Skip to main content

AI agents

In development

AI Email Agents

Classify, summarize, draft, route, and safely act on operational email with an autonomy level the team controls.

Discuss this service with Valentino AI

What this service does

Reduce email-driven freight work with agents that classify, summarize, draft, route, and execute approved actions.

We do not force freight forwarders to replace the systems that run their business. We connect their technology, improve their data, redesign their workflows, and build AI around the realities of freight forwarding.

Who this is for

Operations, customer service, sales, and IT teams managing shared inboxes and email-driven work.

The problem

The shared inbox is the real operating system, and nobody designed it.

Quotes, instructions, document bundles, and exceptions all arrive in the same place, in no particular order, addressed to no particular person. Coordinators spend their day triaging rather than forwarding, and the cost of a missed thread is a missed load.

Shared inboxes triaged by hand, with the same thread read by three people.

Shipment details re-keyed from email into the TMS.

Document bundles checked manually for what is missing.

Exception threads that go quiet because no one owns them.

Response times that vary by who happens to be watching the inbox.

What the service includes

What we build.

Classification

Every inbound thread classified into your taxonomy — quote request, booking instruction, document submission, exception, or routine query — with confidence recorded.

Thread summaries

Long threads reduced to what changed and what is outstanding, so whoever picks it up does not read forty messages to find the ask.

Entity extraction

Shipment references, lanes, parties, dates, weights, and charges extracted into structured fields, validated against your data before anything downstream uses them.

Attachment checks

Document bundles checked against the expected set for that shipment type, with the gaps surfaced to a human rather than silently accepted.

Approved drafting

Replies drafted from approved templates and grounded in approved data and SOPs, held for human review until acceptance rates justify moving up the autonomy ladder.

Routing, tasks, and controlled actions

Threads routed to the right owner, tasks created in your systems, and system-write actions executed only through permissioned APIs behind an approval gate.

Freight workflows

Where it applies in your operation.

  • Quote emails
  • Shipment instructions
  • Document bundles
  • Exception threads
  • Approved routine replies

The outcome

An inbox that arrives sorted, summarized, and assigned.

The team stops triaging and starts deciding. Autonomy moves one rung at a time — observe, assist, act — and only when the measured acceptance rate for that class of work earns it.

How an engagement works

The engagement.

Autonomy ladder — Observe, Assist, Act, ImproveA reinforcing loop of Observe → Assist → Act → Improve, cycling back to the start, with each pass writing back to trust record at the center.Observeread-only, measure accuracyAssistdrafts held for reviewActcontrolled actions, approvedImprovetune taxonomy and rulesTrust recordacceptance rate, per classRUNGEach rung is earned — autonomy widens only where the measured acceptance rate supports it.
01
Observe

Inventory the inboxes, agree the taxonomy, and run the agent in read-only mode. Measure classification accuracy and thread volume against a human-labelled baseline before it touches anything.

02
Assist

Turn on summaries, extraction, and drafting with every output held for human review. Track draft acceptance per class to see where the agent is actually trustworthy.

03
Act

Promote only the classes that have earned it to controlled action — routing, task creation, and approved replies — with permissioned APIs, validation, and an escalation path.

04
Improve

Review misroutes, corrections, and exceptions on a set cadence. Tune the taxonomy and rules, and run the exception runbook when something goes wrong.

What you receive

Deliverables you keep.

  • Inbox inventory
  • Autonomy ladder with promotion criteria
  • Email taxonomy
  • Routing and action rules
  • Working pilot
  • Evaluation suite and exception runbook

Reference architecture and boundaries

Built alongside your systems of record.

The exact stack should follow the client's systems, data rights, skills, budget, and operating model. A useful reference architecture separates transaction authority from data, intelligence, action, experience, and governance layers.

Reference architecture — five layers, systems of record to governanceLayer stack from systems of record at the foundation up through integration and data, intelligence and action, experience, and governance at the top, with intelligence and action highlighted as the focal layer.OVERSIGHTTRANSACTIONS05GovernanceIdentity, permissions, evaluations, traces, incidents, retention, cost, andcontinuous improvement.04ExperienceDashboards, portals, intranet, chat, email, voice, and human-review queues.03FOCALIntelligence and actionRetrieval, semantic definitions, agents, deterministic rules, andpermissioned tools with validation and approvals.02Integration and dataAPIs, webhooks, files, events, normalization, lineage, quality monitoring,and governed freight entities.01Systems of recordTMS, CRM, accounting, rating, carrier, customer, email, telephony, anddocument systems remain authoritative for transactions.FOCAL LAYERIntelligence and action is where agents and rules touch your data — behind explicit permissions and human approval.

Security and human control

The principles every build follows.

  • Use least-privilege identities and explicit tenant, role, and record-level access.
  • Treat email, documents, web content, and model output as untrusted until validated.
  • Keep systems of record authoritative; agents use controlled APIs rather than casually editing copied data.
  • Require human approval for consequential customer, financial, compliance, or system-write actions until reliability is demonstrated.
  • Retain logs, evaluations, failures, approvals, and ownership information for review.

How success is measured

Measured against a defined KPI.

We baseline these before the build starts, so the improvement is a measurement rather than an impression.

Classification accuracy

Draft acceptance rate

Response time

Touch reduction per thread

Misroute rate

Automation rate by class

Frequently asked questions

The questions forwarders ask first.

Is this a replacement for our existing systems?

Usually no. The service is designed to connect and extend the systems of record the forwarder already uses.

How do you prevent unsafe automation?

Start with narrow workflows, explicit permissions, validation, evaluation, human approval, auditability, and a clear rollback or escalation path.

What is required before publishing results or case studies?

Replace general language with verified client evidence, named systems, measured baselines, and approved proof assets. We do not imply integrations, certifications, or outcomes that have not been confirmed.

Further reading

Go deeper on the thinking behind this.

Source notes

These references support the industry and technical framing on this page. They do not validate claims about any specific client. Claims about client results, integrations, or compliance posture are published only with verified project evidence.

Ready to scope it?

No commitment to start. We'll look at your systems and workflows, then come back with a fixed-scope plan.

Discuss this service with Valentino AIAsk a question first