Skip to content
Civware

Agents for multi-step work,with people in control

An agent is AI that works through a task step by step, the way a careful colleague would. It reads the request, looks up what it needs in your systems, proposes the next action and records what it did. People approve the steps that matter.

Group
Build
First step
A pilot for one job
You decide
Go or no-go, after the pilot
Works with
Your CRM, ERP, email and documents

When an agent is the right fit

Agents suit work that is repetitive but not identical. If the steps never change, a plain automation is simpler and cheaper, and we'll say so.

  1. 01

    The same multi-step task is done by hand many times a week, following rules people could write down.

  2. 02

    The work spans several systems, so people copy details from one screen into another.

  3. 03

    Most cases are routine, but some need judgment, and you want those sent to a person.

What we build

  1. 01Request agents that read emails or forms, gather the facts and prepare the reply or change for approval.
  2. 02Back-office agents that keep records in step across your CRM, ERP and ticketing tools.
  3. 03Reconciliation agents that compare records, flag mismatches and draft the fix for someone to approve.
  4. 04Research agents that pull facts from your documents into a short brief, with every source linked.
  5. 05Follow-up agents that chase missing information and keep a case moving until it's complete.
  6. 06A review screen where your team sees each proposed action and approves, edits or rejects it.

How it works

Before an agent is built, we agree a written list of the tools it may use and what it may do in each. Nothing outside that list is reachable.

  1. 01Each tool has a level of access: read, draft or change.
  2. 02Changes wait in the review screen until a person approves them.
  3. 03Every step and tool call is logged, so you can see what the agent did and why.
What the agent can reach A task goes to the agent. The agent connects to four tools, each with its own level of access: read only for the CRM and documents, drafts only for email, and changes to the order system only after a person approves. Every step is written to a log. FIG. 01WHAT THE AGENT CAN REACH TASK AGENTplans the stepsuses 4 tools READ ONLYREAD ONLYDRAFTS ONLY APPROVAL FIRST CRMDocumentsEmailOrder system LOGevery step and tool call recorded
Fig. 01 An agent reaches only the tools its task needs, at the access it needs. Changes wait for a person, and every step is recorded.

Safeguards

Agents act, so their risks are different from a chatbot's. These are the ones we plan for, and how each is handled.

Acting on a wrong assumption
Each proposal shows the facts it relied on, and changes that matter wait for a person's approval.
Doing more than it should
The agent runs under its own named account, with the least access its task needs and no tools beyond its agreed list.
Following hidden instructions
An email or document can contain text written to redirect an agent, known as prompt injection. Agents are designed to treat what they read as information, not instructions, and we test for this before launch.
Small errors that repeat quietly
Every step is logged, results are checked against the measure you agreed, and an agent can be paused at any time.
Costs that creep up
Usage limits and cost alerts are set before launch, so a busy week can't run up an unexpected bill.

How to start

First engagement · Pilot

One agent for one job, built on your data

The pilot delivers:

  • A working agent for one job, running on your real data in a test setup
  • The review screen your team uses to approve, edit or reject its actions
  • An evaluation sheet showing how it did on your own examples, case by case
  • A written go or no-go recommendation, with what it would cost to run
Price
Fixed, agreed before we start
Length
Usually 4 to 8 weeks
Decision pointAt the end of the pilot, you decide: launch with a small group, adjust, or stop.
Book a call

Not sure which job to start with? Begin with a Readiness Assessment.

Questions about agents

What's the difference between an agent and an automation?

A plain automation follows fixed rules, the same way every time. An agent can handle requests that vary, choosing which steps to take within the limits you set. If fixed rules would do the job, we'll recommend the automation.

Which steps need a person's approval?

You decide, and we write it down before anything is built. It's usually any step that changes a record, spends money or contacts a customer. As results build trust, you can choose to approve less.

Can an agent work with our existing systems?

Usually, through each system's API, the connection software uses to talk to other software. Where a system has no API, we'll explain the options and their trade-offs before you decide. See AI Integration.

What happens when the agent gets something wrong?

It's designed to stop and pass the case to a person when details don't match or it isn't sure. Because every step is logged, a mistake can be traced, fixed and tested against your examples before the fix goes live.

How do you measure whether it's working?

Against the measure agreed at the start, such as time per case or cases redone. It's checked on your own examples before launch and monitored after. See Evaluation & Monitoring.

Who owns the agent?

You do: the code, prompts, configuration and documentation.

Tell us one job you'd like AI to do

We'll tell you honestly whether AI is the right fit, and what a sensible first step looks like.

One job is enough to start.

We use your details to reply to you, as described in our privacy policy.