AI agents (GaaS)
Most software waits for someone to use it. An agent does the job itself: it reads the invoice, checks it against the order, updates the system and only asks a person when something looks wrong. That is powerful, and it is also a new way to get breached, because the agent holds real access to real systems. So we build every agent around one rule: it can only do what you would let a careful new employee do on their first day, and every step it takes is written down. Run for you as a service, this is what the market now calls GaaS: AI agents as a service.
01In detail
What our agents do
Process documents
Invoices, applications, forms and contracts read and entered into your system. Anything unclear goes to a person, not a guess.
Answer customers and staff
Answers drawn from your own policies, manuals and records, with a link to the source, and a hand-over to a person when needed.
Sort and route work
Emails, tickets and complaints sent to the right desk the first time, with a draft reply ready.
Update your systems
Records created, orders raised, visits scheduled, through the same connections your apps already use. Anything costly waits for approval.
Run on your own servers
Open AI models on your hardware when data is not allowed to leave the building. Nothing goes to an outside AI company.
Work with what you have
Agents plugged into the PHP, Laravel, Node or Python systems, ERP or CRM you already run. No new tool nobody opens.
02Process
How it runs
Where data is sensitive, the agent runs on open models on your own servers, so nothing is sent to an outside AI company.
Pick one job
Week 1
One repetitive task your team does by hand. We record how long it takes today and how often it goes wrong.
Map access and risk
Week 2
What the agent may read, what it may change, what it must never touch, and where a person signs off.
Build on real data
Weeks 3 – 5
Built against your actual documents and systems, including the messy ones.
Test, measure, launch
Week 6
Security tests, then side by side with the manual process. It goes live only if it clearly does better.
03Deliverables
What you get
- A working agent on one real task, measured against how it is done today
- A written access map: what it can read, change and never touch
- Approval steps for anything costly or irreversible
- A full activity log your team can search
- Security test results, including attempts to trick the AI, with a re-test
- The running cost per task, so you know what each job costs
Tools we reach for
- OpenAI
- Claude
- Llama
- Qwen
- Python
- Laravel
- Node.js
- PostgreSQL
- Docker
Security checks
Six security checks before every launch.
Before anything goes live, we test it the way an attacker would, then write down what we found and how we fixed it. The same six checks run on every project, from a one-page website to a government portal.
- 01
Scan for known weaknesses
We check your code, the libraries it uses and the server for weaknesses attackers already know how to exploit.
OWASP Top 10 · CVE database
- 02
Try to break in
Our engineers attack it the way a real attacker would: logins, permissions, file uploads, pages that should be hidden.
OWASP ASVS
- 03
Check every public link
Every public page and address is crawled for files, folders and test pages that should never have been online.
OWASP Top 10
- 04
Look for malware
The code, the server and anything users can upload are checked for malware and hidden backdoors.
Before every launch
- 05
Try to trick the AI
For AI agents, we try to make them leak data, ignore their instructions or use access they should not have.
OWASP Top 10 for LLM Applications
- 06
Write it all down
Every finding, how we fixed it and the re-test that proves it, in a report your auditors can read.
ISO/IEC 27001-aligned process
Our process follows ISO/IEC 27001 controls. Prospiral is not ISO certified. VAPT means vulnerability assessment and penetration testing: the first two checks above.
05Questions
Questions we get asked
Software that uses an AI model to finish a task on its own: read something, decide what to do, use your systems to do it, and report back. Unlike a chatbot it takes action, which is why we build in limits and approvals from day one.
With ordinary software you pay for a tool and your team does the work inside it. With an agent, the work itself gets done and your team checks and approves it. Your existing systems stay; the agent works through them.
Only with limits. Each agent gets its own account with the least access possible, a person approves anything costly or irreversible, and every action is logged. Before launch we try to break it ourselves.
No. We use provider settings that exclude your data from training, or run open models on your own servers when that is not enough.
Usually about six weeks for one task, from the first call to a measured result. If it is not clearly better than the manual process, we say so and stop.
Start a conversation
Pick one job your team hates.We’ll build the agent for it.
Tell us the task, how often it happens and which systems it touches. In a 30-minute call we will tell you honestly whether an agent is the right answer, and what it would take.