Digital colleagues
that finish the job.
An agent reads the records, makes the call, takes the action across every connected system — and stops the moment a decision needs a person. Every one ships with a documented authority boundary and an immediate override.
May access
Requires approval
Never
Quick answer
What is a custom AI agent?
A custom AI agent is a software system built for one organisation that carries out a multi-step business process from beginning to end — reading data, making decisions and taking actions across connected systems — rather than only generating text in reply to a question. It differs from a general assistant in that its capability is explicitly bounded: the tools it may call, the data it may read and the actions it may take are declared in advance rather than inherited from whatever access the deploying account happens to hold. Responsible deployment requires three controls: a documented threshold above which a human must approve, an immutable log of every decision and action, and an immediate manual override that halts execution including work already in flight.
Most agent projects fail
before the first line of code.
The technology is rarely the constraint. What breaks a deployment is authority nobody wrote down, decisions nobody can explain and thresholds nobody agreed.
Failure 01
Unbounded authority
An agent inherits whatever the deploying account can reach. Nobody wrote down what it may touch, so the boundary is discovered during an incident rather than during design.
Failure 02
No approval threshold
Every action executes at the same confidence level. A routine classification and a fifty-thousand-rupee payout are treated identically until the ledger says otherwise.
Failure 03
Unexplainable decisions
The compliance officer asks why a claim was denied. Without a per-step log recording input, decision, action and outcome, "the model decided" is the only available answer.
Four deliverables, not a feature list.
Each engagement produces artefacts you keep: a mapped process, a bounded agent, an audit record and evidence it behaves before it acts.
01
Process trace and authority map
The process as it actually runs, with every decision point classified by impact and a written threshold above which a human must confirm.
02
The agent, with bounds enforced
Declared tool access, declared data scopes, injection defences tested against every untrusted input path, and refusal behaviour verified rather than assumed.
03
Audit log and kill switch
An immutable record of input, decision, action, outcome and timestamp for every run, plus an override that halts in-flight actions immediately.
04
Shadow-mode evidence
The agent runs against real cases without acting, so you compare its decisions to your team's before it touches production.
Five steps, in this
order, every time.
Autonomy is widened on evidence. Nothing is granted at the start that has not been proven in shadow mode first.
- 1
Map
Trace the process as it runs, not as documented. Count manual steps and what each costs.
- 2
Bound
Declare permitted tools, data scopes and the approval threshold. In writing, before any build.
- 3
Build
Agent, integrations and approval gates together. Audit log and kill switch in version one.
- 4
Prove
Shadow mode against real cases. Decisions compared to your team's before production.
- 5
Widen
Raise the threshold only as the record justifies. Autonomy is earned, never assumed.
Intelligence rules in the Standard, each with a verification method.
Classified Critical: human approval, manual override, no public training on your data.
Agents shipped without a kill switch and an immutable log.
Why capability must be declared rather than inherited.
Rule INT-05 requires an agent's permitted tools, endpoints and data scopes to be stated explicitly rather than taken from the credentials it happens to run under. The reason is containment: an agent that inherits full account access has no failure mode short of total exposure, and a successful prompt injection reaches everything. When capability is declared, an injection can only reach what was granted, and the blast radius of any error is knowable in advance. Verification is direct — attempt an out-of-scope action and confirm refusal.
Source: The Qawex Standard v1.0, rule INT-05 · /standard/intelligence/Answers, not brochures.
How is this different from using ChatGPT at work?
A general assistant produces text you then act on; an agent performs the action itself inside your systems. That difference requires controls a chat interface does not have: declared capability, an approval threshold, an audit log and an override. It also means the agent works from your records rather than from what someone remembered to paste in.
What if the agent makes a mistake?
It is caught at the approval gate or in the audit log. Anything above the documented threshold never executes without a person confirming, and everything below it is logged with input, decision, action and outcome so an error is traceable rather than mysterious. The kill switch halts execution including actions already in flight.
How long does an agent take to build?
A single well-bounded agent typically reaches shadow mode in weeks. What extends the timeline is rarely the build; it is agreeing where the approval threshold sits, which is a business decision rather than a technical one and is worth taking seriously.
Can it work with the systems we already run?
Usually yes. Agents connect through APIs, databases and messaging platforms, so most systems with any integration surface can participate. Where a system has none, systems integration builds one rather than replacing working software.
Will this replace our staff?
No, and we would not build one that tried. Agents take the repetitive portion and hand judgement back to a person with the context already assembled. The measurable outcome is hours returned, and every high-impact decision still requires human approval by design.
Find the process worth automating.
The strategy assessment measures where automation pays back first and what it will cost. You keep the findings whether or not you build with us.