Agents
Five agents.
One job each.
A person decides.
InackGuard's AI agents do the legwork of compliance: they search, read, compare and watch. Each one follows your policy, stops at a finding and hands its work to a person, with every source attached.
all sources searched · findings cited
The loop
How an agent works
The same six steps for every agent, so every output can be read and checked the same way.
- 1
Your policy
Sources, strictness, checklists and thresholds come from the policy version in force.
- 2
Search or read
The agent searches every source in scope, reads the document or runs the rule.
- 3
A finding
Each finding is scored or quoted, with its source, page or list date.
- 4
It stops
The agent never decides. Gaps are named; nothing is shown as clear by default.
- 5
A person decides
The reviewer records an outcome and a reason, with a second reviewer where required.
- 6
Sealed
Inputs, findings, policy version and decision are fingerprinted into the record.
The five
Meet the agents
01 · Screening agent
Every source, one search.
Searches sanctions and watch lists, public-office exposure, registers, courts and the news in one pass. Every candidate is scored and tiered strong, possible or weak, with the reason; a company is matched as a company, a person as a person.
- Optional country and date of birth sharpen the score
- Sources that could not be searched are named
- One name, a file or an API call
- Takes
- A name, a file of names or an API call
- Does
- Searches every source in scope in one pass and tiers each candidate
- Hands to a person
- Strong and possible matches, each with its reason, source and list date
02 · Monitoring agent
Quiet until something changes.
Screens your list once as a baseline, then compares every list update with your records. You hear only about changes that touch a record, once, and a choice you record is remembered.
- Baseline first, so old matches are not new alerts
- One alert per relevant change
- Email, Slack or Teams, or a signed webhook
- Takes
- Your list, once
- Does
- Compares every list update with your records
- Hands to a person
- One alert per change that touches a record
03 · Document review agent
Reads the document, quotes the page.
Reads contracts, data-processing agreements, NDAs, certificates and policies against a checklist for that document type. Every point is quoted with its page, and any quote not found word for word is dropped.
- Missing items listed first, with why they matter
- Risk levels set by the checklist, not the model
- Document text deleted after 30 days
- Takes
- A contract, DPA, NDA, certificate or policy
- Does
- Reads it against the checklist for its type and verifies every quote
- Hands to a person
- A brief: missing items first, each point quoted with its page
04 · Rules agent
Tests before it alerts.
Runs transaction rules for payment firms: structuring, velocity, pass-through, dormant accounts and more. Each draft is backtested on your own past activity before it can raise an alert.
- Nothing is created during a backtest
- Reason and approver on every threshold change
- Scheduled runs and one-transaction scoring
- Takes
- Your transactions and draft thresholds
- Does
- Backtests each rule on the past 90 days, then runs it
- Hands to a person
- Alerts and cases, with the rule version that raised them
05 · Record agent
Keeps the proof.
Seals every search, finding and decision as it happens and chains each entry to the one before. Replays any past decision against its stored evidence or against today.
- Exported bundles verify offline
- Chain re-verified on a schedule
- Legal hold and logged erasure
- Takes
- Every search, finding and decision
- Does
- Fingerprints and chains each entry; replays past decisions
- Hands to a person
- A bundle your auditor verifies offline
Boundaries
What an agent never does
The limits are part of the product, written into every output and enforced by the platform.
- 01
Decide
No agent clears a name, approves a vendor, closes a case or files a report.
- 02
Hide a match
Weak leads are counted and named. Nothing is dropped quietly.
- 03
Invent a source
A finding without a source and a date does not reach you; a quote not found in your file is removed.
- 04
Read a gap as clear
A source that was down, excluded or not licensed is named as not searched.
- 05
Learn from your data
Your names, documents and payments are never used to train a model.
Oversight
Oversight built in
The controls a compliance lead and an examiner expect, on by default.
Approvals
Threshold, policy and list-scope changes need a reason and an approver who is not the proposer.
Second reviewer
Require a second sign-off on escalations or on any case type your policy names.
Workload you can show
Open cases, time to review and outcomes per analyst, exportable for the monthly report.
A full audit log
Every action by a person or an agent, in one exportable log and a feed for your own systems.
Questions
Can an agent clear a name or close a case?
No. Agents stop at a finding. Clearing a name, closing a case, escalating or filing is always done by a person, and the record shows who.
What does an agent do when a source is down?
It records the source as not searched and the output names it. A gap is never shown as clear, and a re-run picks it up when the source returns.
Where is AI used?
In two places only: a model drafts a one-paragraph synopsis of a finished output, and copies quotes out of a document you uploaded for review. A validator checks both against the sources before anyone sees them. Results, match tiers, rule hits and case outcomes come from defined rules and your analysts. An administrator can switch AI off for the whole organisation.
Can we change how an agent behaves?
Yes, within your policy: matching strictness, the sources and lists in scope, rule thresholds with an approver, and who must sign off. Every change is recorded.
Do agents train on our data?
No. Your names, documents and transactions are not used to train any model.
Put the agents on your own names.
A 30-minute walkthrough with a person, on names you choose.

