AI agents that work on your real data.
Most agent pilots look good in a demo and stall on real data: stale sources, missing fields, no way to tell whether an answer is right. We start where they stall. We get the data right and write the evals first, then build the agent in your stack, and stay to run it if you want us to.
When we're a fit, and when we're not.
A good fit if…
- You have a repetitive workflow that reads documents, websites or internal systems and ends in a decision or a draft.
- A pilot works in a demo but not on your real data.
- The agent has to show where every fact came from.
- You want it in your own repos and cloud, not on a black-box platform.
Don't build an agent if…
- A rule or a spreadsheet formula would do. Agents are for the steps that need judgement.
- You want a general chatbot on your website. An off-the-shelf tool is cheaper.
- Nobody on your side can own the workflow and review its output in the first weeks.
- The process changes every week and has never been written down. Fix that first.
What you get.
Deliverables you can check, not hours you have to trust.
- 1.
Workflow map
Every step, decision and hand-off written down with the person who owns the work, including the cases that go wrong today.
- 2.
Data layer
The sources the agent reads, cleaned and kept fresh, with a source and timestamp on every fact. Often the biggest win on its own.
- 3.
Eval suite
Real cases with the expected outcome for each. It runs on every change and blocks a release that gets worse.
- 4.
The agent and its tools
Narrow, typed tools (MCP or API) over your systems, and an agent that uses them. Model choice per task, versions pinned.
- 5.
Guardrails and review
Human approval where the stakes need it, limits on what the agent can touch, and clear failure paths instead of guesses.
- 6.
Monitoring and runbook
Logs, cost and quality tracking, alerts, and a runbook your team can follow. Or we operate it for you.
An eval report you can actually read.
Every release is measured against real cases agreed in discovery. A failing case blocks the release until it is fixed or the case is changed on purpose.
| Case | Expected | Agent result | Check |
|---|---|---|---|
| Supplier quote, VAT line missing | Flag for review | Flagged: VAT missing | Pass |
| Duplicate invoice | Reject, cite the original | Rejected, cites INV-2291 | Pass |
| Price above contract | Escalate to the buyer | Approved | Fail |
| New supplier, no record | Request a registry check | Check requested | Pass |
// eval run → release 0.14
{
"suite": "supplier-quotes-v3",
"cases": 120,
"passed": 117,
"failed": ["Q-0412", "Q-0587", "Q-0601"],
"blocking": true,
"median_latency_s": 6.2,
"model_versions": "pinned",
"run_at": "2026-11-10T09:14:02Z"
}Illustrative example. Your cases and thresholds are agreed with you in discovery.
What teams use it for.
Document triage
Quotes, invoices, applications and tickets read, checked against your records and routed, with a reason for each decision.
Research agents
Market, competitor and supplier research on fresh, sourced data instead of whatever a browser session finds.
Grounded copilots
Sales and support assistants that answer from your data and current market facts, and cite them.
Data-entry workflows
CRM and catalogue updates proposed by an agent, with the source attached and a person approving.
Discovery first. Then a fixed-price build.
The order matters. Without real cases and clean data, nobody can say whether an agent works.
- 1.
Discovery
One week to map the workflow, audit the data, collect real cases for the eval set and write a fixed-scope proposal.
- 2.
Build sprint
The data layer, tools and agent built in your stack, measured against the evals on every change.
- 3.
Launch with review
A supervised start: people approve outputs while the agent proves itself on live work.
- 4.
Run or hand over
We operate it under an SLA, or hand over code, evals and a runbook to your team.
Questions we get
Something else? Ask us directly.
Which models and frameworks do you use?
Whatever fits the task, from several vendors. We pin versions, re-run the evals before any upgrade, and keep the agent portable so you are not locked into one provider or into us.
Who owns the code, prompts and evals?
You do. Everything is built in your repositories and cloud account. If you stop working with us, you keep all of it, with documentation.
How long does a build take?
One workflow usually takes a few weeks after discovery. The proposal at the end of discovery gives the scope, the timeline and a fixed price.
How do you handle our data?
It stays in your systems wherever possible. We work with the access you grant, use least-privilege credentials, do not train models on your data and remove access at the end.
Do you have client case studies?
Not yet, and we will not invent them. Our proof today is our own agent-maintained dataset, Nomad Watch, which runs on the same gate. Our founding customers get a discount on their first build in return for a written case study.
Tell us about the workflow.
Describe the work you want an agent to do and the data behind it. Within 2 business days you get a written take: whether an agent fits, what it would take, and how we'd scope discovery. Or an honest note that you don't need us.
- A written take in 2 business days
- A fixed-scope proposal before any build
- No commitment, no sales sequence
Prefer email? [email protected]