Set up the Investigator and its policy

Upload your policy, turn your SOP into Investigator rules, find the gaps, and test every change on your history, a golden set and custom checks first.

#Upload your policy

In the Your policy step of the setup, press Upload your policy and add a PDF or Word document of up to 20 MB, or a link to one. Didit reads it once, in the background, and drafts an answer to each setup question, quoting the section it came from. Confirm or reject each drafted answer. Questions the document doesn't cover are marked Not in your policy and keep the answer you give them.

The file is kept in private storage that only your organization can open. Deleting the document deletes its file and every drafted answer nobody confirmed yet. An answer you confirmed keeps naming the document and section it came from.

#Your SOP

The SOP tab shows your policy as a readable standard operating procedure, one rule per kind of case. It has eight views:

The last three are explained further down this page.

There are three ways to start a draft:

• Ask Didit lists the kinds of case your team decided most in recent weeks, shows what your team usually chose, and asks what should happen with each. Each answer you change becomes a rule in the draft.
• Paste your SOP reads the procedure your team already follows and drafts a rule for each situation it states, noting the section and sentence each rule rests on. Rules are drafted only where your text gives an answer.
• Adding a gap or changing a message also starts a draft.

When the draft is ready, press Publish. Publishing makes it the SOP the Investigator runs, and the waiting reports are checked against it. If someone published another version while you were editing, discard your draft and start again from the new one.

#Check for gaps

Gaps lists the situations your team reviewed that no rule covers, most reviewed first. Press Check now to check the SOP against the last 90 days of reviews. For each gap you can add its suggested rule to the draft, or dismiss it, in which case that situation keeps going to a person.

#Messages to people

For a rule that asks for more information, Messages to people sets what the person is asked and why:

Write the reason the person reads, in up to 500 characters. Only verification and business verification rules can message the person, and the AML queue never asks the person for anything, so nothing can tip them off. When information requests are on, the message also opens a conversation with the person: see The Investigator, the people you verify and their data.

#Test a change before it goes live

There are two kinds of backtest:

• The Quick check replays your policy on reports the Investigator already prepared over the last 30 or 90 days, without seeing what your team decided, and compares the two. It's part of the setup and of every change of level.
• The AI backtest has the agent re-read past sessions under your draft, so it also tests what the policy can't express as rules. Before it starts it shows how many sessions it will re-read, how long it takes, the price per session and the maximum cost, and you confirm it. If your credit doesn't cover it, you can run a smaller sample. Each session it actually re-reads is charged once: its price is held on your wallet while it's re-read and charged when its report is in. Sessions that fail, are skipped or never start are free, and Sandbox applications cost nothing. You can stop it at any time and only pay for the sessions it already re-read. Its results show the agreement with your team and each overturn, the sessions it would have decided otherwise, with both decisions side by side.

To start an AI backtest, press New AI backtest in SOP → Backtests and set its scope:

The estimate then shows the sessions to re-read, about how long it takes, the price per session re-read and the most it can cost. Sessions sent back for resubmission are left out, because their live session is no longer the attempt your team decided. It's paid from your plan credit, then your balance, or billed under your contract.

Both backtests warn you when the sample may not represent your traffic: fewer than 50 decided items, 80% or more of them from one country or one workflow, fewer than 3 disagreements to learn from, or less than 14 days of decisions. Read the agreement with care when you see a warning, or replay a longer period.

When a backtest finds a disagreement, Ask my team sends it to your team as a question. In the setup's Check it on your history step, open a disagreement to see your team's decision and the Investigator's side by side, with the rule it followed and what it could not see. Then either Change the answer it used, or press Ask my team: the question gets a number, and it waits in Questions and in your Inbox until someone answers it.

#Questions from the Investigator

When an item doesn't fit your policy, or your team keeps overriding the same kind of recommendation, the Investigator asks one question with the item attached. Questions appear in the Questions tab and in your Inbox, and Open in Inbox takes you to them. Preview what an answer would change before you save it. Your answer becomes a new policy version, or you can record that nothing should change, or dismiss the question to close it without an answer.

#Golden set

The golden set, in SOP → Golden set, is a list of reference items with the outcome you expect for each: approve, decline, request information, or always a person. Pin up to 100 items from past reports. Whenever someone publishes a policy or SOP change, or restores an old version, the golden set is checked first:

• If an item's outcome would change and no longer match what you expect, publishing stops and shows you the items that changed.
• You can accept a change on purpose. The item's expected outcome then becomes the new one.
• Each published version also suggests new reference items for the rules it added, which you can accept or dismiss.
• When Didit releases a new version of the Investigator engine, the golden set is checked again automatically.

You can also run the set on demand, and have the agent re-read each pinned item blind to compare its answer. Golden set runs are never billed.

#Custom checks

Custom checks, in SOP → Custom checks, let you add a check in your own words, for example "if the business is a currency exchange, check that it holds a licence in the national register". The builder:

1. asks you clarifying questions until the check is unambiguous;
2. writes the check, reviews it for problems and shows you what it will do;
3. runs it on recent past items it applies to, in shadow and without charge, so you can see which items it would flag;
4. starts running it on matching items once you activate it. You can pause it at any time.

A custom check adds evidence to the report. It never decides an item on its own.

#Workflow recommendations

In SOP → Recommendations, from the last 90 days of decisions, the Investigator suggests changes to your verification workflows where your team keeps deciding the same way, for example raising the face match threshold, adding a proof of address step for one country, or relaxing a warning your team always clears. Each recommendation shows how many items it would have affected and how often your team agreed. Accepting one records that you intend to make the change and takes you to the workflow editor: it never changes a workflow on its own.

#What it costs

The Investigator is available from the Growth plan. Setup, policy questions, shadow mode and the Quick check are always free. Shadow mode covers the whole application: moving any queue to Auto or Full auto takes the Investigator out of shadow mode, and from then on every finished review is billed at your plan's or contract's price. Set a Monthly spending limit in its settings to cap that. See how pricing and credits work.

A review is finished when its item is decided with the report: a person accepts it, as it is or edited, or the Investigator decides it on its own. It's billed once per item and decision, however many times the Investigator reads it. These are never billed: setup, policy questions, shadow mode, the Quick check, overrides, items decided outside the Investigator, and reviews that fail.

The AI backtest is paid: each past session the agent actually re-reads is charged once, at the $0.10 list price less your plan's discount, or at your contract's price when a contract pays, and billed under ID verification. Sessions that fail, are skipped or never start are free, and Sandbox applications cost nothing. Consistency checks, QA of your reviewers and the golden set run in shadow and are never billed.

#Monthly spending limit

Monthly spending limit, in Settings → Cost, caps what the Investigator's reviews cost in a calendar month (UTC), in your billing currency: 100, 250, 500, 1,000, 2,500 or 5,000 a month. It's off by default. The price of every open report outside shadow mode is held back against the limit, because accepting it will charge it. At the limit, the Investigator pauses: no new review starts, and items go to your reviewers without a report until the next month or a higher limit. Activity records Paused at the spending limit, and the investigator.automation.paused webhook is sent. The limit never stops shadow mode.

#If your plan changes to Free

On Free, your team still reviews every item by hand in Verifications, Transactions and Cases. If your plan changes to Free, your policy, questions and settings are kept, and the Investigator is paused until you are back on Growth.