Carlos EscasainCommercial Systems & Process Architect
ENES

Process case / 04 · Process first → AI second

Designing an RFQ intake process before introducing AI

A fictional case: the intake process an AI agent would need—definitions, ownership, confidence thresholds and human validation—designed before the agent.

Fictional scenario8 min readUpdated 28 Sep 2026
Fictional scenario

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

01The outcome

Every RFQ is captured, classified, identified, owned and answered on a clock—with AI preparing drafts inside explicit limits and people holding commercial authority.

The reality

An AI agent is proposed for RFQs before anyone has defined the process it would run.

Key dimension

Process first → AI second

Key decision

Define the intake process and its human authority points first; introduce AI only where a step is defined, reversible and supervised.

ContextA B2B supplier receives requests for quotation by email, in any format. An AI agent is proposed to read them and create quotes. Before any model is introduced, the operating process has to exist: what an RFQ is, who owns it, and which steps need a person.

Disciplines involvedProcess ArchitectureAI & Agentic WorkflowsAutomation
Systems and partiesMailboxAI agentCRMProduct catalogue

02The reality · current state

What actually happens today.

  • RFQs arrive by email as PDFs, spreadsheets and free text
  • Nobody has written down what counts as an RFQ
  • Customer and product identification depend on who reads the email
  • Requests without a clear owner wait in shared inboxes
  • Response times are unknown, because the clock never starts
  • An AI agent has been proposed to ‘handle RFQs’

03Target process

8 stages, 2 roles, 2 systems.

Each stage sits in the lane that owns it and shows what happens in each system at that moment. Select a stage for its anatomy; read the model through controls, exceptions or automation fit.

Operating model / swimlaneRFQ intakeEmail to the commercial mailbox → Validated quote sent and followed up

Stages, owners, system touches and handoffs.

Inside salesIntake, triage & identification
Account ownerCommercial ownership & validation
AI agentProposes — never commits
CRMIntake queue, drafts & records
Lane / stage
Data authority
Inside sales → Account owner
AI draft → committed record
Parse attachmentsClass + confidenceIntake classAccount candidatesAccount linkLines + matchesDraft linesOwner + clockDraft assemblyValidated quoteFollow-up task
Human actionSystem actionDecisionApproval gateLifecycle stateSystem touchHandoffAuthority transferException return
Stage anatomy · 02 / 08

Classification

DecisionAssist — person confirms
  1. Trigger

    Intake record created

  2. Decision

    Is this an RFQ—or an order, a complaint, a price question or noise?

  3. Owner

    Inside sales

  4. System action

    AI proposes a class with a confidence score; rules route confident non-RFQs

  5. Data state

    Intake · Classified

  6. Next stage

    03 Customer identification

Information required
  • Published definition of an RFQ
  • Mandatory data for a quotable request
Controls
  • The definition of an RFQ is written down before a model is prompted or trained
Exceptions
  • Low confidence — Classified by a person; the case joins the evaluation setDetected: Confidence threshold · Owner: Inside sales · Classified by a person
Automation suitability

Assist — person confirms. Classification is where AI helps first—once the classes are defined.

    1. Trigger

      Intake record created

    2. Decision

      Is this an RFQ—or an order, a complaint, a price question or noise?

    3. Owner

      Inside sales

    4. System action

      AI proposes a class with a confidence score; rules route confident non-RFQs

    5. Data state

      Intake · Classified

    6. Next stage

      03 Customer identification

    Information required
    • Published definition of an RFQ
    • Mandatory data for a quotable request
    Controls
    • The definition of an RFQ is written down before a model is prompted or trained
    Exceptions
    • Low confidence — Classified by a person; the case joins the evaluation setDetected: Confidence threshold · Owner: Inside sales · Classified by a person
    Automation suitability

    Assist — person confirms. Classification is where AI helps first—once the classes are defined.

04Key dimension · Process first → AI second

Process first, AI second

Each step is defined as a process rule before any model touches it. Only then is it clear what AI may propose—and what a person must decide.

Step1 · Process rule first2 · What AI may propose3 · What a person decides
ClassificationA written definition of an RFQ and its mandatory dataProposes a class with a confidence scoreClassifies when confidence is low
Customer identificationOne legal entity per request; no silent account creationRanks account candidates, with evidenceConfirms the identity
Product identificationSubstitutions are proposals, never replacementsExtracts lines; matches catalogue and cross-referencesResolves ambiguous products
Commercial ownershipExactly one owner before any record existsNot used—ownership is a ruleCommercial operations closes rule gaps
Quote draftDrafts are invisible to customers and excluded from pipelineAssembles the draft from confirmed dataReviewed in the next step
Human validationOnly a person validates, even at high confidenceShows the source and confidence of each valueValidates or corrects; corrections become evaluation data

Without the left column, an agent would automate a process nobody has agreed. With it, the agent has a job description, limits and a supervisor.

05Ownership & decision rights

One accountable role per decision.

Decision rights for RFQ intake
DecisionInside salesAccount ownerProduct specialistCommercial opsAI product owner
02Define what counts as an RFQC ConsultedNot involvedNot involvedA AccountableC Consulted
02Set confidence thresholdsC ConsultedNot involvedNot involvedA AccountableR Responsible
03Confirm the customerA AccountableC ConsultedNot involvedNot involvedNot involved
04Resolve an ambiguous productR ResponsibleNot involvedA AccountableNot involvedNot involved
06Allow automation to create draftsNot involvedC ConsultedNot involvedA AccountableR Responsible
07Send the quoteI InformedA AccountableNot involvedNot involvedNot involved
  1. 02Define what counts as an RFQ
    AAccountable
    Commercial ops
    CConsulted
    Inside sales · AI product owner
  2. 02Set confidence thresholds
    AAccountable
    Commercial ops
    RResponsible
    AI product owner
    CConsulted
    Inside sales
  3. 03Confirm the customer
    AAccountable
    Inside sales
    CConsulted
    Account owner
  4. 04Resolve an ambiguous product
    AAccountable
    Product specialist
    RResponsible
    Inside sales
  5. 06Allow automation to create drafts
    AAccountable
    Commercial ops
    RResponsible
    AI product owner
    CConsulted
    Account owner
  6. 07Send the quote
    AAccountable
    Account owner
    IInformed
    Inside sales
AAccountableRResponsibleCConsultedIInformedExactly one A per decision

06Exceptions & failure paths

8 exception paths, each with an owner and a destination.

01 Incoming RFQUnreadable attachment

Routed to manual triage with the original file

Detected by
Parser
Owner
Inside sales
Route
Manual triage
02 ClassificationLow confidence

Classified by a person; the case joins the evaluation set

Detected by
Confidence threshold
Owner
Inside sales
Route
Classified by a person
03 Customer identificationCustomer unclear or unknown

Clarified with the sender, or routed to lead intake—no silent account creation

Detected by
Match confidence
Owner
Inside sales
Route
Clarify with the sender, or route to lead intake
04 Product identificationAmbiguous or unknown product

Resolved by a specialist; the resolution enriches the cross-reference list

Detected by
Match confidence
Owner
Product specialist
Route
Specialist review
05 Commercial ownershipNo owner resolvable

Assigned by Commercial operations; the rule gap is logged

Detected by
Assignment rule
Owner
Commercial operations
Route
Ownership queue
06 Opportunity & quote draftNo price available

Draft kept; special price requested

Detected by
Pricing
Owner
Account owner
Route
Special-price request
07 Human validationDraft wrong

Lines corrected; the correction feeds the evaluation set

Detected by
Reviewer
Owner
Account owner
Route
Returns to 04 Product identification
08 Follow-upNo response by the follow-up date

Followed up, then closed with a reason code

Detected by
Follow-up date
Owner
Account owner
Route
Follow-up, then close with a reason

07Process ↔ system

The process question first. Then the system question has an answer.

StageAccountableSystem actionData stateAutomation
01Incoming RFQInside sales (queue owner)Each email becomes an intake record with its attachments and received timeIntake · ReceivedAutomate
02ClassificationInside salesAI proposes a class with a confidence score; rules route confident non-RFQsIntake · ClassifiedAssist — person confirms
03Customer identificationInside salesAI ranks account candidates with evidence; the CRM shows them for confirmationIntake · Customer confirmedAssist — person confirms
04Product identificationInside sales · product specialist when ambiguousAI extracts lines and proposes catalogue matches, including customer cross-referencesIntake · Lines identifiedAssist — person confirms
Accountability transfers — Inside sales → Account owner
05Commercial ownershipCommercial operations (rules) · Account owner (accepts)The CRM assigns by account ownership rules and starts the response clockIntake · OwnedAutomate
06Opportunity & quote draftAccount owner (accountable) · automation (prepares)Automation prepares an opportunity and a quote draft, marked ‘draft — not reviewed’Quote · Draft (not reviewed)Automate
07Human validationAccount ownerA review screen shows each proposed value with its source and confidenceQuote · ValidatedKeep human
Data authority transfers — AI draft → committed record
08Follow-upAccount ownerThe CRM schedules follow-up and records the responseQuote · SentAssist — person confirms
Process question: P1What qualifies as an RFQ, and what information is mandatory?
System question: S1Which classes and required fields does the model return, with what confidence?
Process question: P2Who owns an incomplete or unclear request?
System question: S2Which queue holds it, and how is its ageing shown?
Process question: P3When may automation create commercial records?
System question: S3Which records are created as drafts, and what excludes them from pipeline?
Process question: P4What happens when confidence is low?
System question: S4Which threshold routes to a person, and how are corrections captured?

08The second layer

Questions that change the design.

Definition

  1. What qualifies as an RFQ?
  2. What information is mandatory before a quote can be prepared?
  3. Who owns incomplete requests?

Ambiguity

  1. What happens when customer identity is unclear?
  2. What happens when products are ambiguous?
  3. What happens when confidence is low?

Authority

  1. When can automation create commercial records?
  2. Which actions require human validation, even at high confidence?
  3. Who owns the model’s mistakes?

09Measurement

Process health, defined by owner and action.

No values are shown: in a design, the definition is the deliverable.

LaggingStages 01 → 08
Time to first response

Received to quote sent — median and 90th percentile

Owner
Sales management
Triggers
Find whether the waiting sits in triage, identification or validation
ControlStages 02 → 04
Manual intervention rate

Requests that needed a person before validation, by stage and reason

Owner
Commercial operations
Triggers
Fix definitions and master data before tuning the model
ControlStage 07
Correction rate at validation

Draft fields changed by the reviewer, by field

Owner
AI product owner
Triggers
Change the model only where corrections show a clear pattern
ControlStage 05
Unowned requests

Requests still without an owner after assignment

Owner
Commercial operations
Triggers
Close the gap in the ownership rules
LaggingStage 08
Win rate of email RFQs

Quotes won from email RFQs against other channels

Owner
Sales management
Triggers
Check whether faster answers change outcomes

10What the work produces

Outputs and connections.

  1. 01RFQ definition & mandatory data
  2. 02Intake lifecycle
  3. 03Ownership & assignment rules
  4. 04Human validation points
  5. 05AI boundary & confidence thresholds
  6. 06Evaluation loop from corrections
Disciplines involvedProcess ArchitectureAI & Agentic WorkflowsAutomation