For many compliance teams, Monday morning starts with a familiar picture. A turnover increase rule fired again and again for the same customer over the weekend. A corporate application arrived with its ownership details spread across three email attachments. And an automation that has been discussed for months still isn't switched on, because the team can't yet say how the agent would decide on its own cases.
We built the October release to make that morning lighter. The highlights:
- Agents are measured in shadow mode against your analysts' decisions before they go live.
- Agents are built by describing them in a chat, with each action approved separately.
- Rules can be limited to one alert per day for the same customer and can run on a calendar.
- For corporate customers, the ownership chain can be traced down to real people and used in risk rules.
- The app moved to a new interface that matches the complead name.
AI agents
Shadow mode
The problem. Before a team lets an agent close cases, it wants to see how that agent decides. In a regulated institution "the AI said so" doesn't count as a reason, and the question is not only how often the agent is right but what it would miss. Without that visibility, automation gets postponed.
What changed. Shadow mode lets a new agent run on your real cases without acting on them.
What does the agent do in shadow mode?
It runs on the same schedule and the same cases it would handle live. It reads the case, applies its rules and reaches a decision. Then it stops. It doesn't change a status, add a note, assign the case or send a notification. It records the decision it would have made, for example "I would close this case as a false positive", next to the case.
Meanwhile your analysts keep working as usual. When an analyst decides the same case, the two decisions can be compared.
What does the comparison screen show?
The comparison screen covers Name Screening, Ongoing Monitoring, Advanced Adverse Media and Transaction and Fraud Monitoring. For the period you choose, it shows:
- The agreement rate, with a likely range next to it. With few decided cases a high rate can come from chance, so the screen shows the range the true rate probably falls in.
- Decided, agree and disagree counts. Cases still waiting for an analyst are listed but left out of the rate.
- Missed true positives: cases the agent would have cleared but an analyst confirmed as a real hit. This is the number a compliance team cares about most, so it gets its own tile.
- Breakdowns by AI decision and by rule, so you can see which part of the agent is reliable and which part needs work.
You can filter the list to disagreements only, or to missed true positives only, open each case, and export the comparison to Excel.
When is an agent ready to go live?
The screen runs three checks and shows whether each one passes:
- At least 100 cases have an analyst decision, so the rate rests on enough data.
- The agreement rate reaches the 95 percent target.
- There are no missed true positives.
The checks inform the decision but don't make it. Going live is a manual step taken from the Agentic AI task list, and that step is recorded. An agent in shadow mode doesn't use credits.
Why it matters. The answer to "should we switch it on?" comes from a rate measured on your own data, with the risky disagreements in plain view, and you can show that record in an audit.
Build Your Own Task
The problem. Setting up an agent meant drawing boxes and arrows on a flow canvas. Few people could do it, and they were rarely the people who know the work well.
What changed. Agents are now built in a chat.
How do you build an agent by describing it?
You open a chat and say what you want reviewed, how often, and what should happen with what the agent finds. There is no form to fill in. The assistant asks a short question only when something essential is missing, and replies in the language you write in.
Beside the chat you see a short summary of the agent as it stands: its name, its instructions, when it runs, what it may do on its own and how many questions it may ask you per run. You can edit the instructions by hand in that summary. The assistant builds on your wording rather than replacing it, and the agent is saved only when you save it.
What is the agent allowed to do?
Every action the agent could take on its own is a separate decision. For each one you choose between letting it act unattended and having it ask you first. The assistant recommends the smallest set of actions and leaves anything that changes a case outcome, such as closing a case, changing a risk level or adding a customer to monitoring, to your approval unless you say otherwise.
An agent can run every 15 or 30 minutes, every few hours, or daily, weekly or monthly. When it isn't sure about something during a run, it asks you, up to the limit you set. When the run ends, it reports what it did. If you read a run and say "next time, do this differently", the assistant turns that into a change to the agent's instructions.
Do you have to start from a blank page?
No. Ready-made templates cover common work, such as clearing obvious false positives, surfacing cases that have sat untouched, distributing unassigned cases, watching high-risk customers, following adverse media on high-risk names, explaining overnight transaction alerts and drafting the filing so you only need to review it. You pick the closest one and adjust it in the same chat.
Why it matters. The person who knows the process can build the automation, and the boundaries of what it may do are written in plain language that anyone on the team can read.
Transaction monitoring rules
One alert per day and rule calendars
The problem. Rules such as turnover increase can produce alerts for the same customer many times a day. The analyst opens each one, sees the same thing and makes the same decision. Some rules also only make sense at certain times.
What changed. Each rule can now carry a repeat limit and a calendar.
How does the daily alert limit work?
Each rule in a ruleset has a repeat setting with four options: no repeat limit, once per customer per day, once per customer and account per day, and once per account per day. Once the rule alerts for a customer or account, further alerts from the same rule for the same key are held back for the rest of the day. Other rules keep alerting as usual.
How do rule schedules work?
A rule can carry a calendar made of Allow and Block entries. Within an entry, every criterion you switch on must match. Between entries, any Allow entry that matches lets the rule run, and any Block entry that matches stops it, even inside an Allow window. A rule with no entries runs all the time.
- A window can wrap around: 22:00 to 06:00 covers the night, and days 28 to 3 cover the turn of the month.
- Times follow your organization's time zone and are matched against the transaction's own date, not the moment it reached the platform.
- Rules that depend on a scheduled rule follow the same calendar.
A rule outside its window isn't evaluated at all, and the reason is recorded on the transaction. When someone later asks why a rule didn't fire, the answer is in the transaction detail.
Why it matters. The alerts left in the queue are the ones that say something new, and the reason a rule didn't fire can be traced afterwards.
Corporate customers
Ownership Structure
The problem. With a corporate customer, the real question is who actually stands behind the company. That information often arrives spread across spreadsheets and email attachments, and indirect shares are worked out by hand.
What changed. A new Ownership Structure tab on the corporate customer record keeps the whole chain in one place.
How is the ownership chain built?
On a corporate customer, you add direct owners and their shares. When an owner is itself a company, you add that company's owners under it, and so on down the chain. Owners can also be loaded in bulk from an Excel or CSV file together with company officers. Every add, edit and delete is written to the customer's case history.
The system works out the rest:
- Indirect shares are calculated by multiplying the percentages along the chain. When the same person holds shares through more than one branch, the shares are added together in a single row.
- Ultimate beneficial owners are listed against a threshold you pick: 10, 15, 20, 25 or 50 percent, with 25 as the default. You can also mark a person as a UBO by hand.
- Consistency checks flag a structure whose direct shares add up to more than 100 percent, and show whether ownership is determinable and how much of it can't be attributed to anyone.
- Company officers are managed separately from ownership, so ending one relation doesn't end the other, and they can be included in screening.
An ownership chart draws the whole structure. You can click a node for its details and double-click a linked customer to open it. The tab also shows when an owner or officer is shared with another customer.
Which rule conditions use ownership data?
Customer risk rules can use the number of UBOs, the number of owners and of corporate and individual owners, the highest direct and indirect share, the depth of the chain, the unattributed share, whether ownership is determinable, owner nationalities, the ages of the youngest and oldest owners, whether shares add up to more than 100 percent, whether any owner or officer is sanctioned and how many customers share a party with this one.
Why it matters. Ownership data lives inside the customer record, where it can be screened, charted, audited and used in risk rules, instead of sitting in a spreadsheet beside it.
Interface and teamwork
New interface
At the end of September we became complead, and in October the app took on a look that matches the name. Menus, tables, filters, confirmation dialogs and notifications follow the same layout on every screen. What you learn on one screen applies to the others, new analysts get up to speed sooner, and a short in-app tour walks through the changes.
Smaller changes you'll notice during the day
- Dashboard: Widgets can be resized freely and placed side by side. A new KPI widget shows progress against a target at a glance, and you can ask Nova about the data in any widget.
- Notes in one place: The Memo Reports screen lists a customer's notes from every module, with filters for customer, date range and AI-generated notes, and an Excel export. Teammates can be tagged in notes with @.
- Notification digests: Emails such as approval notifications can arrive instantly, every 1, 2, 4, 8 or 12 hours, or as a single daily digest at 08:00 in your organization's time zone.
- API permissions: API keys can be given Read, Scan and Manage permissions separately for each product. Existing keys keep working as before.
- Bulk work: Previously screened records can be added to ongoing monitoring in bulk, accounts can be added from the interface, KYC form applications can be downloaded as Excel or PDF, and forms can be duplicated.
If you already use complead, the details of each update are in the Help Center inside the app. For earlier months, see the Product Updates category.
