Industries/ Finance/ Policy and procedure search
Answers from your own policies, with the paragraph cited
Staff ask in plain English and get a short answer drawn from your compliance manual, lending criteria, product terms and committee minutes. Every answer names the document, the version and the section, so the person asking and the person reviewing can check it in one click.

- Used by
- Underwriters, advisers, operations staff and compliance
- Takes the place of
- Asking the one colleague who knows, and guessing
- Connects to
- Your document library, intranet and staff sign-in
- Built with
- Our knowledge search service
Where it hurts today
A normal week, before it exists
Every regulated firm has a compliance manual, a set of procedures, product terms in several versions and years of committee minutes. They live on a shared drive or an intranet, named in ways that made sense to whoever saved them. When an underwriter needs to know whether a contractor on a day rate counts as employed, finding the right file takes longer than asking a colleague, so they ask.
That colleague is usually the same experienced person, interrupted a dozen times a day. Their answers are good, but they come from memory, and memory does not always notice that the policy changed last quarter. Newer staff learn not to ask too often and make a reasonable guess instead. Most guesses are right. The ones that are not tend to surface in a file review months later.
Then a reviewer asks a harder question: what did the procedure say in March, when this case was handled? Answering means opening several versions of the same document, comparing them by hand and hoping the right one was saved. It takes a senior person a day, and the result still rests partly on their memory of when the change went live.
What it does
Policy and procedure search, feature by feature
Built for staff in a regulated firm, where the source and the version matter as much as the answer.
- 01
Answers with the source attached
Each answer is a short paragraph followed by the document, its version and the section it came from, with a link that opens the passage highlighted. If the indexed material does not cover the question, the assistant says so plainly rather than producing something that merely sounds plausible.
- 02
Versions and effective dates
Policies and product terms are indexed with the dates they applied between. By default staff get the current version. A reviewer can ask what applied on a past date, and the answer cites the version that was live then, with a note if it has since changed.
- 03
Plain English and exact references
Questions are matched on meaning, so staff can ask in their own words. Product codes, clause numbers and defined terms from your product literature are also matched exactly, so a question that mixes both finds the right paragraph rather than one that only sounds similar.
- 04
It respects who can see what
Credit committee minutes, HR policies and draft procedures stay visible only to the people allowed to open them today. Permissions are checked against your staff sign-in each time a question is asked, so removing someone's access takes effect straight away.
- 05
Contradictions sent to an owner
When two current documents disagree, the assistant shows both and says they conflict instead of choosing one. The conflict is logged and sent to the document owner, so the firm fixes the policy rather than leaving staff to decide which version to follow.
- 06
Questions nobody can answer
Questions that return nothing useful are collected and ranked by how often they come up. Your compliance lead might see that staff keep asking about joint accounts after a death, and that no procedure covers it. That list often becomes the next update to the manual.
A day with it
How the day runs instead
A normal day in a lender's operations and compliance teams once the assistant is live.
- 8.45amUnderwriter
An underwriter has a case with a contractor paid a day rate. She asks whether the income can be annualised. The answer quotes section 4.3 of the lending criteria, version 12, with a link to the paragraph, and she notes the reference on the case.
- 10.20amContact centre agent
A customer on the phone asks about early repayment charges on an older product. The agent asks the assistant, which cites the product terms that applied when the customer took out the loan, not the terms on sale today.
- 1.00pmNew starter
A new operations starter asks how to handle a customer who mentions a serious illness. The answer summarises the vulnerable customer procedure, cites it, and names the team to escalate to, taken from the procedure itself.
- 3.30pmCompliance analyst
A file review needs the complaints procedure as it stood in March. The analyst asks for that date, gets the version that applied and the date it was replaced, and attaches the citation to the review notes.
- Month endCompliance lead
The compliance lead reviews the unanswered questions and the conflicts log. Two procedures disagree on a retention period, and a cluster of questions about joint accounts went unanswered. Both go on the agenda for the next policy committee.
What it connects to
Works with what you already run
The assistant indexes your documents where they already live. Nothing has to be moved first.
Document library or intranet
Policies, procedures and minutes are indexed in place, for example from SharePoint or Google Drive where your firm keeps them, and kept current as documents change, with owners and effective dates recorded.
Staff sign-in
Staff use their existing Microsoft or Google work accounts. The same groups that control who opens a document control what the assistant will quote, so there is no second set of permissions to maintain.
Policy register
Where you keep a policy register with owners, review dates and approval records, the assistant reads it, so each answer can show who owns the document and when it was last approved.
Case management or CRM
Staff can ask from inside the case screen and save the cited answer to the case notes, so the reasoning behind a decision is recorded where reviewers will look for it.
The first version
Start small, then grow it
A first version should be useful on day one and small enough to get right. This is where we would usually draw the line, and what tends to follow once people are using it.
In the first version
- One defined set of documents, such as the compliance manual and product terms
- Answers that cite document, version and section, with a one-click link
- Permissions checked against your existing staff sign-in for every question
- A plain statement when the documents do not answer the question
- A pilot with a small team using the questions they really get asked
Once it is in use
- Point-in-time questions across older versions and effective dates
- Committee minutes, credit policy and regulatory correspondence added to the index
- Asking from inside case and CRM screens, with citations saved to the file
- A monthly report of unanswered questions and conflicting documents
Buyer's guide
Before you commission one
Four decisions worth making before you commission a policy assistant.
- 01
Built-in AI search, or a custom assistant
If your policies live in one place, the AI assistant that comes with your office software may answer well enough, and it already inherits your permissions. A custom assistant earns its place when you need point-in-time answers across versions, citations to the paragraph, logs your compliance lead can review, or documents spread across several systems.
- 02
What to gather first
A week of real questions from the people who get interrupted most, the document set you would start with, and a list of who owns each document. Note where versions live and how a superseded policy is marked today. If nobody can say which version is current, that is the first thing to fix.
- 03
Who should own it
Compliance should own what the assistant may answer and how its answers are reviewed, usually through the policy committee. Each document needs a named owner who resolves the conflicts the assistant reports. Operations or IT can run it day to day, but without compliance ownership it soon becomes another search box nobody quite trusts.
- 04
How you will know it is working
Watch how often staff ask, how often they open the cited paragraph, which answers they reject and how many questions return nothing. Ask your experienced colleagues whether they are interrupted less. The unanswered list matters most: when it shrinks because missing procedures were written, the assistant is doing its real job.
Terms worth knowing
The words you will meet
- Retrieval-augmented generation
- The method behind most document assistants. The system first finds the passages that match the question, then asks a language model to answer only from them. Done well, every answer traces to its source. Done badly, the model fills gaps from general knowledge, which is exactly what a regulated firm cannot have.
- Hybrid search
- Combining meaning-based search, which finds the policy when someone uses their own words, with exact-match search, which finds clause 14.3 or a product code. Each alone fails in ways staff notice quickly. Run together, they find the right paragraph whether the question is plain English or a reference number.
- Point-in-time answers
- Answers drawn from the version of a document that applied on a given date, not just today's. A file review or a complaint often asks what the procedure said when the case was handled, so the index has to know when each version took effect and when it was replaced.
Handled with care
The details that matter here
An assistant that answers policy questions in a regulated firm has to be checkable. These are the details we agree with your compliance lead.
Staff-facing, not customer advice
The assistant answers your staff from your own documents. It does not advise customers or make credit decisions. We agree with your compliance lead where its answers may be used, and anything that would reach a customer goes through the checks you already apply.
Every question and answer logged
Each question, the passages retrieved and the answer given are logged with who asked and when. We agree retention with your compliance lead, so a reviewer can see exactly what someone was told before they acted, and which version of the policy it came from.
Tested against your real questions
Before launch we collect real questions from your staff and score whether the right passage comes back. We share the results with your compliance lead, including the questions it got wrong, and document how the assistant behaves so it can be assessed against your rules.
Owners and review dates
Each indexed document carries its owner and last review date. We agree with your compliance lead what happens when a document passes its review date: the assistant can still cite it, but says it is overdue for review, so staff know to check before relying on it.
Guides
Worth reading first
Questions
What buyers ask first
Will staff trust an AI policy assistant?
They should trust it only as far as they can check it, which is why every answer carries its source. People verify the answers that matter, and one click to the paragraph makes that easy. In a pilot, the answers people reject are as useful as the ones they accept, and we review both with your compliance lead before the assistant goes to more teams.
What happens when the assistant gets an answer wrong?
It will be wrong occasionally, and the most common cause is two of your own documents disagreeing or a policy that was never written down. Because each answer cites its source, a wrong answer can be traced to the passage behind it. We log the rejections, fix what is ours to fix in the search, and pass documentation gaps to the document owner.
Can customers or brokers use the policy assistant?
Not in the first version. A staff tool relies on trained people reading the source. Anything answering customers or brokers needs approved wording for sensitive topics, a clear route to a person and closer review. In the EU, the AI Act has required since 2 August 2026 that AI systems talking directly to people tell them so, unless it is obvious. A broker version for lending criteria is possible later, within limits your compliance lead sets.
Our policies are scattered and out of date. Is it worth starting?
Usually yes, with one honest caveat. Search copes well with untidy filing. What it cannot do is decide which of two current documents is right, or answer from a policy nobody wrote. The first indexing pass produces a list of duplicates, contradictions and stale documents, and clearing the worst of them is often a few days of the document owners' time.
Is it safe to send our policies to an AI model?
It depends on the service and its terms, which is why we agree them before indexing anything. Options range from commercial AI services under contractual terms on training and retention, through processing pinned to a particular region, to models run inside your own cloud account. You get a written data flow that your risk team or a client's due diligence reviewer can examine.
Tell us the questions your experts answer daily
Send us the five policy questions your most experienced people get asked over and over. We reply within 48 hours and will tell you whether search, better documentation or both would help.
More for finance
What else we build for finance and banking
Client portals
A secure login where clients see valuations, statements, documents and application status, upload what you have asked for, and sign or approve without a printer. Permissions are checked on the server for every request, and every view is logged.
Guided onboarding journeys
Application and onboarding flows designed around how customers think: one question at a time, errors explained where they happen, progress saved between visits, and identity and suitability checks handed to your chosen providers at the right step.
Automated management reporting
Ledger, CRM and operational data joined on agreed definitions, so the board pack and investor updates build themselves from the same numbers every month, with unusual movements flagged for a person to explain before anything is sent.
Document reading and triage
Bank statements, payslips, claim forms, valuations and signed agreements read on arrival, key details extracted and checked against the client record, and anything ambiguous or above a set threshold routed to a person rather than processed automatically.
Secure cloud foundations
Hosting arranged so customer and account data are separated, encrypted and reachable only by the roles that need them, with access logged to storage nobody can quietly edit, and card details kept out of your own systems wherever possible.
Everything we build for finance
The same idea, in other sectors
