Preparing account context before a commercial visit
A fictional agentic case: a visit briefing a salesperson can trust—because it shows where every sentence came from.
Independently created. Contains no employer or client implementation detail, internal names or figures.
01The outcome
A one-page briefing assembled from authoritative sources with their freshness shown, every statement cited, inferences labelled, conflicts surfaced—and nothing written to any record without the salesperson.
Salespeople prepare visits by opening six screens—or not at all—and an AI summary is proposed that may invent what it cannot find.
Which sources are authoritative, how fresh must they be—and how does the briefing separate fact from inference?
Assemble context deterministically from owned sources first; let the model only summarize, compare and suggest—labelled and cited—and give it no write tools.
A field salesperson visits several customers a week. The information exists—account data, open opportunities, orders, invoices, performance, open actions, notes from the last visit—but in different systems, with different freshness. A generative summary is attractive and risky: it reads well even when a source is missing or stale.
02The reality · current state
What happens today—or in the pilot.
- Visit preparation depends on each salesperson’s habits
- Performance, orders and open actions live in different places
- A pilot summary mentioned an order that had been cancelled
- Nobody could tell which statements were facts and which were guesses
- Summaries were pasted into notes and became “facts” later
- Data conflicts between CRM and ERP were silently averaged
03Agent workflow
Each step with its actor, its tool and its authority.
- 01Upcoming visit
A visit on the calendar for a known account triggers preparation the day before.
- 02Context assembly
Account, opportunities, orders, invoices, performance, open actions and last visit notes—each from its owner.
- 03Signal selection
Only published signals (segment, coverage, changes since last visit) with their version.
- 04AI summary
Summarizes, highlights changes and conflicts; every sentence cites a source.
- 05Suggested focus
Three questions to ask and open issues to close—labelled as suggestions.
- 06Salesperson review
The salesperson reads, dismisses or keeps suggestions; notes are written by the person.
04Key dimension · Context assembly, fact vs inference
Sources with standing—and a briefing that labels every line
Each source has a standing, a freshness and a use. The briefing below shows how statements are typed and cited.
| Source | Standing | Freshness | Used for |
|---|---|---|---|
| CRM account and contacts | Authoritative | Live | Who, where, relationship owner |
| Open opportunities | Authoritative | Live | Pipeline and next steps |
| ERP orders and invoices | Authoritative | Daily | What was bought and invoiced |
| Performance segment | Derived | Monthly, versioned | On Track, At Risk… with definition version |
| Last visit notes | Authoritative | As written | Commitments made last time |
| Email threads | Excluded | — | Not read: consent and relevance not established |
- FactOrders are 18 % below the same period last year.ERP orders, refreshed this morning
- ChangeThe account moved from On Track to Underperforming this month.Segment v3, published on the 3rd
- InferenceThe drop is concentrated in one product line.Derived from order lines; not confirmed by the customer
- QuestionAsk whether volume moved to a second supplier.Suggested focus—dismiss if not relevant
The pilot’s cancelled order came from a cached export. Freshness is now shown next to every source, and stale sources are left out rather than summarized.
06Ambiguity policy
When the agent is unsure.
- Source older than its freshness ruleStop
Left out of the briefing, with a note that it is stale
- CRM and ERP disagreeContinue
Both values shown as a conflict—never averaged
- Statement cannot be citedStop
Removed from the briefing
- No activity since last visitContinue
Stated as a fact, not filled with guesses
07Evaluation
Scored per dimension—never one accuracy number.
| Dimension | Question | Measure | Target |
|---|---|---|---|
| Evidence | Is every statement cited? | Statements with a valid source | 100 % — release gate |
| Faithfulness | Does each statement match its source? | Reviewer-verified statements | ≥ 98 % |
| Labelling | Are inferences labelled as inferences? | Correct statement type | ≥ 95 % |
| Usefulness | Did the salesperson use it? | Suggestions kept or acted on | Tracked, no target yet |
08The trade-offs
Credible options, judged against these premises.
Let the model search every system freely
Exploration
Cost: Unknown sources, stale data, no citationsA static report, no AI
Stable, simple accounts
Cost: Changes and conflicts still need reading between screensDeterministic assembly, AI summary with citations and labels
Visits that depend on recent change
Cost: Freshness rules and a citation format to maintain09The second layer
Questions that change the design.
Sources
- Which sources are authoritative?
- What data freshness is required?
- Which sources are deliberately excluded?
Truthfulness
- Should recommendations distinguish fact from inference?
- How are citations and evidence shown?
- What happens when data conflicts?
Use
- Who may write what after the visit?
- How is usefulness measured?
- How are wrong statements reported?
10Decisions & outputs
What the work produces.
- 01Context model with source standing
- 02Freshness rules
- 03Citation format
- 04Fact / inference taxonomy
- 05Briefing template
- 06Evaluation set