AI case / 02 · Taxonomy, mixed intent, rule-based routing

Designing AI-assisted intake for unstructured commercial work

Email is only the first medium; the problem is turning unstructured inbound communication into structured, owned work—safely.

The case in brief

Current reality

A shared commercial entry point receives RFQs, pricing questions, complaints, service requests, technical questions, leads and order questions—often several in one message—and whoever reads a message first decides where it goes.

What must become true

Every inbound item is quickly classified; decomposed when it holds several intents; associated with candidate business entities; assigned through explicit ownership rules; put on the right operational clock; represented as structured work with its evidence; and escalated to a person when uncertainty or consequence requires it. AI helps understand the message—it never silently assumes commercial authority.

Design question

Which parts of unstructured intake require reasoning, which belong to deterministic routing rules—and where must the system stop for human judgement?

01The reality · current state

What happens today—or in the pilot.

A regional commercial team shares one entry point. RFQs, pricing questions, complaints, service requests, technical questions, leads and order questions arrive together—and some messages ask for two of them at once. Some belong in the CRM, some in the service desk, some in the RFQ workflow, some nowhere. Whoever reads a message first effectively routes it; urgent complaints wait behind newsletters. The proposal on the table is “let the agent handle the inbox”.

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

What the apparent request hides

“Let the agent handle the inbox” hides the questions that decide whether intake is safe.

  • What taxonomy exists?Without written classes, owners and clocks, the model invents its own.
  • Can one message hold several intents?A price request and a damaged delivery need two owners and two clocks.
  • Who owns each class?Ownership is a rule, not something a model should infer.
  • Which routing is deterministic?Once class, identity and region are known, routing is a rule.
  • Which identity match can be trusted?A sender name is a hint; a known contact on one active account is a key.
  • What may be created—and what only drafted?A reversible work draft is not an opportunity, a case or a reply.
  • Can AI communicate externally?An outgoing message can commit a price, a date or liability.
  • What happens when it is unsure?Unsure goes to a person with the evidence—not to the most likely class.

02Work decomposition

Reasoning, rules and authority—kept apart.

The taxonomy exists before the agent does. With it written, the work divides cleanly: reasoning a model does well, rules a system enforces, and authority that stays with people.

AI reasoning

  • Classify against the written taxonomy
  • Detect and separate several intents
  • Extract references—orders, products, dates
  • Find entity candidates with evidence
  • Prepare structured work
  • Draft a reply where appropriate

Deterministic system rules

  • Deduplicate by source message
  • Filter auto-replies and bounces
  • Check identity keys
  • Assign destination and owner from class, account and region
  • Start the operational clock
  • Route known cases; send the rest to human triage

Human authority

  • Sensitive routing ambiguity
  • Identity confirmation with consequence
  • Customer commitment
  • Pricing
  • Delivery promises
  • Substantive external replies

RFQ is one downstream branch. This case owns front-door understanding and routing; once a message is classified as an RFQ, the specialized RFQ workflow owns extraction, product matching, record preparation and authority from there.

03Taxonomy, identity & ownership

The taxonomy exists before the model.

Every class declares its meaning, owner, destination, clock, what may be automated, which decision stays human and what happens downstream. Context is held to the same tests as any other: relevant, trusted, timely, permitted, identified and structured. Identity comes as candidates with evidence; ownership and clocks come from rules. The classes are generic examples, not a production taxonomy.

Taxonomy

Intake taxonomy: meaning, owner and destination, clock, permitted automation, required human decision and downstream process for each class
ClassMeaningOwner & destinationClockPermitted automationHuman decisionDownstream
RFQA request to quote products or quantitiesInside sales · RFQ workflowQuote clockClassify, hand off with the source referenceEverything the RFQ workflow reservesRFQ processing
Pricing requestA question about a price or conditionAccount ownerResponse clockWork draft with account contextAny price stated to the customerQuote or reply
LeadInterest from an unknown or new contactLead qualificationAcceptance clockLead draft with its sourceQualificationLead-to-customer lifecycle
Service issueA complaint, damage, failure or service requestCustomer serviceService clockCase draft with referencesLiability, remedies, anything sensitiveService resolution
Order questionStatus, change or document for an existing orderCustomer serviceService clockCase draft with the order reference; reply draftAny new delivery dateOrder follow-up
Technical questionProduct application or specificationApplication engineeringResponse clockWork draft with product referencesThe technical answerEngineering reply
Other / uncertainNothing fits, or two classes are equally likelyHuman triageTriage clockNone beyond the source referenceClass and routeDecided in triage

Identity

Identity: each reference found, how it is matched, when it is trusted and what happens otherwise
Reference foundMatched byTrusted whenOtherwise
SenderKnown contact on an active accountThe contact is a key on exactly one accountCandidates shown; confirmed where it matters
Company named in the textName and domain similarityNever on its ownA candidate with its evidence
Order numberExact match in the order systemIt exists for a candidate accountFlagged as unverified
Unknown sender——Lead or unmatched queue; no account is created

Ownership & clocks

Ownership and clocks: what decides each, when it starts and what happens when it runs out
RuleDecided byStarts whenWhen it runs out
Destination & ownerClass, confirmed account and regionThe intent is classifiedRe-routed to the team lead
Service clockClassThe intent is classifiedEscalated to the service lead
Triage clockAny intent routed to human triageThe intent is classifiedEscalated to the triage owner
DuplicateSource reference and intentThe message is receivedThe existing work item is returned

04Agent workflow

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

Systems and partiesShared inboxCRMService deskRFQ workflow

Walk the workflow step by step, or select any step.

Step 1 of 8

01Receive

The message gets a source reference—the key for deduplication and traceability; auto-replies and bounces are filtered by rule.

Who acts
System
Authority
Read

02Classify intents

One or more classes from the written taxonomy, each with the sentence that supports it.

Who acts
Agent
Authority
Execute
Tool
Record classification

03Decompose

One work item per intent, all sharing the same source reference.

Who acts
Agent
Authority
Recommend

04Find candidates

Customer, contact, order, product or opportunity references as ranked candidates with evidence.

Who acts
Agent
Authority
Recommend
Tool
Search entity candidates

05Route by rule

Destination, owner and clock from class, identity, region and business rules—never chosen by the model.

Who acts
System
Authority
Execute
Tool
Retrieve ownership rule

06Create work drafts

A reversible draft per intent, keyed by source and intent, with its evidence attached.

Who acts
System
Authority
Draft
Tool
Create work draft

07Draft a reply

Where the class allows it, a reply draft for the owner—nothing that states a price, a date or liability.

Who acts
Agent
Authority
Draft
Tool
Prepare response draft

08Human triage & reply

Where ambiguity is consequential, a person confirms, re-routes or decides; every substantive reply is sent by its owner.

Who acts
Human
Authority
Human approval

05Key dimension · Taxonomy, mixed intent, rule-based routing

Unstructured message in, structured and owned work out

The model understands; the rule routes. Three branches leave the routing rule, and the human-authority boundary sits outside all of them.

From message to owned work
  1. SystemUnstructured message
    • Source reference
    • Deduplicated
    • Auto-replies filtered
  2. AgentIntent(s)
    • One or more classes
    • Supporting sentence per intent
  3. AgentEntity candidates
    • Customer
    • Contact
    • Order
    • Product
    • Opportunity
  4. SystemDeterministic routing rule
    • Safe route
    • Specialist workflow
    • Human triage
  5. SystemStructured, owned work
    • Owner
    • Clock
    • Evidence
    • Source reference

The routing rule’s three branches

The routing rule’s branches: when each applies, who acts and where the work goes
BranchWhenWho actsWhere the work goes
Safe routeA known class, identity confirmed by a key, no commitment at stakeSystemThe owner’s queue, with the clock already running
Specialist workflowA class with its own governed process—an RFQSystemThe RFQ workflow, which owns extraction, matching and authority from there
Human triageUncertain class, identity uncertain with consequence, sensitive content or a possible external commitmentHumanA person, with the evidence and the proposed routing
Human authority—no route crosses it
  • Customer commitment
  • Pricing
  • Delivery promises
  • Substantive external replies
  • Sensitive routing ambiguity
One message, several intents
“Please quote 400 units of the usual part—and by the way, last week’s delivery arrived damaged.”
  • RFQHand-off to the RFQ workflowSpecialist workflow
  • Service issueDamage case draft with the delivery referenceHuman triage

One work item per intent, each carrying the same source reference. The damage report can create liability, so it goes to human triage—the quote request does not wait for it.

Confidence changes how a candidate is presented, not who owns the work. The owner comes from a rule; the model never invents one.

06Tool & authority boundary

Tools that prepare work—none that send it.

Five bounded capabilities, never raw system interfaces. Each contract fixes its input, preconditions, output, side effects, authority and behaviour on failure. The only side effects are reversible and keyed by the source message, so a second run finds what the first created. No tool sends a message. Tool names are descriptive, not implementation names.

Tool contracts: purpose, input and preconditions, output, side effects, authority and behaviour on failure
ToolPurposeInput & preconditionsOutputSide effectsAuthorityOn failure
Record classificationTag each intent with its classOne source message · the written taxonomyClasses with the supporting sentenceA reversible tagExecuteUncertain → human triage
Search entity candidatesFind customer, contact, order, product and opportunity referencesExtracted references · the team’s visibilityRanked candidates with evidenceNoneReadNo candidate → unmatched queue
Retrieve ownership ruleReturn destination, owner and clockClass, confirmed account, regionThe rule’s result and its versionNoneReadNo rule applies → human triage
Create work draftCreate the work item for one intentSource reference and intent · owner assignedA draft keyed by source and intentA reversible draftDraftKey exists → the existing draft
Prepare response draftPrepare a reply for the ownerA class that allows it · no price, date or liabilityA draft in the owner’s queueA reversible draftDraftNot allowed → nothing drafted

Authority by action

Actions and the authority the agent has for each
ActionAuthorityWhy
Classify and tag intentsExecuteReversible, and visible wherever it matters
Split mixed intent into work itemsRecommendEach item keeps the shared source reference
Find entity candidatesRecommendA candidate match is not authority to assert identity
Assign destination, owner and clockReadA deterministic rule assigns; the agent reads the result
Choose or override the commercial ownerForbiddenThe model never invents ownership—rules assign it
Create work item draftsDraftReversible, keyed by source and intent
Draft a replyDraftPrepared for the owner, never sent
Confirm identity where it has consequencesHuman approvalA wrong customer spreads to every later step
Send a substantive replyHuman approvalExternal; it can commit a price, a date or liability
Commit price, delivery or any customer promiseForbiddenCommercial commitment stays with accountable people
Create an accountForbiddenIdentity is owned by the customer lifecycle, not the agent

07Ambiguity policy

When the agent is unsure.

  • More than one intentContinue

    Split into work items sharing the source reference; any consequential part goes to human triage

  • Class uncertainEscalate

    Human triage with the likeliest classes and their evidence

  • Identity uncertain with consequenceEscalate

    Candidates shown with evidence; a person confirms

  • Unknown senderContinue

    Lead or unmatched queue; no account is created

  • Sensitive content (legal, safety)Stop

    Routed to a named owner immediately; nothing is drafted

  • Same message againStop

    The existing work items are returned; nothing new is created

08Human review

Review where ambiguity becomes consequential.

Not after every low-risk classification: a known class with a key-confirmed customer routes without a person. A person steps in at four points—each designed for a quick, well-evidenced decision.

Review points: why review sits there, what the person sees, what they can do and what happens next
Review pointWhy hereWhat the person seesWhat they can doWhat happens next
Mixed intent with a consequential partOne part may create liability or a commitmentThe message, each intent with its supporting sentence, the proposed items and routesAccept, re-route or merge items; ask for more contextWork items are created with owners and clocks; the source reference stays shared
Identity uncertain with consequenceA wrong customer spreads to pipeline, service and reportingRanked candidates, the evidence for each and what differsConfirm one, choose none, or send to the unmatched queueThe rule assigns the owner from the confirmed account
Possible external commitmentA reply could state a price, a date or liabilityThe draft reply, what it would commit and the context behind itEdit, approve and send—or discardThe owner sends; the sent reply is recorded against the work item
Sensitive or unclassifiable contentLegal, safety or reputational content needs a named ownerThe message, the reason it was held and the suggested ownerAssign, escalate or decide the classThe work item is created with the owner the person chose

09Evaluation

Scored per dimension—never one accuracy number.

Evaluation dimensions: what each proves, its question, measure and target
DimensionQuestionMeasureTarget
ClassificationTask qualityRight class for each intent?Correct class per intent on the evaluation set≥ 92 % (illustrative design target)
Mixed-intent recallTask qualityWere all intents found?Intents found ÷ intents present≥ 95 % (illustrative design target)
Entity matchingEvidence & correctnessRight candidates—and none asserted without a key?Correct candidate set; confirmations without a key≥ 95 % · 0 (illustrative design target)
Routing correctnessTask qualityRight destination, owner and clock?Items routed as the rules require≥ 98 % (illustrative design target)
Duplicate workTask qualityOne work item per intent, however often the message arrives?Duplicate items per source message0 (illustrative design target)
Abstention & escalationAuthority complianceDid it stop and escalate when it should?Required escalations honoured≥ 99 % (illustrative design target)
EvidenceEvidence & correctnessDoes each class and candidate show why?Items with a supporting sentence or evidence100 % (illustrative design target)
Unauthorized external actionsAuthority complianceNothing sent, promised or created beyond authority?Unapproved replies, commitments or records0 — release gate

10The 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 handles the inbox end to end, replies included

Pure acknowledgements

Cost: Promises nobody approved; wrong customers
Rejected

Agent classifies and also chooses the owner

None

Cost: Ownership becomes a guess nobody can audit
Selected

AI classifies and prepares; rules assign and time; people decide where it is consequential

Mixed commercial entry points

Cost: A taxonomy with an owner, a triage queue and an evaluation set

Design decision

Use AI to classify, decompose and prepare unstructured intake; use deterministic rules for ownership and routing; treat uncertain identity as evidence-backed candidates rather than established fact; require human review when ambiguity becomes consequential; and never permit an unreviewed external commercial commitment.

Key principleAI classifies; rules assign.

11Audit, governance & learning

Consequential activity can be reconstructed afterwards.

Any routed item can be explained afterwards: what arrived, what the model understood, which rule decided and who acted.

Reconstructable for every consequential run
  1. The source message and its reference
  2. Each intent with its class and supporting sentence
  3. Entity candidates and their evidence
  4. The ownership rule and version applied
  5. Tools called and drafts created
  6. The reviewer’s decision and reason
  7. What was sent, and by whom
  8. The outcome—resolved, re-routed or reopened

Re-routes and corrections improve intake. They do not widen what the agent may do.

Outcomes feed
  • The evaluation set
  • Taxonomy review
  • Routing-rule changes, made by the rule owner
  • Prompt changes, released only against the evaluation set
They never
  • Let the model assign owners
  • Allow an unreviewed reply
  • Retrain on each correction without review

Governed byThe taxonomy owner, with a periodic review alongside sales and service leads

12The second layer

Questions that change the design.

Taxonomy

  1. Who owns the taxonomy—and how does it change?
  2. Which classes may route without a person?
  3. What happens with mixed intent?

Identity & ownership

  1. Which key confirms a customer?
  2. Which rule assigns the owner—and who changes it?
  3. When does each clock start?

Authority

  1. What may be drafted—and what never?
  2. Who sends, and who is accountable for what was sent?
  3. What blocks a release?

13Decisions & outputs

What the work produces.

  1. 01Intake taxonomy
  2. 02Decomposition rule
  3. 03Identity candidate rules
  4. 04Ownership and clock rules
  5. 05Tool contracts
  6. 06Review points
  7. 07Evaluation set