AI case / 03 · Taxonomy, routing & mixed intent

Turning commercial email into structured work

A fictional agentic case: from a shared inbox nobody owns to classified, owned work with a clock.

Fictional scenario7 min readUpdated 29 Sep 2026

Fictional scenario

Independently created. Contains no employer or client implementation detail, internal names or figures.

01The outcome

Every email is classified, matched, owned and on a clock within minutes; safe classes route automatically; mixed or uncertain emails go to a person; the agent drafts replies but never sends them.

The reality

A shared commercial inbox mixes RFQs, pricing questions, complaints, technical questions, leads and order questions—and response time depends on who looks first.

Architecture question

Which classes are safe to route automatically, what happens with mixed intent—and may the agent create tasks or send replies?

Key decision

Route automatically only classes that are deterministic and reversible, split mixed-intent emails into separate work items, and keep every outgoing reply behind human approval.

ContextA regional commercial team shares one inbox. Several hundred emails a week arrive in six broad kinds. Some belong in the CRM, some in the service desk, some are noise. Items wait until someone recognizes them; urgent complaints sit behind newsletters. An agent is proposed to “sort the inbox”.

Disciplines involvedAI & Agentic Workflows Process Architecture Automation

Systems and partiesMailboxAI agentCRMService desk

02The reality · current state

What happens today—or in the pilot.

Nobody owns an email until someone opens it
Complaints and order questions wait behind marketing mail
The same email is answered by two people—or none
Customer identity is guessed from the sender name
One email often asks for a price and complains about a delivery
A pilot sent automatic replies that promised dates

03Agent workflow

Each step with its actor, its tool and its authority.

01
Receive email
Message ID as key; auto-replies and bounces filtered by rule.
SystemRead
02
Classify
One or more classes from the taxonomy, each with evidence.
AgentExecutesetClassification
03
Identify entities
Customer, contact, order or opportunity references—as candidates.
AgentRecommendsearchAccount
04
Assign owner
Ownership by rule from class, account and region—not by the model.
SystemExecute
05
Draft action
Creates the work item and, where useful, a draft reply for the owner.
AgentDraftcreateTaskDraft
06
Route
Automatic for safe classes; triage queue for mixed, uncertain or sensitive ones.
HumanHuman approval

04Key dimension · Taxonomy, routing & mixed intent

A taxonomy with routes, owners and actions

Each class declares whether it can route without a person, who owns it and what the agent may prepare.

Email taxonomy: route, owner and prepared action
Class	Route	Owner	Agent prepares
RFQ	Automatic	Inside sales	Hand to the RFQ intake agent
Lead enquiry	Automatic	Lead qualification	Create lead draft with source
Order question	Automatic	Customer service	Service case with order reference
Pricing request	Human	Account owner	Task with context; no price in any draft
Technical question	Human	Application engineering	Task with product references
Complaint	Human	Customer service lead	Case flagged urgent; draft acknowledgement for review

Mixed intentA mixed-intent email becomes one work item per intent, linked to the same message—and goes to the triage queue if any intent is human-routed.

“Automatic” means routing and preparation. No class allows the agent to send a reply without a person.

05Authority

What the agent may do—and what it may not.

Actions and the authority the agent has for each
Action	Authority	Why
Classify and tag the email	Execute	Reversible and reviewed in the queue
Match customer and references	Recommend	Owners confirm identity on their items
Create work item drafts	Execute	Drafts, keyed by message ID
Assign owner	Read	A rule assigns; the agent reads the result
Draft a reply	Draft	Prepared for the owner, never sent
Send a reply	Human approval	External, can promise dates or prices

06Ambiguity policy

When the agent is unsure.

More than one intentContinue
Split into linked items; triage queue if any is human-routed
Class confidence lowEscalate
Triage queue with the top two classes
Unknown senderContinue
Lead or unmatched queue; no account created
Sensitive language (legal, safety)Stop
Routed to a named owner immediately

07Evaluation

Scored per dimension—never one accuracy number.

Evaluation dimensions: question, measure and target
Dimension	Question	Measure	Target
Classification	Right class or classes?	Exact class set on the evaluation set	≥ 92 %
Mixed intent	Were all intents found?	Recall of intents per email	≥ 95 %
Routing	Right queue and owner?	Correct destination	≥ 97 %
Authority	Nothing sent without approval?	Unapproved outgoing messages	0 — release gate

08The trade-offs

Credible options, judged against these premises.

Situational

Rules on subject lines and senders

Stable, templated senders

Cost: Misses free-text and mixed intent

Rejected

Agent classifies and replies automatically

Pure acknowledgements

Cost: Promises nobody approved

Selected

Agent classifies and prepares; rules route; people reply

Mixed commercial inboxes

Cost: A taxonomy with an owner and a triage queue

09The second layer

Questions that change the design.

Taxonomy

Which classifications are deterministic enough to route automatically?
Who owns the taxonomy?
What happens with mixed-intent emails?

Identity

How is customer identity matched?
What happens with unknown senders?
Which references must be confirmed?

Authority

Can the agent create a task?
Can it send a reply?
When is a human required?

10Decisions & outputs

What the work produces.

Architecture decisionADR 013 — Grant agent authority per action and consequence, never by confidence

01Email taxonomy
02Routing model
03Authority model
04Mixed-intent rule
05Triage queue design
06Evaluation set

Disciplines involvedAI & Agentic Workflows Process Architecture Automation