Stage 4 · Deploy · into real workflows

AI Agents

Most of what is sold as an AI agent is a chatbot with a longer memory. A real agent takes a task from start to finish inside your systems: it reads the request, finds what it needs, acts, and hands consequential decisions to a person. We design what it may know, what it may do and where it must stop.

Talk to an AI architect See the packages

Scoped in writingActs inside your systemsStops at consequenceEvery action loggedAny model

deployed onBe Cloud

What it does

Digital workers that do one job properly, and ask when they should.

An AI agent is software given a goal, the tools to reach it and the rules it must work within. Given “resolve this order enquiry” it opens the order in your ERP, checks the courier record, reads the customer’s last two emails and writes back, then decides whether it may close the case or should hand it to a person. Unlike a chatbot, it does the work rather than describing it.

The words “one job” matter. The agents that pay for themselves do one recurring task well, with a written scope that says what they may read, what they may change and where they must stop and ask. An agent that can refund, commit to a date or write to a customer without anyone looking is a liability with a friendly name, so we write that line down before anything is built.

Where it fits

Three jobs an agent does well in a New Zealand business

Customer service

The account questions that fill a shared inbox

A distributor with a few thousand trade accounts: the agent answers order status, invoice copies and balances from the CRM and accounting system, closes what it is sure of, and hands credit questions and complaints to a person with the history attached.

Research

A client brief before the first meeting

A professional-services firm: the agent assembles a brief from your own files, the Companies Office record and the client’s public material, citing every source, so the partner walks in prepared instead of spending the morning searching.

Workflow

Enquiry to quote to booking

A trades or services business: the agent reads the enquiry, checks the calendar and the rate card, drafts the quote and books the job once a person has approved the price. The customer has an answer the same day.

How we build it

Every agent follows the five rules of the architecture

Knowledge before models. The agent answers from your CRM, your policies and your documents. A model only reasons over what those return.
Consequence needs a person. The scope lists what the agent may do alone (look up, draft, schedule), what waits for approval (refund, commit, send), and who approves.
Integrate, don’t replace. The agent works inside Microsoft 365, your CRM and your ERP. Nobody learns a new system.
Every action leaves a record. What it read, decided and did, and who approved it, in a log you could show a customer, an auditor or a court.
No vendor is load-bearing. The model behind the agent is changed in a configuration file. The scope, the knowledge and the approval design are yours.

One task, start to finish

What a workflow agent does with an enquiry at four in the afternoon

1
An enquiry arrivesWeb form or email. The agent reads it and pulls the customer record if there is one.
2
It gathers what it needsCalendar availability, the rate card, travel time from the last job.
3
It drafts the quote and a replyFrom your template, in your words, with the assumptions listed.
4
A person approves the priceOr the agent sends it unaided if the quote sits inside a band agreed in the design.
5
Booked, confirmed, recordedCalendar, confirmation and CRM updated, and every step in the log.

Compare

An agent inside a product you own, one built in-house, or a Be AI agent

	Agent feature in a product you already pay for	Built by your own developers	Be AI agent
Reaches across your systems	Only that product	If you design it	Yes
Stop line written before the build	Rarely	If you design it	Yes
Record of every action	Sometimes	If you build it	Yes
Survives a change of model or provider	No	If you design it	Yes
Sensible when	The task lives inside that one product	You have developers and an architecture	The task crosses systems and carries consequence

Questions

What people ask about AI Agents

What happens when the agent is wrong?

Sometimes it will be, which is why the design decides what it may do alone. A wrong draft costs a minute of review; a wrong refund costs money, so refunds wait. The log shows what it read and why it decided, so you fix the scope rather than guess.

Can it act without approval?

Only where the scope says so. Look-ups, drafts and scheduling usually run unattended; anything that moves money, commits you to a date or reaches a customer usually waits. The line moves as trust is earned, by decision rather than by default.

Which model does it use?

Whichever suits the task and your data rules: OpenAI, Anthropic, Google and Microsoft models are all options, as are open-source models in your own tenancy. The architecture chooses on cost, quality and where the data may go, and the choice can change later.

Is this about replacing staff?

It removes the part of a job nobody was hired for. The agent takes the re-keying and the twentieth identical question; people keep the exceptions, the relationships and the judgement. What you do with the hours it returns is your decision.

One job, done properly, with a person where it matters.

Tell us the task you would hand a good new hire in their first week and we will tell you whether an agent can do it, and what has to be designed first.

Start your AI journey →Get my AI Opportunity Score Or call 0800 23 3444