AI for financial services

Ask what customers complained about last month, and AI for financial services answers from your complaints log, tagging each complaint to a root cause and any vulnerability indicator, and your complaints team checks a sample. MI commentary comes back drafted with every figure footnoted to its query, and an analyst gets a customer's alert history in one view.

Last reviewed ยท Virgo Analytics

One question, traced, example

Claims delay complaints rose most, 84 to 126, with the net rise in home insurance

Example data

You asked

Draft last month's complaints commentary for the MI pack.

Drafted October's complaints commentary6 steps

  1. Read the question0.2s

    You want the complaints paragraph for the October pack. Last month is 1 to 31 October 2026, compared with September as the pack does.

  2. Decide whether to query0.1s

    Every figure in a pack has to come from the tables at the pack's snapshot date. Nothing can come from memory.

  3. Pick the dataset0.3s

    Used complaints_tagged, one row per complaint, with its root cause tag read from the complaint text and checked on a sample, its product line and its snapshot date.

    complaints_taggedsnapshot 5 Nov 2026

  4. Write the queries0.6s

    Query 1 counts complaints by root cause for each month. Query 2 splits the root cause that rose most by product line.

    Query 1

    WITH tagged AS (
        SELECT
             c.complaint_id
            ,c.root_cause_tag
            ,DATE_TRUNC('month', c.received_date) AS month_received
        FROM mi.complaints_tagged c
        WHERE 1=1
            AND c.snapshot_date = DATE '2026-11-05'
            AND c.received_date >= DATE '2026-09-01'
            AND c.received_date < DATE '2026-11-01'
    )
    SELECT
         t.root_cause_tag
        ,COUNT(CASE WHEN t.month_received = DATE '2026-09-01' THEN 1 END) AS september
        ,COUNT(CASE WHEN t.month_received = DATE '2026-10-01' THEN 1 END) AS october
        ,COUNT(CASE WHEN t.month_received = DATE '2026-10-01' THEN 1 END)
            - COUNT(CASE WHEN t.month_received = DATE '2026-09-01' THEN 1 END) AS change
    FROM tagged t
    GROUP BY t.root_cause_tag
    ORDER BY change DESC, t.root_cause_tag

    Query 2

    WITH claims_delays AS (
        SELECT
             c.complaint_id
            ,c.product_line
            ,DATE_TRUNC('month', c.received_date) AS month_received
        FROM mi.complaints_tagged c
        WHERE 1=1
            AND c.snapshot_date = DATE '2026-11-05'
            AND c.root_cause_tag = 'Claims delays'
            AND c.received_date >= DATE '2026-09-01'
            AND c.received_date < DATE '2026-11-01'
    )
    SELECT
         d.product_line
        ,COUNT(CASE WHEN d.month_received = DATE '2026-09-01' THEN 1 END) AS september
        ,COUNT(CASE WHEN d.month_received = DATE '2026-10-01' THEN 1 END) AS october
    FROM claims_delays d
    GROUP BY d.product_line
    ORDER BY october DESC
  5. Run the queries1.0s

    Both finished against the read-only view. Nothing was written.

    5 rows3 rows

Write the answer0.4s

Each figure in the draft carries a footnote to its query. Why claims delays rose is not in these tables, so that sentence is left for the owner of the claims delays tag.

Complaints received rose from 300 in September to 340 in October.1 Claims delays rose most, from 84 to 126, and made up 37.1% of October's complaints.1 The net rise was in home insurance, where claims delay complaints went from 55 to 97. Pet rose by one and travel fell by one.2 Complaints about unclear policy documents rose from 52 to 61.1 The other three root causes fell.1 [Reason for the rise in home claims delays: to be written by the tag's owner.]

Footnote 1: 126 rows tagged Claims delays, received in OctoberLeft blank: the cause of a rise isn't in the tables
  1. Query 1, mi.complaints_tagged, snapshot 5 November 2026, complaints received 1 September to 31 October 2026, 5 rows.
  2. Query 2, same table and snapshot, root cause "Claims delays", by product line, 3 rows.

8 rows0.4 GB scanned2.6s total

Result 1
Root causeSeptemberOctoberChange
Claims delays84126+42
Policy documents unclear5261+9
Payment and direct debit errors6058โˆ’2
Price at renewal5654โˆ’2
Staff conduct4841โˆ’7
Total (sum of the five rows)300340+40
Result 2
Product lineSeptemberOctober
Home5597
Pet2021
Travel98
Total (sum of the three rows)84126
Example data: an invented insurer's complaints log for September and October 2026, read at the 5 November snapshot. Tags were set from the complaint text and checked on a sample before this query ran. To watch a trace like this one play out, with each step timed as above, open the AI-led analytics demo.

What works now in AI for financial services

Complaint text tagged to root cause and vulnerability indicatorsThe AI reads each complaint's free text and gives it a root cause from your own list, plus any vulnerability indicator the customer's words mention, such as a bereavement or a health condition. It stores each tag with the sentence it came from. Each month a complaints handler rereads a random batch, and any tag they overturn goes back to its owner with the reason. It needs the complaints log with the customer's own words, the root cause list, the product line and the outcome.
MI commentary footnoted to its queriesThe paragraph above is the form: footnote 1 names Query 1, mi.complaints_tagged, the 5 November snapshot and its 5 rows, and every other figure's footnote follows the same form. It needs MI tables with a snapshot date on every row, and the pack's timetable and definitions, set out on the management reporting page.
Alert history gathered for an analystWhen a transaction monitoring alert opens, the AI puts the customer's earlier alerts, how each was closed and the case notes into one view, each line cited to its alert or note ID. The analyst reads it and makes the decision. It needs the alert table with outcomes and the case notes, both keyed on customer ID.

What does not work yet

Deciding a complaint or its redressA tag says what a complaint is about. A complaint handler decides whether it is upheld and what the customer gets.
Vulnerability nobody wrote downThe tagger reads what the customer said, so a customer who never mentioned their circumstances gets no indicator, and the count of tagged complaints is a floor.

A drafted paragraph with no footnotes is one more thing to check by hand before sign-off.

What has to be true of your financial services data first

  • The complaints log keeps free text, root cause and outcome

    If the customer's words are summarised into a code at the point of entry, the tagger has nothing to read and nothing to check a tag against.

  • Every MI table carries a snapshot date

    A footnote reading "snapshot 5 November 2026" is only worth printing if someone can rerun its query a month later and get the same 126.

  • Each tag has a named owner

    Root cause wording changes over time, so one person approves each change and September's "Claims delays" means the same as October's.

The AI readiness assessment takes last month's pack, runs these three checks on its complaints log and MI tables, and then answers ten of your MI team's questions by hand.

Regulation and risk for AI in financial services

The sources we read on 20 November 2026, quoted or named.

The FCA's page "AI and the FCA: our approach"

The page says:

"We do not plan to introduce extra regulations for AI. Instead, we'll rely on existing frameworks, which mitigate many of the risks associated with AI."

It also says:

"Our rules emphasise accountability for senior managers and are relevant to the safe use of AI."

FCA, AI and the FCA: our approach. Checked 20 November 2026, when the page showed it was last updated on 18 November 2026.

The Consumer Duty

FCA, Consumer Duty information for firms.

Checked 20 November 2026.

The Money Laundering Regulations 2017

In the Money Laundering, Terrorist Financing and Transfer of Funds (Information on the Payer) Regulations 2017, regulation 28 covers customer due diligence measures and regulation 40 covers record-keeping.

Checked 20 November 2026.

Automated decisions are covered once, on the hub page under What every sector shares.

This is a list of the sources we read, with the date. It is not legal advice, and your own adviser decides what applies to you.

Every drafted figure names its query and the snapshot it read, so a reviewer can rerun any number before sign-off.

Where the method is explained

Questions people ask

Can AI tag our complaints to root causes?

Yes, from the complaint's free text, against the root cause list your complaints team already uses, and each tag is stored with the sentence it came from. A complaints handler rereads a monthly batch and logs the reason for any change.

Can AI write the commentary in our MI pack?

It can draft it. Each figure in the draft carries a footnote to the query and rows it came from, so the pack's owner can check any number before sign-off. The person who owns that part of the pack writes why it moved.

Does AI decide whether a transaction is suspicious?

No. The AI gathers a customer's alert history into one view, and an analyst decides. Whether to make any report, and what it says, stays with your own people.

What does the FCA say about rules for AI?

The FCA's page "AI and the FCA: our approach" says: "We do not plan to introduce extra regulations for AI. Instead, we'll rely on existing frameworks, which mitigate many of the risks associated with AI." Nothing here is legal advice.

Why does every figure in the draft read the 5 November snapshot?

Because complaint records keep changing after month end, when a tag is corrected or an October complaint is logged in November. Reading the 5 November snapshot means the 126 in the draft and the 126 a reviewer reruns next month are the same count, and each footnote names that snapshot.

Does every figure in your complaints MI trace to its log?

With the complaints paragraph from your last pack and the log behind it, the first call traces each figure to its rows and lists any that won't trace.

Book a call

Free 30 min call