Case Study — AI Employee
Deploying an AI Employee in an Insurance Brokerage
How we'd put Hermes Agent to work on the admin load an independent agency can't hire its way out of.
This is a worked example, not a delivered engagement — the vertical, workflows, architecture and guardrails are real and how we'd actually build it; the brokerage profile is illustrative.
The business
An independent commercial lines brokerage. A principal, two producers, three account managers, a small service team. Profitable, well-run, and completely bottlenecked on administration. None of the work below is skilled work — all of it was done by skilled people, because there was nobody else to do it, and the agency couldn't make the economics of another hire work at their premium volume.
Certificates of insurance
A contractor needs proof of coverage before they can start on site, today. Five to ten minutes each — find the policy, confirm the holder, generate the certificate, write the email. Several hundred a month. Zero revenue attached to any of them.
Renewal preparation
Every policy needs work starting ninety days out. Pull loss runs, summarise coverage, note exposure changes, assemble a proposal. Genuine expertise sits in the decisions, but eighty percent of the hours go into gathering and formatting.
Carrier submissions
The same client data re-keyed into four to six different carrier portals, each wanting it in a slightly different shape.
Why an agent, and why this one
We build a lot of conventional automation, so we ask this on every engagement before writing any code.
The disqualifying constraint: their core system has no usable API. The agency runs on an Agency Management System — the category that includes Applied Epic, AMS360 and HawkSoft. For a firm this size there is no practical integration path. No webhooks, no affordable API tier, no export that isn't manual.
That single fact rules out the entire no-code automation category. Zapier and n8n need an endpoint to talk to. There isn't one. What there is, is a piece of desktop software that a human operates by clicking — and Hermes Agent with computer-use access can operate it the same way: reading the screen, navigating the interface, pulling the policy record. Not elegant. It is the only thing that works.
Three other reasons Hermes specifically
Self-hosted and MIT-licensed
Client PII, financial detail and coverage information. Routing it through a third-party agent service was never going to pass the principal's own risk assessment, let alone their E&O carrier's. Hermes runs on the agency's hardware, data in a local store.
The principal is never at a desk
He's at client sites, in carrier meetings, driving between both. An agent reachable from Telegram, that can send him a draft to approve from a car park, is the difference between a tool that gets used and one that doesn't.
It accumulates skills
Every carrier portal has its own quirks. Every certificate holder has preferences. Hermes writes reusable procedures from successful runs, so the fifth submission to a given carrier goes better than the first.
Architecture
Requests come in over email and Telegram. Hermes reads the skill and fact store, operates the AMS by computer-use inside a defined boundary — and everything that governs it (spend caps, the kill switch, audit logging, the approval gate) sits deliberately outside that boundary, deterministic and non-negotiable. That boundary is the argument of the whole piece.
Where we drew the line first
In insurance the boundary between administrative work and licensed activity is bright, legally meaningful, and the first thing we established.
The agent never binds coverage. Never quotes. Never advises on coverage adequacy.
All three are licensed activities and all three stay with licensed humans, full stop. The agent gathers, formats, drafts and routes. A person decides. Once you accept that constraint, the design follows: everything the agent produces is a draft, every draft has a named human approver, and the audit trail records who approved what and when.
Least privilege, per tool
Read access to the inbox and draft-creation rights. No send rights. Read-only into the AMS for the first eight weeks; limited write later, for specific fields, on specific record types.
A human gate on anything that leaves the building
Certificates, emails, submissions. The approver taps approve on their phone. Seconds per item, and it removes the entire category of failure where something wrong reaches a client and nobody notices.
Spend caps and a kill switch
Hard daily and per-task ceilings on model spend, alerting well before the limit. One command stops everything. Both existed before the agent touched a live record.
Full audit logging
Every tool call, every action, every output, timestamped. The thing that turns "we have an AI doing our certificates" from a red flag into a defensible process.
Isolated execution
Container, not the principal's laptop.
Staged rollout
Eight weeks read-only, watching what it proposed without letting it act. Then drafts with approval. Then a narrow set of low-risk actions. Trust earned in stages.
What it does
Four workflows. Scope discipline is the difference between an agent that's still running in month six and one that got switched off in month two. The initial list had nine — two of the remaining five turned out not to be wanted once these four were running.
Certificate of insurance issuance
A request arrives by email. The agent parses it, pulls the policy record from the AMS, generates the certificate, drafts the reply. It lands in the account manager's queue with the source request attached. Where the request is ambiguous, the agent flags it rather than guessing — an agent that confidently produces a wrong certificate is worse than no agent.
Renewal preparation pack
Ninety days before expiry, the agent assembles the file: coverage summary, loss runs tabulated, exposure changes flagged against last year, a draft proposal in the agency's own format. The producer opens a prepared file instead of a blank one.
Carrier submission assembly
One intake, formatted into each carrier's required shape, learned as skills over the first few weeks. Submissions are prepared, never sent — a licensed person reviews and submits.
Morning brief
7:00am, in Telegram. Renewals hitting the ninety-day mark, certificate requests that arrived overnight, carrier responses outstanding past their expected date. The value isn't summarisation — it's that the principal stopped opening four systems to find out whether anything was on fire.
The self-improvement loop, described honestly
Hermes writes reusable skills from successful task runs and keeps persistent facts across sessions. It's real, and it's the main reason we chose Hermes over a stock framework. It's also routinely oversold, so here's what it actually looks like.
What genuinely improved
The agency's house style, and which holders want certificates in which format
Each carrier portal's particular sequence
The specific phrasings this principal uses and doesn't
What didn't
It did not become able to do things it couldn't do before
Skill accumulation is compounding personalisation, not capability growth
What needs managing
Skill libraries accumulate near-duplicates and stale procedures
Left alone, the agent can pick a superseded skill and produce last quarter's format
What was hard
Case studies that only contain wins are advertisements.
Computer-use against legacy desktop software is brittle
The AMS was never built to be driven by anything but a human. Modal dialogs appear unpredictably, screen state changes between versions. What made it workable: fixed screen resolution, generous verification between actions, and a hard rule that the agent confirms it's on the expected screen before acting. It still fails sometimes — loudly, handing off to a human, which is the only acceptable behaviour.
AMS vendor terms had to be checked first
Automating a desktop application you licence is not automatically permitted, and finding that out after deployment would have been an expensive mistake.
Document extraction is imperfect
Loss runs and ACORD forms arrive as scans of varying quality. Extracted figures are presented for confirmation rather than used silently — anywhere a number mattered, a human sees it before it's relied on.
Silent failure is worse than loud failure
An agent that quietly does nothing is dangerous, because the business assumes it worked. Every workflow reports completion — including failures — into a channel someone reads.
Teaching the team to manage it
Nobody had a mental model for a colleague that works at 3am, never asks a clarifying question unless told to, and is occasionally confidently wrong. We wrote a one-page operating guide: what to delegate, what to check, what to never assume was handled.
What we deliberately don't let it do
The most useful section for anyone considering this.
Nothing that requires a licence — no binding, no quoting, no coverage advice
No autonomous sending — every outbound item is a draft with a named approver
No financial transactions — no payment credentials, no premium handling, no trust account access
No client-facing chat — the agent talks to staff, staff talk to clients
No cold outreach without compliance sign-off
No new capability without a workflow that needs it — the default answer to a permission request is no
Commercial model
Setup covers discovery, deployment, hardening, workflow build and staff training. A monthly retainer covers model spend, monitoring, skill curation and workflow changes.
The retainer isn't padding. An unmaintained agent degrades — skills go stale, a carrier portal changes layout and a routine breaks, a model deprecates, the business changes what it needs. Maintenance is the product working.
Where it stands
4
Workflows live
Most firms being sold "AI employees" right now are being sold a demo. The demo always works. What happens six weeks later, when a portal changes its layout and nobody notices for four days, is the actual product. We deploy agents with permissions scoped, spend capped, approval gates on anything that leaves the building, an audit trail that survives a compliance review, and someone maintaining it afterwards. If your core system has no API and you've been told that means automation isn't possible — it usually is. Let's talk.
Ready to Get Started?
Book a free discovery call and let's explore how Dev can accelerate your business with AI-first solutions.
🍪 We use cookies
We use cookies to analyze site traffic and improve your experience. No personal data is sold.