Service
Custom AI assistants that carry out multi-step tasks with your tools and your approval rules
Some tasks need several steps and a decision at each one: look up a record, read a document, draft a message, update a system. We build an AI assistant for one such task, with a fixed set of tools, limits on what it may do and points where a person must approve.
- Built for
- Companies that have already automated simple workflows
- Scope
- One assistant, one task, a fixed set of tools
- Staff role
- Approve actions that send, pay or delete, and review the log
- Price
- Request a quote
For a judgement-heavy task that still follows a procedure
A fixed workflow handles work where every step is the same each time. A multi-step AI assistant suits the next kind of task, where the steps are known but the order and the tool depend on what the assistant finds.
Examples are preparing a quotation from an enquiry, matching orders against payments, or researching a prospect before a sales call. Each one can be written down as a procedure a new staff member could follow.
- Companies that have already automated simple workflows
- Teams with a multi-step task that follows a describable procedure
- Owners who want every action logged and reviewable
How it works
Task
One task is chosen
One task with clear inputs, outputs and success criteria. A narrow job is easier to test and to trust.
Permissions
Tools and permissions are defined
The systems the assistant may read, the ones it may write to, and its spending and action limits are agreed in writing.
Build
The assistant is built
An AI model, the tool connections and the guardrails are assembled in n8n or custom code.
Test
Test cases are run
Your test cases are run against expected results, including failure cases and misuse cases.
Approve
Approval steps are placed
A person approves before any action that sends a message, makes a payment or deletes a record.
Monitor
It runs with a full action log
Every lookup, draft and update is logged, so the task owner can review what the assistant did and why.
From an enquiry to a quotation a person approves
- Enquiry arrives (Email or form), Automated
- Customer record looked up (CRM, read only), Automated
- Price list read (Sheet or accounting), Automated
- Quotation drafted (Draft only), Automated
- Staff approve before sending (Approval), Your staff
- Each step written to the log (Action log), Automated
What we need from you and what we handle
We need from you
- A clearly described task, written as a procedure
- Access to the systems the task touches
- Test cases with expected results
- An owner who reviews the assistant's work
We handle
- Writing the task definition and success criteria with you
- Tool connections with the least access the task needs
- Guardrails, limits and approval steps
- Running the test cases and reporting the results
- The monitoring log and written handover notes
What a first project includes and what it leaves out
Included
- One assistant for one task
- Tool connections
- Guardrails and approval steps
- Test results
- A monitoring log
Left out
- Open-ended assistants asked to do anything
- Payment or deletion without a person approving
- Tasks with no stable procedure
Limits we state plainly
- An assistant can take a wrong step or misread an instruction. Actions that cannot be undone need a person's approval.
- Text from emails and web pages can contain instructions that try to redirect an assistant. Inputs are treated as untrusted and the tools it can use are limited.
- AI usage costs rise with task length. They are estimated before launch and billed by the AI provider.
- An assistant acting on personal data falls under the Data Privacy Act of 2012 (RA 10173). Decisions with legal or financial effect on a person stay with staff.
Tools and channels involved
Claude, OpenAI or Gemini API with tool use, n8n or custom code, and connections to your CRM, email, sheets, accounting and internal databases.
What an assistant can read and update depends on each system's API and plan. See the GoHighLevel, HubSpot, QuickBooks, Odoo and Google Sheets integration pages.
What it costs
Cost depends on the scope of the task and how many systems the assistant has to work with.
Custom AI assistant development
Request a quote
Quoted per project after a short review of your channels, volumes and systems.
Request a quoteWhat moves the price
- Scope of the first project
- Number of channels or systems
- Number of steps and approval points
- Test cases to cover
- AI usage, billed by the provider
- Support level
Questions we get asked
Which tasks suit a multi-step AI assistant?
Tasks with several steps, a describable procedure and a result a person can check. Preparing a quotation, matching orders to payments and researching a prospect are typical. A task that changes every time, or that nobody can describe, is not a fit.
How are permissions and approvals set?
Before the build, we agree in writing which systems the assistant may read, which it may update and what it may never do. Approval steps go before any action that sends, pays or deletes.
How is its work logged and reviewed?
Every tool call, draft and update is written to a log with the input that led to it. The task owner reviews the log, and failures raise an alert.
Is this the same as a chatbot?
No. A chatbot answers questions in a conversation. This assistant works through a task across your systems. If you need answers from company documents, see the custom AI and integrations service.
Who owns the assistant after handover?
You do. It runs in accounts in your name, with written notes on its tools, limits and approval points.
Next step
Describe the task as you would explain it to a new hire.
Tell us the steps, the systems involved and where a person must decide. We will say whether a multi-step assistant fits or a simpler workflow would do.