Answer page
AI agents for businesses: what they do and what humans decide
An AI agent is a program that receives a goal, works out the steps itself, and uses tools to get there: reading files, querying systems, writing records. Unlike a chatbot it does not only answer, it acts. APPIQ Solutions builds agents with fixed approval gates so a person decides before anything reaches the outside world.
How is an AI agent different from a chatbot?
A chatbot answers questions. It has text and maybe a knowledge base, nothing more. Classic automation runs a fixed chain of steps, the same one every time.
An agent gets a goal instead, such as: prepare this request for sales. From there it picks which tools to use in which order, checks its interim result, and corrects course when a query comes back empty.
What is an agent allowed to do in my company?
Exactly what you grant it. Permissions get assigned explicitly, never inherited. You set three layers before the agent runs for the first time.
- Tools: which systems it may address, for example read the CRM, write to the calendar, read-only on invoicing.
- Data: which folders, mailboxes, and records it sees. This follows your existing permission model.
- Gates: at which points it has to stop and wait for a human decision.
What does gated autonomy mean exactly?
Gated autonomy means the agent works on its own up to a gate, where a person decides. You place gates wherever a mistake gets expensive.
An example: the agent reads a customer request, finds the matching reference, fills the quote, and puts it in front of sales. All of that runs untouched. Sending is the gate, and a person clicks there.
The advantage over full automation shows up in liability. Whoever approved is on record, with timestamp and name.
Which agents actually run in a business?
The useful agents are unglamorous. Each one attaches to a clearly bounded flow and removes manual work there.
- Request agent: reads the inbox, classifies, creates the contact, prepares the reply.
- Knowledge agent: answers questions from your manuals, contracts, and minutes, and cites the passage.
- Quote agent: fills the template, pulls prices from the price list, flags deviations.
- Reporting agent: gathers figures from several systems and files the monthly report as a draft.
How secure are AI agents?
The serious security questions around agents concern permissions and untrusted input. An agent with write access to a production system is a user account, and we treat it as one: its own account, the narrowest permissions that work, a log of every call.
The second question is prompt injection, where someone smuggles instructions to the agent through a mail or a document. The defence is to treat foreign content as data and to keep any tool with real effect behind a gate.
Then comes privacy: where the model runs, what leaves your building, how long transcripts are kept. We settle that before the build, not after.
How does the APPIQ Cognitive Engine orchestrate agents?
ACE is our own platform, a desktop application backed by a service that runs around the clock on a server. It coordinates specialised agents across the software lifecycle: analysis, implementation, review, acceptance.
Two things separate it from a single assistant. First, memory in the form of a knowledge graph, so project knowledge survives between sessions. Second, a check against technical debt, the kind that piles up when models produce code nobody understands any more.
The same gates apply throughout: you see what an agent proposes and decide what gets accepted.
Frequently asked questions
Does an agent need access to all our data?
No, and it should not get it. An agent receives exactly the sources its task requires. The request agent sees the inbox and the CRM, not the accounting system.
Can an agent spend money or place orders on its own?
Only if you explicitly allow it, and we advise against it. Anything with a payment effect belongs behind a human approval.
What happens when an agent gets stuck?
It stops, reports its state, and hands over to a person. An agent that guesses its way forward when uncertain does more damage than one that halts.
Can I trace why an agent decided something?
Yes. Every run logs the tools used, the sources read, and the result of each step. That lets you trace an output back to where it came from.
Do agents work without an internet connection?
That depends on the model. With a locally hosted model, yes. With a vendor model, no. We make that call together with your requirements on privacy and answer quality.
How many agents does a mid-sized company need?
One, to begin with. A single agent on a process that holds up in daily work beats five half-finished ones. Others follow once the first has proven itself.