Offering · Agent build
We build an agent for one workflow inside the systems you already run, with clear operating limits, traceable decisions and a structured handover. Your business and control owners set the limits and retain the decisions that carry risk.
What you see in production
An agent prepares, checks and drafts. When it acts for a person, it reads only what that person may see; when it runs in the background, it works under its own approved identity and scope. Either way it acts only within limits an owner set, and hands anything uncertain to a person with the reason stated.
The record behind each step shows the data it used, how current that data was, the version that ran and who approved the result. That is what your risk, finance and operations teams will ask for.
Follow one governed request
Select a step to see what the system checks and what evidence remains. The example is a claims-preparation request; the same pattern applies to other governed workflows.
Selected stage
Identify the person, role, channel, workflow purpose and case. If an agent is acting in the background, use its registered workload identity and approved scope.
Evidence retainedUser and agent identity, purpose, case ID and request time.
Insurance
Insurers hold the context an examiner or underwriter needs, but it sits in several systems and a procedure library. These agents assemble it and cite it.
Prepares assigned-case status, events, parties and the applicable procedure for examiner review. The worked example is synthetic and read-only.
Reads broker submissions, extracts the exposures and checks them against appetite before an underwriter opens the file.
Prepares routine changes such as address, payment or beneficiary updates, and checks the documents and authority behind each one.
Banking
Bank analysts spend much of a review gathering and reconciling records before any judgment starts. These agents do the gathering and leave the judgment where it belongs.
Assembles the records, screening results and changes since the last review into a draft file for the analyst.
Pulls financial statements, exposure and covenant history into the bank’s memo template, so credit officers review analysis instead of assembling it.
Legal and tax
For in-house legal and tax functions and for law firms. These agents work from your approved playbooks and research sources, keep each matter’s material within its permissions, and never give advice on their own.
Reads an incoming contract, compares each clause with your playbook and clause library, and prepares a marked-up review for the lawyer.
Gathers the facts, prior positions and relevant authorities from approved research sources, and drafts a research memo for the tax team.
Prepares a new matter for conflicts review by gathering the parties, related entities and prior matters, and checking them against the firm’s records.
Finance and operations
Any firm with time, billing or a close can use these, from professional services to portfolio companies.
Watches unbilled work against thresholds the business and finance agreed, and does the routine follow-up when one is crossed.
Sorts the queue so routine items are prepared and only true exceptions reach a controller, with the reason stated.
Growth and portfolio companies
Smaller firms need the same owner, limits and evidence, delivered in weeks and run by a small team.
Answers customer questions about where an order is and why it is held, so service agents stop searching three systems.
Prepares a one-page briefing before a customer call from the CRM, open tickets, orders and recent correspondence.
These are examples of the agents and patterns we build, not client case studies.
How an agent reaches production
The model is one component among the data it reads, the tools it can call, the identity it acts under and the people who rely on it. We approve that whole composition for a stated scope.
A change to any part, a new source, a new tool or a new model version, can change what was approved, so it goes back through the gate.
Where AI should not act
Two questions set the bound for each task: what an error would cost, and how much human judgment the task requires. Each quadrant carries a control, agreed with the business owner before release.
The quadrant follows the use, not the task type. A summary for an internal note and a summary placed in a medical, legal or customer file carry very different consequences, so classify the specific use before choosing the review.
Human judgment by design
Full autonomy is oversold. Human involvement is not a final approval click; it is designed in at three points and scaled to what an error would cost.
People set the task, the sources the agent may use and the actions it may never take.
Owners watch for drift: a changed source, a superseded document still being retrieved, a rising rate of overrides.
People step in where the cost of an error or the uncertainty of a decision calls for it.
Where the business is a human relationship, such as advice delivered by financial professionals, a fully autonomous agent is a change of business model, not a better workflow. We design those agents to strengthen the professional: preparing the conversation, servicing the account and following up on what was promised.
Agent authority
By default an agent gathers evidence and drafts. Next it proposes changes for an accountable owner to accept. It acts on its own only for a narrow action with independent authorization, tested preconditions and a way to recover. Some actions are never automatic.
How it fits together
The agent works between your systems and the people who own the process. Nothing is ripped out: it uses the systems of record and data platform you already run, under its own identity and permissions.
AI-ready data foundation AI security and governance evidence
People
Owners, reviewers and operators
Set limits, approve what carries risk, redirect the work.
Agents
Prepare, check, draft, alert
Each with its own identity, tested before every release.
Your systems
Systems of record and data
Claims, policy, core banking, ERP, CRM, documents and your data platform.
Included with every agent
One person answers for the result, its limits and when the agent may do more.
Answers come from approved, current sources, filtered by each user’s permissions before the model sees them.
Versioned test sets with pass thresholds, rerun whenever the model, data or procedure changes.
A named person who can stop a release, and a failing test that proves the gate works.
Who asked, under what authority, which sources and versions were used, and what came back.
Accuracy on your cases, how often people step in, escalations, drift and cost per completed task. Interventions should fall as trust is earned. A rising rate means the controls need attention; a rate near zero can mean nobody is really reviewing.
Read next: Worked guide: claims preparation Agent Onboarding Adoption by Design How we work
Next step
A free sixty-minute session with Richard to examine one workflow, what is blocking it and whether there is a useful next step.