The broker's official client record may sit in the BMS, while the live working file is scattered across Outlook, Teams, SharePoint, OneDrive, insurer portals, PPL records, PDFs, spreadsheets and personal folders.
The stated impact is search time, duplicate work, missed instructions, slower file review and E&O or audit risk. For a 50-person commercial broking team, this is a strong candidate for AI-assisted automation because the cost is largely fixed while the benefit scales across every account handler, executive, claims handler, renewals colleague and compliance reviewer affected by fragmented files.
Problem: Split Client Files
The operational problem is not that the broker lacks a system of record. The problem is that the BMS holds the official policy, client and accounting record while the evidence needed for daily work is distributed across the tools staff already use.
Proposed Solution: AI-Assisted Client File Control
Create a single client file control layer that automatically gathers, labels and reconciles key documents and messages from the places staff already use.
The goal is not to replace the BMS. The BMS remains the official policy, client and accounting record. The AI-supported workflow sits around it and makes sure that important working documents and communications are visible, correctly filed and traceable.
This is a fixed-scope project model, not a large recurring integration programme.
What the Solution Would Do
The workflow controls the file by finding relevant items, classifying them, matching them to the likely client or policy, and routing exceptions where human confirmation is needed.
Recommended Operating Model
Use a controlled automation workflow as the core, with narrow AI support for interpretation.
A fully autonomous AI agent is not the right primary design for this problem because the broker needs predictable control, auditability and low operational risk. The process involves regulated records, client instructions, policy evidence and potential E&O exposure. A controlled workflow is better because each step is measurable, repeatable and auditable.
AI support is still useful, but only for judgement-heavy tasks such as reading documents, recognising client instructions, spotting conflicting versions and summarising file history. The final record movement, alerts and approvals should follow a fixed workflow.
Best-fit design: 85% structured workflow + 15% AI interpretation.
Recommended operating model
Target Workflow
The target process starts whenever a new email, document, Teams file, SharePoint upload or portal output appears.
New emails, Teams files, SharePoint uploads, OneDrive documents and portal outputs enter the workflow.
Reads documents and messages to identify client, policy, insurer, document type, transaction and importance.
Matches the item to the likely client and policy file, then recommends filing, review or alert handling.
Medium-confidence, high-risk and uncertain items go to handlers or reviewers for confirmation.
Daily exceptions, unresolved evidence gaps and team statistics are surfaced to managers.
Trigger and AI Review
Examples include client acceptance, underwriter clarification, quote revisions, updated property schedules, fleet lists, claims histories, statements of fact, local OneDrive documents that should be in the working file, portal outputs and BMS reference data.
The AI reads the item and identifies the client name, policy reference, insurer, document type, transaction type and importance of the record.
| Data point | Example |
|---|---|
| Client name | ABC Manufacturing Ltd |
| Policy reference | COM/123456 |
| Insurer | Aviva, AXA, Zurich, etc. |
| Document type | Quote, schedule, certificate, instruction, endorsement |
| Transaction type | Renewal, MTA, new business, claim, complaint |
| Importance | Client instruction, underwriter condition, acceptance, declinature, material change |
Filing Recommendation and Confidence Check
The item is matched to the likely client and policy file. The system should not initially make unrestricted changes to the official BMS record. Instead, it should recommend filing for standard items, auto-route important items to review, flag missing evidence and create file note suggestions where a summary would help the handler or reviewer.
This keeps the process controlled and avoids incorrect filing. Account handlers and team leads receive only exceptions, not another general inbox.
Implementation Scope for a 50-Person Team
The first version should stay commercially credible and fixed in scope. It should prove value in renewals, MTAs and client instructions before expanding into higher-risk or broader integration work.
| Workstream | Included |
|---|---|
| Workflow design | Define document categories, file-matching rules and exception logic. |
| Source setup | Outlook, SharePoint, Teams and a client/policy reference list from the BMS. |
| AI document reading | Identify client name, policy reference, insurer, document type and instruction type. |
| Filing recommendation | Suggest the right client or policy location. |
| Exception queue | Send uncertain or high-risk items to handlers. |
| Missing-evidence alerts | Flag client acceptance, underwriter subjectivity, changed terms and important unfiled evidence. |
| Basic reporting | Weekly count of filed items, exceptions and unresolved evidence gaps. |
| Staff training | Practical training for handlers, executives and team leads. |
| Testing and tuning | Accuracy testing on real broker documents. |
Excluded from the First Version
| Excluded item | Reason |
|---|---|
| Full BMS write-back | Higher cost and higher audit risk. |
| Insurer portal integrations | Adds complexity; not needed to prove value. |
| Historic file clean-up | High volume, weak immediate ROI. |
| Complex management dashboards | Start with simple exception reporting. |
| Fully autonomous document movement | Too risky before accuracy is proven. |
| All process areas at once | Start with renewals, MTAs and client instructions. |
Expected Efficiency Impact
The base case assumes 50 users, 220 working days per year and a target saving of 20 minutes per person per day at a fully loaded staff cost of £45 per hour.
| Metric | Assumption |
|---|---|
| Users | 50 |
| Average working days per year | 220 |
| Current time lost searching or reconstructing files | 20-40 minutes per person per day |
| Target time saved after automation | 20 minutes per person per day |
| Fully loaded staff cost | £45/hour |
| One-time implementation cost | £35,000 |
| Monthly maintenance | £1,200 |
Annual time recovered is 50 users x 220 days x 20 minutes, or 220,000 minutes saved per year. That equals 3,667 hours saved per year. At £45 per hour, the annual labour value is £165,000.
The base-case calculation is £115,600 net benefit divided by £49,400 year-one cost, producing 234% ROI. Payback is £49,400 divided by £165,000, multiplied by 12 months: 3.6 months.
Scenario View
| Scenario | Saving assumption | Year-one net benefit | ROI | Payback |
|---|---|---|---|---|
| Conservative case | 15 minutes saved/user/day | £74,350 | 151% | 4.8 months |
| Base case | 20 minutes saved/user/day | £115,600 | 234% | 3.6 months |
| Strong case | 25 minutes saved/user/day | £156,850 | 318% | 2.9 months |
| Year two | 20 minutes saved/user/day | £150,600 | 1,046% on maintenance | Ongoing |
The project only needs to save about 6 minutes per person per working day to cover its year-one cost. That is a credible threshold for a split-file problem where staff currently search across BMS, email, Teams, SharePoint, portals and local folders.
Additional Rework and Risk Reduction
If the 50-person team avoids only 500 hours per year of rework from duplicate document creation, missing evidence and file review failures, that adds £22,500 additional annual value. Search-time saving plus rework saving produces £187,500 total annual benefit and a 3.2-month payback.
Preventing two material file-quality escalations per year, each requiring 25 internal hours at an average £70 per hour, adds a further £3,500. That is not the main ROI driver, but it strengthens the operational risk case.
Why This Should Not Be a Fully Autonomous AI Agent
A fully autonomous AI agent could search, decide, move files, update records and create summaries with limited human intervention. That is attractive in theory, but not ideal here.
For broker file control, the risk is not just operational efficiency. The system is handling evidence around advice, client instructions, subjectivities, acceptances, changed terms and policy documentation. A wrong decision could affect file quality, compliance review or E&O exposure.
| Option | Fit | Reason |
|---|---|---|
| Fully autonomous AI agent | Low-medium | Too much judgement risk for client instructions, policy evidence and audit trails. |
| Simple automation only | Medium | Reliable for filing rules, but weak at reading messy emails, PDFs and version conflicts. |
| Controlled workflow + narrow AI support | High | Combines predictable process control with AI document understanding. |
The best solution is therefore a controlled file-control workflow where AI assists with reading, matching, classifying and flagging, but staff retain approval over uncertain or high-risk items.
Recommended AI-Agent Position
- Controlled workflow first. The broker gets predictable process control and auditability.
- Narrow AI interpretation. AI reads messy emails, PDFs and version conflicts where deterministic rules are weak.
- BMS remains authoritative. The workflow improves file control without replacing the policy, client and accounting record.
- Exception-only reporting. Account handlers and team leads receive unresolved issues rather than another general inbox.
- Fixed commercial scope. The first version proves value before full write-back, portals or historic clean-up are attempted.
| AI-supported task | Control mechanism |
|---|---|
| Read email/document content | Confidence score and review queue. |
| Identify client, policy and document type | Match against client/policy reference list. |
| Flag client instruction or underwriter condition | Handler confirms importance. |
| Detect possible missing evidence | Exception report routes to owner. |
| Summarise file timeline | Staff can verify source documents. |
| Identify conflicting document versions | Workflow asks for confirmation. |
Success Metrics
The deployment should be measured with clear before-and-after operational metrics rather than vague AI adoption language.
| Measure | Current baseline | Target after 6 months |
|---|---|---|
| Average time to find key renewal evidence | 10-20 minutes | Under 3 minutes |
| Documents correctly classified | Not consistent | 70-80%+ |
| Correct client/policy matching | Manual / inconsistent | 75-85%+ |
| High-risk unfiled items detected | Not consistently measured | 90%+ detection queue coverage |
| File review failures due to missing evidence | Baseline from QA sample | 30-50% reduction |
| Duplicate document creation | Baseline from staff sampling | 25-40% reduction |
| Handler time spent reconstructing files | 20-40 mins/day | 10-15 mins/day |
| Break-even saving required | N/A | 6 mins/user/day |
| Target saving | N/A | 20 mins/user/day |
Recommended Rollout
Implement the workflow across one 50-person commercial team or one 50-person group split across commercial handling, account executives, renewals and compliance review.
- Phase 1Controlled deployment
8-10 weeks across one 50-person team, focused on renewals, MTAs, client instructions, subjectivities, quote and schedule version control.
- Phase 2Tune and stabilise
Review false matches, add document wording patterns, refine exception categories, track time saved and compare file review outcomes.
- Phase 3Optional expansion
Add another team, claims documents, BMS write-back, portal-specific workflows or historic clean-up only after value is proven.
Decision Rule for Rollout Success
Proceed to wider rollout only if the 50-person deployment achieves the following minimum targets.
| Criterion | Minimum target |
|---|---|
| Time saved per user per day | 15 minutes |
| Correct document classification | 70%+ |
| Correct client/policy matching | 75%+ |
| Reduction in where-is-the-evidence queries | 25%+ |
| Reduction in missing-evidence file review issues | 30%+ |
| Payback period | Under 6 months |
Recommendation
Implement a fixed-price AI-assisted document control workflow for a 50-person team.
This is a strong candidate for AI-assisted automation because the problem is repetitive, document-heavy, high-volume and costly, but still needs controlled human approval where client instructions or regulated records are involved.
The best design is not a fully autonomous AI agent. The right approach is a controlled automation workflow with narrow AI support, because it preserves auditability, keeps implementation cost under control and focuses the AI on the parts where it adds most value: reading documents, identifying key evidence, matching files and flagging exceptions.