A four-person employee benefits consultancy in Surrey was running out of road. Forty-five SME employer clients, each expecting a polished annual scheme review, personalised benefit benchmarking, and employee communication materials — and a team of four to deliver it all. Every quarter, they deferred new business enquiries they could not take on. By September 2026, 29 hours of every working week were disappearing into admin that followed a clear, repeatable pattern. Three AI agents changed the maths.
The Bottleneck: Where the Hours Were Going
Employee benefits consultancies run on a repeating delivery cycle. Every employer client gets an annual scheme review — a structured meeting backed by a written report comparing their current benefits package against FCA suitability standards, employee uptake data, and market benchmarks. Between reviews, the firm handles renewals, mid-year queries, and employee communication requests when employers change or add benefits.
For a four-person firm serving 45 clients, this is a full-time operation. The pattern was consistent: each annual scheme review required four to six hours of preparation — pulling the previous year's data, researching current market rates for group life, income protection, private medical, and employee assistance programmes, drafting the report itself, and preparing the employee communication materials that came out of it. Spread across 45 clients and a calendar year, that is one full-time role dedicated to scheme review prep alone.
The three specific tasks consuming the most time were:
- Benefit benchmarking research — comparing clients' current benefit levels against market norms for their sector and headcount. On average 90 minutes per client per year, plus ad hoc research when a client asked whether their private medical excess or income protection deferral period was competitive
- Scheme review report drafting — pulling together data, structuring the analysis, and writing the report section covering FCA Consumer Duty suitability, total remuneration context, and recommendations. Two to three hours per review
- Employee communication materials — letters, intranet copy, and benefit statements explaining changes after a renewal or scheme redesign. One to two hours per batch, four to six times per year per client
Nineteen of their 45 clients were due an annual scheme review in a single quarter. The team completed eleven. The other eight were delayed — not because of a lack of information or expertise, but because there were not enough hours in the week to do the preparation work.
We knew exactly how to do every piece of the work. We just had too much of it. Every review we delayed was a client relationship risk — and a growth ceiling we could not push through.
Three Agents, One AI Operating System
The design principle was simple: identify the three repeating tasks consuming the most prep time, and build an agent for each one. Not a replacement for the consultant's judgement — a preparation layer that handles the information gathering and first-draft production so the consultant can focus on the analysis, the client relationship, and the recommendations that only a human should make.
All three agents were built on n8n, drawing on a shared knowledge base containing the firm's past scheme reviews, provider market data, and communication templates. The governance layer — following the three-layer framework for UK service businesses — was built in from the start, with clear data classification rules for what client data could flow into which part of the system.
Agent One: The Scheme Review Agent
The Scheme Review Agent runs in advance of every annual review meeting. When the consultant triggers it — typically two weeks before the client meeting — the agent pulls the client's current scheme data from the firm's CRM (provider, benefit levels, employee headcount, uptake rates, premium costs), retrieves the previous year's scheme review report from the knowledge base, runs a structured analysis against the firm's FCA Consumer Duty suitability checklist, and drafts the written scheme review report covering the executive summary, year-on-year comparison, suitability assessment, and a draft recommendations section.
The output lands in the consultant's Slack channel as a formatted draft, flagged clearly as AI-prepared. The consultant reads it, edits the recommendations to reflect their professional judgement, and adds the qualitative observations they have built up through the client relationship. The average review and edit time: 45 minutes. Pre-agent, the average preparation time was three hours.
Agent Two: The Benefit Benchmarking Agent
The Benefit Benchmarking Agent handles the market research that previously dominated scheme review preparation. When triggered with a client's sector, headcount band, and current benefit structure, it queries the firm's regularly updated provider data files to identify current market norms by sector and company size, produces a structured benchmarking table comparing the client's benefits against market median and upper quartile, flags any areas where the client is materially below market, and generates a natural-language commentary on the benchmarking findings for inclusion in the scheme review.
This agent connects directly to the delegation matrix logic the firm used to design the system: benchmarking research has clear inputs, a well-defined output format, and zero ambiguity about what good looks like. It was the highest-volume, clearest-spec task on the list — exactly where AI delivers most reliably.
Agent Three: The Employee Communication Agent
The Employee Communication Agent drafts the materials employers need when benefits change — the letters, intranet copy, and benefit statements that explain renewals, new schemes, or changes to coverage. When the consultant submits a brief (employer name, change type, tone, any specific points to emphasise), the agent retrieves the employer's communication tone preferences and approved terminology from the knowledge base, drafts the full communication pack covering the letter, benefit statement summary, and FAQ section for common employee questions, and formats the output for immediate use in the employer's preferred channel.
All three agents were built with a hard human-in-the-loop checkpoint before any output reached a client. No AI-generated scheme review, benchmarking analysis, or communication ever left the firm without a consultant's review and sign-off. That boundary — clearly defined from day one — was the reason the firm could use AI in an FCA-regulated context without governance risk.
The Numbers After Six Months
Six months after deploying all three agents, the firm ran a structured review of their output numbers against the same period the previous year.
- Scheme reviews completed: 59% increase in annual scheme reviews completed on time within each quarter
- Time recovered: 24 hours per week across the team — equivalent to just over three working days
- On-time delivery rate: Up from 65% to 96% — the first time the firm had hit near-full on-time delivery
- New clients onboarded: Seven new employer clients in six months — the first net growth in two years
- Running cost: £90 per month in LLM API costs and n8n hosting
The seven new clients were the outcome the firm had not fully modelled when they started the build. They had framed the problem as clearing the backlog. They discovered that clearing the backlog freed enough consultant time to take sales calls, follow up on warm leads, and actually onboard new business. Capacity, not demand, had been the growth constraint — and they had not fully realised it until the AI system removed it.
We thought we were solving a delivery problem. We were actually solving a growth problem. The agents did not just recover hours — they changed what those hours could be used for.
What Made It Work
Three decisions made the difference between a system that worked and a pilot that stalled.
Knowledge base quality first. Before writing a single line of automation, the firm spent two weeks curating their knowledge base — past scheme reviews, provider market data sheets, their FCA Consumer Duty suitability templates, and their communication style guide. The agents are only as good as what they can draw on. The benchmarking agent depends entirely on current, structured provider data. That curation work is ongoing — the knowledge base is updated monthly — and it is the reason the system keeps improving rather than degrading over time.
Clear task selection. The three tasks chosen for automation had one thing in common: well-defined inputs, a clear output format, and established quality standards the consultant could verify in under an hour. Tasks requiring contextual judgement — the actual recommendations in a scheme review, the strategic conversation with an employer about their benefit objectives — stayed human. The ROI framework the firm used to evaluate each agent before deployment confirmed this: every agent deployed produced a measurable return within the first month of use.
Compliance built in, not bolted on. Operating in an FCA-regulated environment, the firm designed the governance layer before the agents. Client data was classified as red-tier from day one — handled only through API deployments with zero data retention and explicit DPAs with every tool in the chain. The human-in-the-loop checkpoint before any client-facing output was non-negotiable. This was not a constraint that limited what the system could do — it was the design decision that made the system safe to deploy at all.
Employee benefits is a sector built on trust and regulatory credibility. Any AI deployment that compromised either would cost more than the hours it saved. The firm's approach — automate the preparation, protect the judgement, govern the data — is the model any FCA-regulated advisory business should follow when building AI into their operating system. The same principles apply whether you are running an IFA practice, a commercial insurance brokerage, or a pensions advisory firm.
If you run an employee benefits consultancy or any professional advisory firm where AI could clear the admin backlog without compromising the advice — get in touch. We design AI operating systems for UK service businesses that are built to perform in regulated environments from the start.