AI case / 04 · Evidence-backed summary, manager decides

AI-assisted account review with evidence

A fictional agentic case: a review that starts from evidence instead of a slide someone built the night before.

Fictional scenario

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

01The outcome

Each account in the review has a prepared page: the published numbers, what changed, open actions, anomalies and questions—every item cited; decisions and actions are taken and recorded by the manager.

The reality

Monthly account reviews start with managers assembling numbers; the discussion that follows has little time and no shared evidence.

Architecture question

What may the agent prepare for a performance review—and where must it stop short of commercial judgement?

Key decision

Let the agent read published signals and records, summarize and ask; forbid it from recommending commercial strategy, pricing or plan closure.

ContextThe monthly performance review (see Revenue Operations) covers accounts that are At Risk or Underperforming. Managers spend the first half of each review reconstructing context: what changed, what the recovery plan said, which actions are overdue. An agent can prepare this—if it stays on the evidence side of the line.

Systems and partiesData platformCRMPlanning

02The reality · current state

What happens today—or in the pilot.

  • Managers prepare reviews from exports and memory
  • Changes since the last review are not visible
  • Overdue recovery actions are discovered in the meeting
  • A pilot summary recommended a price reduction
  • Numbers in the summary did not match the dashboard
  • Questions for the account owner are improvised

03Agent workflow

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

  1. 01
    Published data

    Target, actuals and segment from the governed data platform—same version as the dashboard.

    SystemReadgetSignals
  2. 02
    Signals

    Changes since last review: segment moves, coverage, overdue plan actions.

    SystemRead
  3. 03
    Context

    Recovery plan, recent activity, open opportunities.

    SystemReadgetAccountContext
  4. 04
    AI summary

    A cited summary, anomalies and three questions for the account owner.

    AgentRecommend
  5. 05
    Manager review

    The manager reads, challenges and decides in the forum.

    HumanHuman approval
  6. 06
    Action

    Plan outcomes and new actions recorded by people, in the CRM.

    HumanHuman approval

04Key dimension · Evidence-backed summary, manager decides

A summary that stops at the evidence

Each item is typed and cited. The limits below are enforced by what the agent can read and by the output format.

  • FactPerformance 76 % of YTD target (published, v3)Performance to target, data platform
  • Open actionTwo of three recovery actions overdueRecovery plan, due dates
  • AnomalyOrders stopped on one product line in AugustOrder lines by product family
  • QuestionHas the customer qualified a second supplier?Suggested for the account owner
The agent never
  • No pricing, discount or strategy recommendation
  • No closing, escalating or changing a recovery plan
  • No numbers that are not the published version
  • No statements about people

The pilot’s price-reduction suggestion was fluent and unsupported. Strategy is the forum’s job; the agent prepares the evidence for it.

05Authority

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

Actions and the authority the agent has for each
ActionAuthorityWhy
Read published signals and recordsReadThe same version the dashboard shows
Summarize, find anomalies, ask questionsRecommendPreparation for a human forum
Recommend pricing or strategyForbiddenCommercial judgement belongs to the forum
Close or escalate a recovery planForbiddenPlan outcomes are decided in review

06Ambiguity policy

When the agent is unsure.

  • Numbers differ from the published versionStop

    The page is not generated; the data owner is notified

  • No recovery plan for an At Risk accountContinue

    Stated as an open action, not filled in

  • Anomaly without clear causeAsk

    Phrased as a question for the owner

07Evaluation

Scored per dimension—never one accuracy number.

Evaluation dimensions: question, measure and target
DimensionQuestionMeasureTarget
FaithfulnessDo numbers match the published version?Exact match100 %
EvidenceIs every item cited?Items with a source100 %
AnomaliesWere the anomalies a reviewer would flag found?Recall on labelled reviews≥ 85 %
AuthorityNo strategy or pricing advice?Out-of-bounds statements0 — release gate

08The trade-offs

Credible options, judged against these premises.

Rejected

Agent recommends the commercial response

None in this context

Cost: Judgement without accountability
Situational

Dashboard only

Small review lists

Cost: Context still reconstructed in the meeting
Selected

Cited summary, anomalies and questions; manager decides

Recurring performance forums

Cost: Output format and limits to govern

09The second layer

Questions that change the design.

Evidence

  1. Which version of the numbers does the summary use?
  2. How are anomalies evidenced?
  3. What if data is late?

Limits

  1. What may the agent never recommend?
  2. Who reviews the output format?
  3. How are out-of-bounds statements detected?

Forum

  1. Who reads the page before the meeting?
  2. Where are decisions recorded?
  3. How is usefulness measured?

10Decisions & outputs

What the work produces.

  1. 01Review page template
  2. 02Signal and source list
  3. 03Statement types and citations
  4. 04Agent limits
  5. 05Evaluation set
  6. 06Forum integration