.
How the path goes
- 01
A task reaches the agent
The trigger is not a button: a new email, a new row, a time on the clock.
- 02
It walks the steps
Every step is either a fixed rule or a decision handed to a model, and the line is drawn beforehand.
- 03
An exception goes to a person
It stops and arrives as one message: the case, what the agent has done, what it is missing.
What decides whether this works or annoys people
The exception list matters more
A process described the way it runs when everything goes right breaks on a refund, a duplicate, an empty field.
Where a person has to confirm
An action that cannot be undone, an email sent or money taken, goes through a confirmation decided before the code.
A silent agent looks fine
Stopping without an alert goes unnoticed, so the agent checks in on a schedule and every step has a waiting limit.
What to measure once it is live
- tasks closed without a person
- exceptions per day
- time from event to result
- model cost per month
Why this comes out faster
The shape of this one is known: the states, the edge cases and the things that usually go wrong have been decided before. Nothing here is a template, and the saving is not in your half of the work. It goes into your process, your content and the systems this has to talk to, which is the part nobody can have solved in advance.
A likely stack for this
Picked against the task when we scope it, not decided in advance. This is the shelf it usually comes off.
- Agent and modelsPythonFastAPIOpenAIAnthropicpgvector
- Steps, queues and retriesTemporaln8nRedisPostgreSQLPlaywrightDocker
- Where the result landsTelegram Bot APIGrafanaS3
Price it yourself, right here
Five steps, and you can see the number without leaving a contact. The estimate accounts for the kind of work, what you already have and what it will need inside.
These bills do not come from us
- AI model calls, billed by the provider
- hosting and servers
- subscriptions to your CRM, mail, telephony
- the domain and sending address for email
Asked before the first call
- What happens when the agent gets it wrong?
- Every action is written to a log that shows which step and which rule produced it. Irreversible steps sit behind a human confirmation, so a mistake costs a redone task, not money.
- Does it need access to our systems?
- Yes, and only to the systems the process touches. Keys are created in your own accounts, scoped to specific actions, and you can revoke them at any moment without us.
- Do we actually need AI here, or a plain script?
- Often a script is enough, and we say so. A model is needed only where the input is unstructured: a request in plain words, a document laid out differently each time.