AI Agent Setup
Not a general chatbot — connected to your systems, bounded in authority, with a record of what it did.
- Company data
- System integration
- Bounded authority
Making AI useful inside a company starts not with choosing a model but with connecting it to the right data and the right authority. That connection is what we build: the agent reads your documents, works inside your systems, and can show afterwards what it did.
The work first, the model second
We start from the work, not the model: which question is asked over and over, which form is filled in by hand, which piece of information is hunted for in five places. We write down together what is to be automated, and decide up front how far the agent may decide on its own and where it must ask a person.
- Identifying the repeating work
- Writing down the agent's decision boundary
- Marking the steps that need human approval
- Defining what success means before we start
Answers grounded in your own documents
The agent speaks from your documents rather than general knowledge gathered from the internet: product catalogue, price list, technical specification, contract template, internal policy. It states which document an answer rests on, and it is set up to say it does not know rather than invent something the documents do not contain.
- Indexing company documents
- Citing the source of an answer
- Refreshing when a document is updated
- Keeping confidential documents out of scope
An agent that does work, not just talks
An agent's value is not in answering but in attaching that answer to a task. We connect it to the systems you already run: open a record in the ERP, find the customer in the CRM, route an incoming message to the right person, read a form on your site and prepare a draft quotation. If a system cannot be connected, we say so at the start.
- ERP, CRM and accounting connections
- Messaging and form channels
- Website and store content
- Early notice when a system cannot be connected
Authority, cost and record
The agent cannot do everything on its own: what data it may see, which actions it may take without approval, and how much it may spend in a month are all bounded. Every action it takes is recorded — which question, which source, which answer — and can be read back later. To correct a behaviour you dislike, you first have to be able to see what happened.
- Bounded data and action permissions
- Spending ceiling
- Action log and retrospective review
- Approval-gated actions kept separate
Delivered working, not set up and left
Commissioning happens on real work: the agent runs under supervision first, its answers are read together, and wrong turns are corrected. We show your team how to ask it and how to correct it. We can keep operating it, or you can take it over entirely.
- Supervised trial period
- Reading answers together and correcting
- Team training
- Ongoing operation or full handover
An agent that says it does not know is worth more than one that invents
The agents we build cite their source and do not invent what the documents do not contain — because a system that invents once is never trusted again.
- Process discovery and scoping
- Indexing of company documents, with source citation
- ERP · CRM · accounting · messaging integrations
- Permission boundaries, approval steps and spending ceilings
- Action logging and retrospective review
- Supervised commissioning and team training
- Model chosen to fit the task; setup remains open to a change of provider
Let's Talk About Your Project
Contact us for free technical consulting.