AI & Agents
We build agents for specific jobs inside your organization, like triaging requests, assisting caseworkers, and reviewing documents against rules. Then we prove they work on your real cases and put the controls in place to run them: who owns each agent, what it can touch, and what needs a person's approval.
Led by an Army veteran. Built by engineers who have shipped production systems for NASA and Cigna, and at Amazon Web Services (AWS), Bloomberg, and Capital One.
The agent lifecycle
Most agents stall between a good demo and a system people rely on. We cover the whole lifecycle: build it for one job, prove it on your real work, and govern it once it's live.
01BUILD
We start from one job your staff does every day and build an agent that does the repetitive part, inside the systems they already use. Three examples of the kind of work that fits:
Reads incoming requests, pulls out the details, sends each one to the right team, and attaches a draft response for review.
Answers staff questions from policy and case history with citations, and prepares the next document in the case.
Checks documents or regulations against requirements and flags conflicts, gaps, and outdated language for a person to decide.
You decide what runs on its own, what waits for a reviewer, and when it stops. Here is how a typical intake agent handles each case. Anything incomplete goes to a person instead of being guessed.
Model providers' own SDKs and managed retrieval where it fits. No heavy frameworks a security reviewer has to untangle.
Each step is a defined state that can pause for approval, resume, and write to the audit log.
An agent gets only the systems and actions its job needs, and nothing it can't explain.
Approved commercial APIs, AWS GovCloud, or models hosted entirely on your own hardware.
02EVALUATE
We agree on what "working" means before we build, then test against it. You get a report you can take to your leadership and your security reviewers, including the results that aren't good yet.
Accuracy, review effort, handling time, and cost per item, with targets set together.
Real examples, including the hard ones and the edge cases your staff already know about.
Prompt injection, attempts to reach data the agent shouldn't, and requests outside its job.
Approvals show sources and reasoning, not just a confident draft. Fluent answers can persuade people even when they're wrong.
No update ships if it makes a passing case fail. Corrections from live use become new test cases.
| Criterion | Target | Result | |
|---|---|---|---|
| Routed to the right team | 95% or more | 96.8% | Pass |
| Required fields extracted | 98% or more | 98.9% | Pass |
| Reviewer edits per item | 1.0 or fewer | 1.4 | Fix |
| Cost per item | 5 cents or less | 3 cents | Pass |
| Prompt-injection attempts | 0 succeed | 0 of 40 | Pass |
03GOVERN
Once agents are live, the questions change: who owns this one, what can it reach, when did we last test it, and how do we stop it. Governance answers those questions on any given day, not just at launch.
| Agent | Owner | Can access | Needs approval for | Last evaluated | Status |
|---|---|---|---|---|---|
| Request triage | Operations lead | Request inbox, routing table | Any reply sent outside the office | Nov 14 | Live |
| Policy assistant | Benefits manager | Policy library, read-only | Nothing; it answers staff questions only | Nov 10 | Live |
| Records review | Records officer | Case files, read-only | Any change to a record | Oct 30 | Paused |
Every agent with its owner, purpose, access, and evaluation history. It also feeds the AI use-case reporting agencies already do.
Least-privilege access per tool and per system, reviewed whenever the agent's job changes.
Set per action, not per agent: what runs on its own, what needs a reviewer, and who that reviewer is.
Every input, decision, approval, and change, kept in your environment and ready for review.
Quality, cost, and error rates tracked against the launch criteria. The owner is alerted when they drift.
Pause any agent at once, see what it already did, and restore the last approved version.
Aligned to the four functions of the NIST AI Risk Management Framework: Govern, Map, Measure, and Manage. Security and evaluation
Handover
If the builder is the only one who can fix an agent, you don't really own it. Handover is planned from the start, and it isn't finished until your team has proven they can handle a failure on their own.
We run it alongside your team and keep an issues log. Everything we learn goes into the handover package.
For the people who run the infrastructure: architecture, runbook, deployment, and the evaluation harness.
For the staff who use it every day. They try to break it while we watch, and we fix what they find.
Your team handles a failed run with us unavailable: notice it, pause new actions, check what already happened, and restore the last good version with their own access.
YOU KEEP
Prefer that we keep running it? Ongoing support is available, with the same documentation either way.
Starting point
The Production Readiness Review covers controls, evaluation, running cost, and adoption for one workflow, and ends with a go, fix, or stop recommendation. It's fixed price, and the findings are yours whether or not you hire us to build.
Step 1: two weeks
Step 2: you have a pilot
Step 2: starting fresh
FAQ
Often you shouldn't, and we'll tell you. Custom work makes sense when a workflow crosses systems your tools don't connect, needs approval rules or audit records they don't support, or has to run somewhere they can't, such as GovCloud or a network with no internet.
A map of the workflow, a test set built from your real cases with measured results, an estimate of running cost, and a written go, fix, or stop recommendation with a fixed-price plan for the next step.
One workflow, a sample of real documents or requests, and a few hours a week with the people who do the work. System access only as far as the review requires.
Senior engineers who have built production systems for NASA and Cigna, and at Amazon Web Services (AWS), Bloomberg, and Capital One. The people who scope your project are the people who build it.
Sparingly. We call model providers directly, use managed retrieval where it fits, and write the orchestration ourselves so every step is visible and easy to review. Fewer dependencies also means less for your security team to approve and maintain.
No. With private models, your data never leaves your environment. With commercial APIs, we use enterprise terms that exclude training on your data, and only where your policies allow.
Yes. Open-weight models can run entirely on your own hardware, including on networks with no internet access.
Contact
Send a short note: what the system does today, what isn't working, and any deadlines. An engineer will reply within two business days.
contact@autonar.ai