AI case / 01 · Authority, tools & human commit

Designing an AI-assisted RFQ intake agent

A fictional agentic case: the authority, tools, review and evaluation that let an agent prepare RFQs without being able to commit them.

Fictional scenario

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

01The outcome

Every RFQ is drafted in minutes with evidence for each value; ambiguity is flagged, not guessed; nothing reaches pipeline or the customer without a named person’s approval; every run can be reconstructed.

The reality

RFQs arrive as free-text email and attachments, and an agent is expected to “handle them”—with nobody having said what it may do.

Architecture question

When is a customer match trusted, what happens with an ambiguous product—and what may the agent never do, however confident it is?

Key decision

Give the agent execute authority only over reads and reversible drafts; make customer confirmation a recommendation and every commit a human approval—enforced by the tools, not by the prompt.

ContextA B2B supplier receives requests for quotation by email in every format. The intake process has already been designed (see the process case): what an RFQ is, who owns it, which steps need a person. This case designs the agent that runs the reversible part of that process—and the boundary it cannot cross.

Systems and partiesMailboxAI agentCRMProduct catalogue

02The reality · current state

What happens today—or in the pilot.

  • Inside sales retype RFQ lines into the CRM from PDFs and spreadsheets
  • Customer identification depends on who reads the email
  • Product references use customer and competitor codes
  • A proof of concept let the model create opportunities directly
  • Nobody can say why the proof of concept picked a product
  • The same email forwarded twice produced two opportunities

03Agent workflow

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

  1. 01
    Receive email

    The message and attachments get a source request ID—the idempotency key for the whole run.

    SystemRead
  2. 02
    Classify request

    RFQ, order question, complaint or other—against the written RFQ definition.

    AgentExecutesetClassification
  3. 03
    Identify customer

    Ranks account candidates by domain, signature and history, with evidence.

    AgentRecommendsearchAccount
  4. 04
    Identify products

    Matches each line to catalogue or cross-reference; flags lines with more than one plausible product.

    AgentRecommendmatchProducts
  5. 05
    Prepare CRM draft

    Creates a draft with every value and its source; invisible to pipeline and customer.

    AgentDraftcreateOpportunityDraft
  6. 06
    Human review

    The account owner approves, edits, rejects or asks for more context—with a reason code.

    HumanHuman approvalrequestReview
  7. 07
    Commit

    The system—not the agent—turns the approved draft into an opportunity.

    SystemExecute

04Key dimension · Authority, tools & human commit

Seven questions the agent design must answer

Each answer is a rule the tools and the review model enforce—not a sentence in a prompt.

  1. 01When is a customer match trusted?

    When a key matches (a known contact on an active account) or a person confirms a candidate. A high similarity score alone is a recommendation.

  2. 02What if product identification is ambiguous?

    The line is flagged with its alternatives and evidence. The agent never substitutes; the reviewer chooses or asks the customer.

  3. 03Which records can be drafted?

    Opportunity drafts only—invisible to pipeline, forecast and customer, and expiring if not reviewed.

  4. 04Which actions require approval?

    Creating the opportunity, any customer communication and anything that commits price or delivery.

  5. 05What evidence is shown?

    For every value: the source (email line, record, history) and the confidence—plus what approval will do.

  6. 06What if no customer is found?

    A lead in a holding queue with an owner. The agent never creates an account.

  7. 07What if confidence is high but the consequence is high?

    Human approval. The score changes how the draft is presented, not who commits it.

The proof of concept failed on the last question: a 97 % match on the wrong legal entity created a real opportunity. Confidence was treated as authority.

05Authority

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

Actions and the authority the agent has for each
ActionAuthorityWhy
Classify the requestExecuteReversible, visible and checked by the reviewer
Search accounts and catalogueExecuteRead-only, within the requester’s visibility
Confirm customer identityRecommendA wrong match spreads to every later step
Create opportunity draftExecuteInvisible and reversible by design
Create opportunityHuman approvalEnters pipeline and forecast
Reply to the customerHuman approvalExternal and irreversible
Create accountForbiddenIdentity is owned by the customer lifecycle, not the agent

06Ambiguity policy

When the agent is unsure.

  • Classification below thresholdAsk

    Routed to inside sales for classification

  • Two customer candidatesEscalate

    Both shown with evidence; the owner confirms

  • No customer candidateStop

    Lead in a holding queue; no draft is created

  • Ambiguous product lineContinue

    Draft continues with the line flagged and alternatives listed

  • Mandatory data missingRetrieve

    Checks history; otherwise asks the requester

  • Same source request seen againStop

    Returns the existing draft; nothing new is created

07Evaluation

Scored per dimension—never one accuracy number.

Evaluation dimensions: question, measure and target
DimensionQuestionMeasureTarget
ClassificationIs it an RFQ?Correct class on the evaluation set≥ 95 %
ExtractionAre lines, quantities and dates captured?Required fields correct≥ 90 %
Entity matchingRight account, right products?Correct top candidate; correct set in top three≥ 90 % · ≥ 98 %
AuthorityDid it stop when it had to?Required stops honoured100 % — release gate
EvidenceIs every value sourced?Values with a cited source≥ 95 %

08The trade-offs

Credible options, judged against these premises.

Rejected

Agent creates opportunities directly above a confidence threshold

Low-value, known customers, standard products

Cost: Wrong entities reach pipeline; confidence becomes permission
Rejected

Extraction only; people do everything else

Very low volume

Cost: Most of the clerical effort remains
Selected

Agent drafts with evidence; people commit

Mixed customers and ambiguous product references

Cost: A review step and a review card to design

09The second layer

Questions that change the design.

Identity

  1. Which key confirms a customer?
  2. What happens with a new contact at a known account?
  3. Who owns the holding queue?

Authority

  1. Which tools can write—and what exactly?
  2. Which checks live in the tool, not the prompt?
  3. Who can widen the agent’s authority?

Quality

  1. Which cases are in the evaluation set?
  2. How do corrections become new cases?
  3. What blocks a release?

10Decisions & outputs

What the work produces.

  1. 01Agent workflow
  2. 02Authority matrix
  3. 03Tool contracts
  4. 04Ambiguity policy
  5. 05Review card
  6. 06Evaluation set
  7. 07Audit model