Home/Agentic AI

Agents that do the work. Not demos that describe it.

Three classes of agent we build and operate, wired into the systems you already run — helpdesk, CRM, warehouse, data warehouse — and judged on resolution rate and cost per ticket rather than on how the demo felt.

RAGTool useEvalsGuardrailsHuman in the loopCost tuning
01 · The roster

Three agents,
live in production.

Pick one to watch it work. The console is a sketch of the real thing — the metrics underneath are the ones we actually report on.

Agent roster3 classes live
live · session 4f2a-9d31agents/knowledge
02 · In detail

What each one actually does.

Same engineering discipline as any other product we ship: real data, real permissions, real failure modes.

/01

Knowledge agent

Answers from your own material, with citations

Policies, manuals, contracts, catalogues and tickets become a retrieval layer an agent can answer from — with the source shown next to every claim, so staff can check it and customers can trust it.

  • Retrieval over private documents and databases
  • Inline citations and confidence scores
  • Re-indexing on a watch folder or webhook
  • Access control inherited from your systems
Plugs into
SharePoint · Drive · S3Confluence · NotionContract repositoriesProduct catalogues
/02

Service agent

Tier-one support that closes tickets, not deflects them

Persona-driven support across web, app and messaging, with triage, tool access and a clean human hand-over that carries the full transcript instead of starting the conversation again.

  • Triage and routing across channels
  • Real tool access — orders, accounts, bookings
  • Escalation with full context attached
  • Judged on resolution rate and cost per ticket
Plugs into
Zendesk · FreshdeskWhatsApp · web chatCRM and order systemsTelephony hand-off
/03

Insight agent

Ask the warehouse a question in plain English

Natural language to SQL over your warehouse, with the query shown, the chart rendered and a written brief in the inbox each morning — so the analyst is freed for the questions that actually need one.

  • Text-to-SQL with the query surfaced
  • Anomaly flags on the metrics you name
  • Scheduled written briefs
  • Dashboard-aware, joins inferred
Plugs into
Postgres · BigQuerySnowflake · Redshiftdbt modelsMetabase · Looker
03 · Where it pays

Automation by function.

The work we are asked for most often, in the order it usually pays back.

Support

Tier-one resolution, order and account lookups, refunds within policy, and escalation that arrives with the transcript already summarised.

Sales

Inbound qualification, proposal drafting from your own past work, CRM hygiene and follow-up sequences that reference the real conversation.

Finance & back office

Invoice and document extraction, three-way matching, reconciliation exceptions surfaced for a human instead of a monthly hunt.

Operations

Dispatch suggestions, exception handling, SLA watchers and the small decisions that currently wait for someone to open a spreadsheet.

People & internal

Policy and benefits answers, onboarding walkthroughs, and an internal helpdesk that cites the handbook rather than paraphrasing it.

Engineering

Codebase and runbook retrieval, incident summaries, release notes and the triage pass before a human picks up the pager.

04 · How we ship one

Four beats, six to ten weeks.

From a named workflow to an agent in one channel with a score you can point at.

01

Pick a workflow with a number attached

We start where the cost is measurable — tickets per week, hours per close, days per enrollment. If nobody can name the number, the agent has nothing to be judged against.

02

Build the retrieval and tool layer first

Most agent failures are data failures. We get the documents, permissions and tool calls right before any prompt engineering, because that is what makes answers correct rather than plausible.

03

Evals before rollout, not after

A test set from your real traffic, scored on accuracy, refusal and escalation behaviour. You see the score before the agent sees a customer.

04

Ship narrow, widen on evidence

One channel, one queue, a human in the loop. Scope widens when the numbers justify it — and every widening is a decision you make, not a default.

05 · Guardrails

The boring part that
keeps it in production.

Every agent we run ships with all six. None of them are optional extras.

01

Human in the loop where it matters

Refunds, medical, legal and anything irreversible route to a person by design. The agent prepares the decision; it does not make it.

02

Grounded or silent

Answers come from retrieved sources with citations. When confidence is low the agent says so and hands over, rather than inventing a plausible sentence.

03

Evals as a standing test suite

The scored test set runs on every prompt, model or data change — so a model upgrade is a measurable event rather than a surprise on Monday.

04

Cost and latency budgets

Per-conversation cost and response time are tracked like any other SLO, with model routing and caching tuned against them.

05

Your data, your tenancy

Models run against your infrastructure and accounts. No training on your data, retention set by you, audit log on every tool call.

06

A kill switch that works

Every agent can be paused per channel or per queue in one action, falling back to the human path without a deploy.

06 · Built with

The AI stack in production.

Model-agnostic by design — routing is a config decision, not a rewrite.

Anthropic ClaudeOpenAILangChainLlamaIndexQdrantpgvectorPythonTypeScriptPostgresRedisAWS BedrockKubernetesLangSmithOpenTelemetry

Name the workflow.
We’ll tell you if an agent should touch it.

Thirty minutes with the engineers who run these in production — including the honest answer when automation is the wrong tool.

Book a 30-min callGet a project quote