Show Your Work: Every Transaction Has a Trail

nov. 18, 2026
Partager

Every trader learns the phrase early. Show your work. The process matters as much as the answer. In financial services, an answer without a trail behind it is a liability waiting for someone to ask a follow-up question.

That standard applies to money. It rarely applies to data. Every transaction carries a trail: who approved it, when, and against what. The data trail is the one nobody keeps. It’s the record of where a piece of data came from, what it is today, and what or who has touched it since.

The request comes in, and then the scramble starts

A former client emails the bank asking for every piece of personal data it holds on them. GDPR gives them the right to all of it, usually within one month: the data itself, where it came from, who received it, and how long it'll be kept. That covers the CRM record and the KYC file. It also covers the call recordings MiFID II required the bank to keep, and the spreadsheet a relationship manager exported in 2021 and never deleted.

Materiality requests and incident investigations start the same way. The clock runs, and someone rebuilds the answer from scratch. A typical materiality request takes about 80 hours for two people. That happens several times a quarter, every quarter, indefinitely.

Finding the data is the hard part. Picture a bank with thousands of data stores, most built long before anyone imagined a request like this. None were designed to be searched the way you search the web. The usual workaround is to limit the search to systems known to hold customer records. That breaks down when the same client appears under one name in the CRM and a different ID in the KYC file.

Eighty hours goes into reconstructing something the institution once knew and lost track of: where the data lives, what it actually is now versus what someone assumed it was, who's touched it since. None of that should take an 80-hour excavation. It does, because nobody's kept the data trail current.

Retention creates a second version of the same problem. Say a bank’s policy is to keep security logs for a year. To avoid running past that limit after a technical delay, the team purges well before the one-year mark. The harder question is which of those logs captured personal data by accident, such as a credit card application pasted into a log file. Nobody can answer that without looking.

Do this work continuously instead of on demand, and the cost changes shape entirely. Audit and compliance costs typically drop 60 percent once the answer already exists instead of getting rebuilt every time someone asks. Across an institution's full data footprint, that same discipline, applied everywhere at once, adds up to hundreds of millions in avoided cost and reduced exposure. Someone can finally point to the data and say exactly what it is. That's worth more than any abstract security posture claim.

AI agents add new names to the trail

Agents read data and act on it at machine speed, often under a service account that was set up for something else. An agent assembling a credit memo could pull from a data warehouse that renamed half its fields during a migration two years ago. It answers with the same confidence whether those fields still mean what they used to or not.

Banks already have a control for decisions that matter. Maker-checker means one person prepares and a second approves, which catches typos and insider fraud, though it can still fail when two employees collude. It works differently with agents. You can sit a person down and ask why they made a call. An agent can produce a confident explanation that has nothing to do with what it did, then hold to it. It also lacks the instinct that says a file looks a year old when the team refreshes it every six months.

Accountability still lands on a named human, and regulation trails adoption by years. An examiner will ask a person to explain a decision made by a system that can’t explain itself. The data trail is the only evidence that person will have.

 

Each agent is also a new line in the data trail that an examiner can ask about. Which agent read this client’s records? What did it do with them? Who approved its access? Plenty of teams can’t answer those questions for employees yet. Agents only make those questions multiply.

The institutions moving fastest on agents will be the ones that can answer first, because their data trail already exists.

Four steps to a trail you can show

  1. Find and classify what you hold - Start with the systems an examiner would ask about first: client records and anything copied out of them. Label each dataset by type and also by whose data it is.
  2. Map who and what can reach it - Human or machine, service accounts and agents included. Remove access nobody uses.
  3. Record what happens next - Keep a running log of who and what touched each dataset, what it did, and where the data moved. That log is the trail itself. 
  4. Run a mock access request - Pick one client and ask for their full data trail the way they could under GDPR. Time the answer. Repeat until it takes hours, not weeks.

How Cyera helps

Cyera handles the first three steps as it runs and gives you a starting point for the fourth.

Steps one and two. Cyera discovers and classifies sensitive data across cloud, SaaS, and on-prem stores, and flags excessive permissions to it. Agent Guardian extends that view to agents: which ones exist, what data and tools they can reach, and what actions they can take. 

Step three. Access Trail ties every access event, by a person or an AI agent, to the data’s sensitivity, ownership, and location. Data lineage shows what employees and agents did with files, including what was changed and where the data moved.

Step four. A data risk assessment shows where one line of business stands before a request arrives. If a clock is already running, Cyera’s Incident Materiality Assessment gives expert-led answers on what data was exposed and how sensitive it is. 

The next request will arrive the way the last one did, without warning and with a deadline. Have the answer ready before it does.

Find out how long your next request would take. Book a data risk assessment.

Already on the clock? Ask about an Incident Materiality Assessment.

Partager