30-Day AI Design Pilot Plan for a Regulated Reporting Team

Learn how to scope a 30-day AI design pilot for a regulated reporting team around a committee decision, with named documents, review roles, baselines and exit criteria fixed before day one.

Published Nov 17, 2026AI in Reporting

30-Day AI Design Pilot Plan for a Regulated Reporting Team

Summary

A pilot scoped as "try the AI" gives a risk committee opinions and no evidence. A 30-day plan names the documents, team and period, measures time to a reviewable draft, correction rounds and errors caught in review, and states before day one what would stop it. AI-assisted templates follow normal change control. For CTOs, heads of reporting and change managers at banks, insurers and wealth managers.


Consider a head of reporting at a mid-sized insurer who agrees to "try the AI" for a month. Four designers get access and use it as they see fit. One builds a new notice template with it, one uses it to restyle existing reports, one tries it twice and goes back to the editor, and one does not open it. At the end of the month the risk committee asks three questions. Was it faster? Did it produce errors? Was any customer data sent anywhere? The answers are that the designers liked it, that it felt faster, and that nobody is sure about the rest. No one measured the time before the pilot, so "faster" has nothing to be compared with. Errors were fixed as they were found, so nobody counted them. The committee asks for another quarter.

Nothing went wrong in that pilot, and nothing was learned from it either. It was scoped around a tool and not around a decision. This post sets out a way to scope a 30-day pilot so that, at the end, the committee can say continue, extend or stop, and show why.


What a Pilot Has to Answer

A pilot in a regulated team has an audience beyond the people running it: a risk committee, a procurement panel or an internal audit team who were not in the room. What they need is evidence that was defined before the work began. Three questions cover much of what they will ask.

  1. Is it faster to a draft someone will review? The measure is time to a reviewable first draft, compared with a baseline.
  2. Does it stay inside our controls? The measure is whether each AI-assisted change passed through the same review and approval as a manual one, and whether data handling matched what was agreed.
  3. Do the errors it makes get caught? The measure is the number and type of errors found in review, and whether any reached release.

A pilot built around those three questions needs three things fixed in advance: a named set of documents, a named team and a fixed period. The first stops the pilot drifting to whichever template is convenient. The second makes the review roles clear. The third forces a decision at the end and not a slow fade into permanent trial.


Choosing the Pilot Documents

Pick three templates that differ in difficulty: one simple, one moderate and one awkward. A pilot on easy documents flatters the tool, and a pilot on a single hard one is hard to interpret.

Tier Example What it tests
Simple A one-page confirmation letter with a few fields Whether the basic loop of describe, draft and review works at all
Moderate A monthly statement with a transaction table and a conditional notice Data binding, repeating rows and conditional content
Awkward A multi-page schedule with variants by product, long records and shared sections Edge cases, layout under stress and the cost of correction rounds

For each document, assemble a small set of sample records in advance: a typical record, a record with no activity, a record that runs over several pages, one with a long name and one with an optional field missing. The records should have the shape of real data. During design sessions, use invented or anonymised values. Real values belong in the later reconciliation step, in an environment with the permissions that apply to them. This keeps the question "was customer data sent anywhere?" easy to answer, because the plan says it was not and the log can show it.

Before day one, record a baseline for each document. If the team has change tickets or timesheets for similar templates, use those. If not, ask a designer to build one comparable template by hand in the week before the pilot and time it. A number from a single template is a weak baseline, but it is better than the impression that the last one felt slow.

For a migration pilot, where the documents come from a legacy reporting tool, the same pattern of a few representative reports applies; see AI-assisted migration from Crystal Reports.


Success and Exit Criteria, Written Before Day One

One page carries the plan: the page that says what will count as success and what will end the pilot. Write it before the first session, have the committee read it, and do not change it during the month without recording why.

Measure How to define it Baseline Continue if (example) Stop or pause if (example)
Time to reviewable first draft Minutes from the first instruction to a draft a reviewer agrees is worth reviewing From week 0 Median at most half the baseline on simple and moderate documents No improvement on any tier
Correction rounds Review cycles before sign-off From past tickets Median of two or fewer Rounds rise above the baseline
Errors found in review Count by type: layout, data binding, wording, missing required content Not applicable Logged and fixed before sign-off Not applicable
Errors reaching release Errors found after sign-off, by any route Not applicable None Any, since it triggers a pause
Control exceptions Changes that skipped the normal change process Not applicable None Any
Data handling Customer data seen by the assistant or sent with an attachment Not applicable None, as planned Any

The thresholds are examples. Each team should set its own, and should set them with the people who will have to defend them.

Two points about this table. First, errors found in review are not a failure of the pilot. They are the review doing its job, and a pilot with no logged errors is more likely to have a weak review than a flawless tool. The signal that matters is the errors that escape review. Second, the stop conditions matter a great deal to a risk committee, and teams are the least likely to write them down. Decide in advance that a change released outside change control, or customer data in a conversation that was meant to be synthetic, ends the pilot or pauses it until the cause is understood.

A 30-day plan can be laid out as five stages.

Stage Days Work
Setup Before day 1 Approve scope and criteria, set up the environment and access, record baselines, assemble sample records
Simple document 1 to 7 Build and review; practise the logging routine
Moderate document 8 to 14 Build and review; test binding checks against source data
Awkward document 15 to 21 Build and review; record where the tool struggles
Reconcile and decide 22 to 30 Reconcile outputs to source, complete the log, write the summary, hold the decision meeting

Who Reviews What

AI-assisted templates should pass through the same change control as manual ones. The change ticket does not need to describe how the template was made, but it helps the pilot to add one field, "AI-assisted: yes or no", so the two groups can be compared afterwards.

A pilot needs five roles, and one person can hold more than one in a small team, with one exception below.

  • Designer. Uses the assistant and produces the draft.
  • Template owner. Reviews structure: right sections, right order, right layout on the awkward records.
  • Data reviewer. Checks the data bindings against source values. This is a separate review from the visual one, because a table that looks right can be bound to the wrong field.
  • Compliance reviewer. Owns the list of content that must appear and confirms it does.
  • Pilot lead. Keeps the log and runs the decision meeting.

The exception is that the designer should not approve their own template. That rule applies without the pilot, and it matters more when drafts arrive faster.


What to Record

A pilot is useful to the committee when it leaves a record they can read without the team in the room. Record, as the work happens and not afterwards:

  • The scope, criteria and thresholds as approved before day one.
  • For each document: the baseline, the time to reviewable draft, the correction rounds and the reviewers.
  • An error log with the type, how it was found, by whom and whether it reached release.
  • What the assistant was shown: which records, which attachments and which data sources.
  • The configuration: which provider and model were used, who had access and what instructions the assistant was given.
  • The decision and the reasons for it, including what remains unanswered.

A negative or mixed result is a good outcome if the record is clean. A committee that sees an honest log is more willing to approve a second, wider phase than one that sees a summary with nothing behind it.


How This Maps to CxReports

Here is how the plan maps to CxReports, using what the documentation states.

A separate pilot workspace. A workspace is an independent environment, and workspaces are not interconnected. A root user creates one and adds its users. The AI Assistant needs an AI Integration, which a workspace administrator configures with an engine (OpenAI, Anthropic, Google or DeepSeek), a model name and the organisation's own API key. Data Export and Import do not carry AI Integrations, database connections, users or roles, so each is set up by hand. That helps in a pilot, because the pilot workspace can hold no production database connection. The assistant can browse external databases and inspect runtime data, so a workspace without a connection to production has nothing from production to read.

Who gets the assistant. Access needs the AI permission on the user's role, at two levels: Read for read-only chat and Full Access for full use. The documentation describes the levels in those few words, so test what each can do in the setup stage before assigning roles. The plan's roles map to workspace roles, with the assistant restricted to the designers.

Instructions the pilot team shares. AI Skills are plain-text instructions given to the assistant automatically. A global skill goes into all conversations in the workspace, while a report skill applies to conversations opened from that report. For a pilot, a skill that lists the fixed content and house conventions for one document, scoped to that report, keeps the instructions visible and avoids leaking one document's rules into another. Skills are plain text read by the model, so they are not the place for credentials or customer data. The skill text belongs in the evidence pack, since it is part of what the assistant was given.

Sample records as data sources. A JSON or CSV file can be attached to a message, up to 256 KB per file and three files per message. When the assistant creates a data source from an attached file, the data source holds the file's exact content, so the awkward-record set can be attached once and each variant is tested against it. Images and PDFs can also be attached, for example a PDF of the current document as a reference, and a PDF is uploaded to the AI provider, which keeps it for a period that depends on the provider. Attach documents that the plan allows to leave the building.

Reviewing what the assistant did. Each message gives the assistant an overview of the report, and its actions appear in the chat as expandable rows, so a reviewer can see what was added or changed. The documentation does not state how long conversations are kept or what assistant use is logged. The pilot lead should therefore keep the pilot's own record of what was asked and what changed, and not rely on the product to hold it.

Reconciling output. A report can be previewed against the sample records, and the API export call accepts an inline data payload together with params, so the same awkward records can be rendered to PDF repeatedly as the template changes. The data reviewer compares figures against the source system. This tests the template whoever drafted it.

A retrievable copy and promotion. Data Export saves selected reports, templates, themes, dictionaries and other elements as a file, so exporting before and after a change gives the change record a retrievable copy. If the pilot succeeds, Data Import loads the file into another workspace, with a choice to stop, keep the existing version or replace it when an element already exists. Elements that reference an external database fail to import unless the database exists in the target. Importing into production would still go through the bank's normal approval.

The criteria, the thresholds, the review roles, the change control and the decision remain with the organisation. CxReports does not set a pass mark. The provider contract also remains with the organisation, because the organisation supplies the provider key. For the design assistant's scope and limits, see AI-assisted report design.


Getting Started

Pilot need CxReports mechanism What stays with the organisation
Contain what the assistant can read Separate workspace with no production database connection The decision on what data the pilot may use
Limit who can use it AI permission on roles; AI Integration per workspace Which roles get access, and the provider contract
Shared instructions Report-scoped AI Skill The content rules in the skill
Repeatable test records Attached JSON or CSV turned into a data source The sample set and its realism
See what changed Action rows in the chat The pilot log and its retention
Reconcile output Preview and API export with inline data The comparison to the source system
Retrievable copies Data Export before and after each change The change record and approvals

Documentation:

To set up a pilot workspace with your own sample documents, request a demo.


This article covers technology evaluation practices. The thresholds are examples and not recommendations. The supervisory expectations that apply to your institution depend on its charter, size and jurisdiction, and are determined by your compliance and risk functions. Nothing here constitutes legal or regulatory advice.

Modal Fallback