Agentic Automation: AI Agents for Business Processes

10 agentsbuilt and proven inside Complexica first
ISO 27001:2022certified information security management
25+ yearsof applied AI research
Model-agnosticworks with any frontier model

The basics

What is agentic automation?

The difference between types of AI automation comes down to one question: who decides the next step?

01 · Assist

Chat

You ask, it answers, and you still do the work.

Next step decided by you

02 · Automate

Rule-based workflows, such as RPA

The same fixed steps run the same way every time, such as keying each order email into the ERP. When the input or the process changes, the automation stalls.

Next step decided by fixed rules

03 · Operate

Agents

An agent is given a goal. It plans the work, acts in your systems, checks the result and keeps going until the job is done.

Next step decided by the model, within the limits you set

What an agent really isAgent = Model + Harness + Enterprise

Model: the reasoning. It reads context, plans and chooses the next step.
Harness: the system around the model. It connects to your systems, remembers state, tests, rolls back and stops.
Enterprise: your rules. Permissions, audit, human approvals and change control.

Who it is for

Is agentic automation a fit for your process?

Agents earn their keep on complex, governed work. It is worth a conversation if your process looks like this:

The work has many steps and spans several systems, such as your ERP, CRM, inbox and shared files.
Some steps need a person's approval, and every action needs a clear audit trail.
Rule-based automation, such as RPA, breaks when the input or the process changes.
It has to run inside your IT, under your permissions and change control.

Sound familiar?

Talk us through one process. We'll tell you honestly whether an agent is the right tool.

Book a discovery call

Use cases

Proven on our own processes first

Every agent below runs on Complexica's own work. Each one shows where the same pattern fits in your business.

Support: from ticket to fix

The problem

Before anyone can reply to a support ticket, someone has to dig through the code, the database, the logs and the files to find the cause.

What the agent delivers

Support becomes a chain of agents. The Support Agent investigates in a sandboxed, read-only environment and works the diagnosis through with the engineer in Slack. If a code change is needed, the Dev Agent makes the approved fix and the QA Agent checks it before anything goes live.

Same pattern in your business: customer self-service, and any request that has to be investigated across several systems before someone can reply.

Support Agent
Investigates the ticket, diagnoses the cause and drafts the reply.
Dev Agent
Turns the approved fix into a code change. Never publishes itself.
QA Agent
Checks the root cause, the change and the reply before anything goes live.

Human gate: an engineer approves before anything changes. Judgement in the loop.

Legal: contract questions and reviews

The problem

A legal question means hours of searching the drive. Reviewing a contract means reading and comparing every clause by hand, then rounds of rework over email.

What the agent delivers

Ask the Legal Agent in Slack. It checks signed agreements, redlines and emails and answers consistently, and reviews a draft against your standard positions to return a memo and a tracked-changes redline. Every conclusion links back to its source; high-risk and commercial decisions stay with the authorised reviewer.

Same pattern in your business: supplier document checks, invoice matching and approval drafting: any document checked against your own standards.

Dimension	Before	With the Legal Agent
Trigger	Email, meeting or attachment	Natural-language request
Information gathering	Manual search and follow-up	Agent retrieves relevant sources
Analysis	Lawyer reads and compares	Agent compares and flags exceptions
Quality control	Reviewer checks manually	Automated checks and a human gate
Rework	Email and redline cycles	Agent revises immediately
Output	Memo, redline or email	Memo, redline and source trail
Elapsed time	Dominated by queues	Dominated by approval

The goal

Move legal effort from document handling to judgement, exceptions and approval.

Website changes from one Slack message

The problem

One web change took eight steps across two teams, with hidden costs in waiting, context switching, repeated explanations and inconsistent testing.

What the agent delivers

A change is now one Slack message. The agent edits the site, checks every affected page visually and functionally before and after it goes live, and replies in the thread.

Same pattern in your business: living SOPs, reporting and commentary: content that has to be updated, checked and published reliably.

1
Locked in
Our site lived inside HubSpot's CMS.
2
Mirrored
The agent rebuilt it as native HTML, testing every page.
3
Ours
The code now lives in our repository, hosted on our side.
4
Wrapped
The Website Agent now handles every change.

The current complexica.com home page, maintained by the Website Agent

# cpx-website-agent

CPX team

@CPX Website Agent which pages are near-duplicates of each other?

CPX Website Agent Agent

14 pairs of near-identical pages, competing in Google.

Cyber security: prove before you report

In development

The problem

Knowing which security vulnerabilities are real, and fixing them without breaking anything.

What the agent delivers

The Cyber Agent finds where you're exposed, proves which issues are real with red-team testing, and brings evidence and a fix plan to your engineer. Only approved issues are patched, and each fix is re-tested until it holds.

Same pattern in your business: any check where findings must be proven, and approved, before anyone acts on them.

1
Scan
Finds where you're exposed.
2
Validate
Red-team testing proves which issues are real.
3
Report
Evidence for each issue, plus a plan to fix it.
4
Your engineer approves
Nothing is patched without a yes.
5
Fix
Writes the patch for each approved issue.
6
Verify
Re-tests until the fix is proven to hold.

Legacy software, brought up to date

In development

The problem

Older desktop applications that are hard to read and hard to change.

What the agent delivers

The Legacy Management Agent updates and maintains legacy codebases. In this example, a car-distribution optimisation system became a modern web app that is clear at a glance.

Same pattern in your business: risky spreadsheets rebuilt as governed apps, with approvals and audit trails.

Before: the legacy desktop car-distribution system, a dense map of routes with small charts — BeforeDesktop app, hard to read and hard to change

After: the same car-distribution system as a modern web app with clear metric cards and a route map — AfterModern web app, clear at a glance

Also in discovery with customers

Knowledge and answers

Ask the business
Living SOPs
Customer self-service

Documents and approvals

Supplier document checks
Approval drafting
Invoice matching

Operations and markets

Operations monitoring
Market and disruption watch
Process mining

Finance and control

Reporting and commentary
Spreadsheets to governed apps
Contractor compliance

Have a process like these? Talk it through with the team that builds the agents.

Book a discovery call

What we do and how we work

One team, from first conversation to a running agent

We design, build and operate the agent end to end, and keep improving it once it is live.

1
Discover
We look at your processes with you and pick the ones where an agent can take on multi-step work.
2
Design
We decide which systems the agent reads and acts in, what it may do, and where a person approves.
3
Build and validate
We build the agent and the system around it: connections, memory, multi-step execution, testing, rollback and the ability to stop. Every release is versioned and evaluated before it goes live.
4
Run with people in control
We operate the agent inside your IT, with permissions, approvals for high-risk steps, and traces and alerts on every run.
5
Learn and improve
A scheduled review turns finished work into memory, so each job makes the next one easier.

Learns from every job

A scheduled review turns finished work into memory

The work

Agent and engineer work the ticket

captures

Logs

Tickets, threads and branches

informs

reviews

Memory

Best practice, client knowledge and navigation

consolidates

Review

A scheduled reviewer consolidates the lessons

It improves fastest when engineers say what was wrong, and why.

Why Complexica

Built on research, proven on ourselves

25+ years of applied AI research

Our agents build on more than 25 years of applied AI research at Complexica.

Enterprise integration experience

Agents that sit inside IT, not beside it, connected to your systems under the permissions, audit and change control you already run.

Model-agnostic

We work with any frontier model, so you are not tied to one AI provider.

Ten agents, built and proven inside Complexica first

Support Agent
Diagnoses tickets and drafts replies
Dev Agent
Turns approved fixes into code
QA Agent
Verifies fixes before anything goes live
CPX Website Agent
Makes and verifies changes to complexica.com
Commercial Agent
Drafts and reviews scope proposals
Legal Agent
Drafts and reviews legal contracts
Demo Design Agent
Makes and verifies changes to demos
Customer Support Agent
Diagnoses and resolves customer issues
In development
Cyber Agent
Scans and validates security vulnerabilities
In development
Legacy Management Agent
Updates and maintains legacy codebases
In development

Who runs the unit

Spiros KayeHead of AI AutomationView profile →

Adam MichalewiczManager, AI AutomationView profile →

Security and governance

Governed from the first deployment

Every deployment is governed by Complexica's ISO 27001:2022 certified information security management system, with human oversight and controls appropriate to each process.