Skip to content
Service · AI Agents

— 02 / always on AI Agents
agentic workflows that do the work, not the talking.

Agentic workflows that do the manual work your team does today. Built on your data, wired into your systems, in production with a full audit trail. Not chatbots. Not pilots. Our fastest shipped in 9 days — and still runs.

Format
Build sprint
Live in
weeks not quarters
Fastest
9 days
Runs
24/7 monitored
— What we build

One agent does a task. A workflow runs the process. A workforce runs the desk.

  1. 01

    Agent

    One job, end to end. Reads the document, extracts the fields, checks the rule, routes the exception, logs every step. Fastest to production: 9 days.

  2. 02

    Agentic workflow

    Agents chained across a whole process — the NAV close, the reconciliation break queue, the KYC screening review. The machine carries the clean flow; your team takes what it escalates. Where the compounding starts.

  3. 03

    Digital workforce

    Workflows composed across the operation. Orchestrated and monitored as one layer, governed to a standard your auditor signs. This is what Backbone runs on.

44% of finance teams already run agentic AI in 2026 — six times more than a year ago (Wolters Kluwer). The question stopped being whether. It's which process first.

— Why us

we ship workflows that run, not demos that impress.

Most "AI" in asset management is a chatbot bolted onto a slide. Ours are agentic workflows in production at regulated managers in Chile and Peru — triaging regulatory documents, reconciling data, catching exceptions before a human looks.

A workflow earns its keep only if it closes the loop: pull the document, extract the fields, check the rule, route the exception, log every step. We build that whole loop — model, tools, guardrails — and wire it into your PMS, your CRM, your document store.

Every workflow ships with audit trails, source citations and redaction by default. When your auditor asks what the agent decided and why, the answer is one query away. Claude, GPT, open-source LLMs, computer vision, NLP — the model sized to the task.

— In the catalogue

Six workflows. Pick the process, not the promise.

WorkflowWhat it takes overStatus
Regulatory document triage Classification and extraction of ~1,200 inbound docs/month In production · Chile
Reconciliation break clearing Routine multi-custodian reconciliation exceptions In production
KYC screening review Hit clustering; obvious false positives cleared, matches escalated In production
NAV close assistant Pre-close validations and audit evidence assembly In production · Backbone
Compliance filing assembly Regulatory packs assembled from the captured record In production · AML/KYC
Corporate actions processing Capture, validation and application with human review Roadmap

Your process isn't here? That's what the discovery call is for — we'll tell you in 30 minutes whether a workflow can run it.

— How we build

Automation that survives production — five rules we won't bend.

  1. — 01

    Production, not pilots

    No POCs, no demos that die in the boardroom. Built to run from day one.

  2. — 02

    It does the whole job

    A workflow that needs a human at every step isn't automation. People handle the exceptions, not the routine.

  3. — 03

    Measured before it ships

    Every workflow comes with an eval suite proving it works on your data. Vibes don't pass procurement.

  4. — 04

    Wired into your stack

    It calls your PMS, CRM and custodians. Automation only counts if it reaches the systems you actually run.

  5. — 05

    Auditable and cost-anchored

    Audit trails and provenance by default. The bill scales with usage, not headcount.

— Process

Four phases — weeks to a working workflow.

Building an agent is a fixed-scope sprint. We agree what "working" means up-front, build against it, and deploy when the evals say it's ready — not when the calendar says so.

  1. WEEK 01
    — scope

    Brief & Success Criteria

    Which task the agent takes over, and how we'll know it works. We assemble an eval set from your real cases and set the pass thresholds before writing a line of agent code.
    Outputs
    • Automation brief
    • Eval set (≥ 200 cases)
    • Success thresholds
    • Cost ceiling
  2. WEEK 02–09
    — build

    Build the Agent

    The agent, the tools it calls, the retrieval layer and the human-review UI — iterating against the eval set on every push. Weekly demos in your environment, on your data.
    Outputs
    • Agent code in your repo
    • System + tool integrations
    • Review UI (if needed)
    • Run logs
  3. WEEK 10–11
    — evaluation

    Eval & Stress Test

    Run the agent against the full eval set, edge cases and adversarial inputs. If it misses the thresholds, it doesn't ship.
    Outputs
    • Eval report
    • Failure-mode catalogue
    • Mitigation plan
  4. WEEK 12
    — deploy

    Deploy & Runbook

    Production cutover with monitoring, alerts and an on-call playbook. We hand over to your team or keep running it under a managed contract.
    Outputs
    • Production deployment
    • Runbook + on-call playbook
    • Eval-suite cron
    • Cost dashboard
— In production

In production.

Live since Q1 2026 · Chile
AGF · Santiago (confidential) Chile · Asset Manager · UAF-regulated

Regulatory document classifier in 9 days. From a stack of PDFs to a production triage workflow.

The compliance team was triaging ~1,200 inbound regulatory documents a month by hand. We built a Claude-based classifier with RAG provenance, an eval harness across 6 document types, and a human-in-the-loop review queue. Nine days from brief to production. Manual triage down 78% in month one.

“We'll put you on a call with them.”
9d
Brief to production
Including the eval harness, not just the demo.
78%
Manual triage cut
Compliance ops time-back, measured month one.
94%
Eval pass rate
Across 6 document types. F1 not vibes.
0
Hallucinations live
RAG with provenance + retrieval-only answer mode.
— Next step

Tell us the process — we'll tell you if a workflow can run it.