What happens, in order
1. Scope one process with a known baseline. Renewals on one line of business is the usual starting point: high volume, explicit rules, and a current cost everybody already understands. We need a baseline to be measured against.
2. Configure the product. Your coverages, parameters, business rules, underwriting logic, pricing structure and compliance requirements go into Sherpa+ as configuration rather than code. This is the reference every agent will check against, so it comes first.
3. Benchmark against your own people. We run the agents across a slice of your historical cases and compare their recommendations with what your underwriters actually decided. The disagreements are the valuable output, and this is a far better test than a demo.
4. Set the human gates. Confidence thresholds, escalation routes and a named human owner for every agent, agreed in design rather than added in remediation.
5. Go live on one process, then extend. New business runs on the new model from the outset. An existing book migrates incrementally, highest-value lines first, with a documented manual fallback behind every automated step.
How fast is fast?
It depends on the shape of the work, and the honest answer has two numbers.
An embedded programme: about six weeks. On one live deployment, technical kickoff to go-live ran roughly six weeks, with functional design approved inside four weeks and user acceptance testing over five days. Commercial alignment, capacity and regulatory approvals sit outside that technical build and can take longer.
An enterprise back-office programme: phased, across quarters. A tier-one carrier's accident and health and motor operation covering ten operational modules and more than 200 active products is not a six-week job. That work runs in delivery phases with an agent workforce built out module by module.
What both share: nothing waits on a core replacement, and nothing waits on a capital approval cycle.
How do you connect to our systems?
Four ways, and we use whichever exists.
APIs, where your systems expose them.
Batch files, SFTP or bordereaux, where they do not.
Screen navigation, where the only route in is the user interface a person would use.
Robotic paths into legacy and mainframe environments, for the systems nobody wants to touch.
A meaningful share of insurance operations still runs behind portals and desktop applications with no usable API. Requiring clean APIs before starting would rule out most of the work worth doing.
The questions that come up first
Why do so many agentic AI projects get cancelled?
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls.¹
In insurance operations those three symptoms usually trace back to one cause. Most projects start from a written standard operating procedure, and that document becomes the agent's instructions. Day one is fine. Then a coverage limit moves, an exclusion is added, appetite tightens in one territory. The product changed and the document did not, so the agent carries on executing last quarter's product. The gap widens quietly and gets found by a customer or an auditor rather than by the operations team.
Cost then escalates, because the fix is continuous manual reconciliation. Value stays unclear, because the touchless rate plateaus as exceptions accumulate. Controls look inadequate, because the agent can prove it followed the written step and cannot prove it applied the current rule.
Our answer is architectural rather than procedural. Agents validate against your live product configuration, so a rule change on Tuesday changes Wednesday's decisions with nothing rewritten.
Does this take underwriting control away from our team?
No. Underwriting and claims decisions stay with your people.
Agents prepare, validate, check against your guidelines, route and execute. Anything outside authority escalates with the case assembled and the reasoning attached, so a human decides in one step rather than reconstructing the file. Confidence thresholds route borderline and sensitive actions to a person before they have any real-world effect.
What changes is the shape of the work. Your most experienced underwriters stop touching every routine renewal and start defining and supervising the rules, which puts their judgment across the whole book rather than the cases they had time to open.
Will it work with our legacy systems?
Yes, and legacy is the normal case rather than the exception.
A meaningful share of insurance operations still runs behind portals and desktop applications with no usable API. Requiring clean interfaces before starting would rule out most of the work worth doing, which is why we integrate four ways and use whichever exists. Nothing waits on a core replacement.
What will our regulator want to see?
Complete records rather than samples. Every agent action is logged with its inputs, the rule that triggered it, the decision taken, the confidence level and the outcome, traceable to a named human owner and queryable for any period.
Every regulated function is performed by a licensed party, and you choose which one. We are partner-agnostic, so if you already work with a third-party administrator or broker you trust, we operate alongside them. Where you need one in a given market, we bring options from our partner ecosystem. Piece 4 of this guide sets out the structures and how the model maps to emerging supervisory guidance on AI.
How much do we have to commit upfront?
One process, and no capital.
Deployment starts on a single line of business rather than the whole operation, and we structure engagements so the majority of the fee moves with volume rather than sitting in an upfront implementation charge. That means we carry risk if the work does not materialise, and it keeps the decision out of the same approval queue as a core replacement programme.
If the first process does not produce a result worth extending, that is useful information, and the commercial structure should not punish you for acting on it.
What changes for the people doing the work today?
Less of the work nobody wanted. The routine 80% of cases, the rekeying, the chasing of missing information and the reconciliation move to the agents. What stays with your team is rule design, supervision at portfolio level, and the genuinely difficult cases where judgment changes the economics of the book.
We recommend involving the operations lead and the underwriting lead in scoping rather than at go-live, because they are the people who will own the agents.
Frequently asked questions
How long before we see something working?
Weeks, not quarters. The benchmark against your own historical cases produces evidence before anything touches production.
Who does the implementation work?
Our own delivery team, working alongside yours, and in some markets alongside a licensed partner where regulated activity is in scope.
Which process should we start with?
Renewals on a single line, in most cases. High volume, explicit rules, and a current cost everybody already understands, which makes the result easy to read.
What does a deployment cost?
Pricing is set case by case, against the scope of work and what that work costs you today. We put a number to it during scoping rather than publishing a rate.
What do you need from us to start?
A process, a baseline, a slice of historical cases to benchmark against, and access to the systems that process touches. The first conversation is about the workflow rather than the architecture.
¹ Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", 25 June 2025.