Skip to content

Playbooks · White paper · November 14, 2026 · 8 min read

Teaching AI how we work

A proposed, human-reviewed model for turning discovery conversations into client briefs.

Playbooks · White paper

Teaching AI how we work

A strategist finishes a discovery call and opens a blank brief. The client has explained why the last campaign disappointed them, which customers they want to reach, and what their sales team keeps hearing. A writer joining the project next week needs enough of that conversation to make useful decisions without asking the client to repeat it. The writer also needs the reason an earlier direction was rejected, the claim legal still needs to verify, and the difference between a requested launch date and a date the agency agreed to meet.

That handoff is where we start with AI at Voyager. We recommend a model that gives an assistant a defined assignment, approved information, explicit limits, and a named human reviewer. The assistant prepares a draft. Our people resolve uncertainty, make the creative decisions, and approve anything that leaves the team.

Video summary · 1:35

Teaching AI how we work

A proposed discovery-to-brief model. A short walkthrough of the four chapters below. Captions are on by default.

Narration is AI-generated. The diagrams and examples are illustrative, and the model described is a proposal, not a measured result.

Full transcript with visual descriptions

1. Client and agency knowledge 0:00–0:23

On screen: Three separate boxes show client knowledge for the assigned account team, agency procedures for authorized staff, and assignment context for the named project team. Source, owner, and review date appear below. A boundary note prohibits cross-client pooling.

  • At Voyager, we recommend a model for turning discovery
  • notes into a client brief. These diagrams and examples
  • are illustrative.
  • Keep client knowledge, agency procedures, and assignment context separate.
  • Give each approved fact a source, an owner, and
  • a review date.
  • Limit access to the assigned team. Label assumptions and
  • leave conflicting evidence unresolved until the account lead checks
  • it.

2. Instructions and examples 0:23–0:47

On screen: An illustrative instruction card lists objectives and deliverables, approved sources and citations, unknowns, and decision owners. An illustrative discovery record labels distributor sales as client-reported and the call to action as unconfirmed. The account lead must resolve it.

  • Give the assistant one bounded assignment: prepare a draft
  • brief from approved sources, with objectives, deliverables, citations, and
  • open questions.
  • For example, a client may report distributor sales. That
  • does not confirm which call to action the team
  • should recommend.
  • Mark the missing answer as unknown. Test the instructions
  • against permitted examples, including incomplete notes and contradictory sources.

3. Review and approvals 0:47–1:12

On screen: Five connected stages show approved discovery notes, source checks, an assistant draft labeled Draft only, lead review labeled Human decision, and an approved brief with its version recorded. A note requires human approval before client delivery. Full-job costs include preparation, generation, review, corrections, and run cost.

  • Follow five stages: approved discovery notes, source checks, assistant
  • draft, account-lead review, and an approved brief.
  • The account lead checks claims, dates, and unresolved questions,
  • then approves a specific version. Drafting permission does not
  • authorize client delivery.
  • Compare the full job, including preparation, generation, review, corrections,
  • and run costs. Keep negative results when assistance costs
  • more.

4. Revisions and new assignments 1:12–1:35

On screen: Five boxes connect correction, proposed change, retest, named approval, and versioned release. A dashed return line indicates the revision loop. The team retains the previous approved instructions. The final note directs viewers to the full article and its interactive diagrams.

  • When reviewers find a repeated problem, propose an instruction
  • change. Keep it separate from approved instructions while the
  • owner retests it.
  • A named reviewer approves the revision. Record the version
  • and decision, and retain the previous instructions if the
  • change fails.
  • Keep client-specific corrections within that client's permissions. Use the
  • full article for the detailed model, interactive diagrams, and
  • review criteria.

Client and agency knowledge

A discovery transcript records what people said. A working knowledge set also records which statements are confirmed, where each statement came from, and who can decide whether it applies to the brief. We recommend keeping client knowledge, agency procedures, and assignment context in separate, labeled areas. A writer can then see whether a sentence comes from an approved client fact, an agency rule, or a working assumption for this assignment.

An approved fact needs a source reference, an owner, and a review date. A product description might point to a client-approved page and identify the person responsible for confirming future edits. A discovery note might cite the transcript, speaker, and relevant time marker. The record should show when the fact was captured, when someone last checked it, and which work may use it.

Assumptions need a different label and an owner who can resolve them. In the discovery-call example, the client may describe distributor sales while the team wonders whether a direct online purchase matters. The assistant can flag that question and retain it in the brief's open-items section. It cannot turn the team's guess into a client requirement. If two notes conflict, the record should preserve both sources and mark the point unresolved. The account lead checks the evidence and asks the client when the decision affects the work. A newer note does not become authoritative by age alone.

Agency procedures belong in a separate area because they apply across work without carrying one client's confidential material. A procedure can explain how we structure briefs, label evidence, and record approval. Assignment context stays narrower. The current discovery notes, source log, draft, and review comments belong to the named project team. Access follows the work people are assigned to do; a different client team does not receive those files as a convenient example library.

Each document needs a maintainer, an approval state, and a review date. Meeting transcripts and approved decision notes serve different purposes: one preserves the discussion, while the other records the decision, its owner, and the follow-up. Before a transcript enters a knowledge system, the team confirms recording consent, sets access and retention rules, and removes material that the assignment does not require. A folder labeled final_final_v7 gives an assistant a fair introduction to agency life.

Knowledge boundaries

Useful context, kept in its lane

Three distinct knowledge areas make provenance visible. Assignment context stays scoped to the named job and its permitted team.

Conceptual map, not a quantitative chart. Select a boundary to inspect its example documents and permissions.

Client knowledgeAgency proceduresAssignment context

Boundary 01 · Selected area

Client knowledge

Facts, preferences, and decisions belonging to one client.

Maintainer
Account lead and named client owner
Approval status
Client-confirmed facts; review date recorded
Permitted role
Assigned account team only
Example documents and records
  • Approved discovery notesMaintainer: account lead · Approval: client-confirmed
  • Brand voice and terminologyMaintainer: client owner · Approval: approved for this account

Hard boundaryNo cross-client pooling. Client knowledge is neither combined with another client’s material nor made available to other client teams.

Text equivalent: client knowledge is limited to its assigned account team; agency procedures are available to authorized agency staff; assignment context is limited to its named project team. Each example has a named maintainer and approval state.

Selected area: Client knowledge. Maintainer: Account lead and named client owner. Approval status: Client-confirmed facts; review date recorded. Permitted role: Assigned account team only. Example documents: Approved discovery notes, Maintainer: account lead · Approval: client-confirmed; Brand voice and terminology, Maintainer: client owner · Approval: approved for this account.

Instructions and examples

Once the sources have clear boundaries, the team can write instructions for one recurring task. For a discovery-call brief, the instructions should ask the assistant to identify the business objective, list requested deliverables, preserve the client's terminology, and distinguish confirmed requirements from assumptions. Each important statement should carry a source reference when the source is available. A brief should also name the decision owner and record the date the assignment was approved.

Missing information needs a defined destination. The assistant can mark a field as unknown, cite the part of the conversation that leaves it open, and add a question for the account lead. It should not fill a gap with a plausible deadline or a familiar industry pattern. If the missing answer changes the strategy, audience, claim, or scope, the instructions should direct the writer to pause the recommendation until a person resolves it. Routine omissions can stay in a review list when they do not change the next decision.

Examples make those rules concrete. We recommend choosing past briefs and discovery materials that the client agreement permits the team to reuse, then removing details the new assignment does not need. An approved example can show how a writer separates the business problem from the requested asset. A returned draft can show why a research question stays open and why a sales date in a call does not become a delivery commitment. Each example needs an explanation of the decision it teaches.

The team can test the instructions against permitted past examples before relying on them for a live assignment. The reviewer compares each draft with the original notes and records whether it preserves required details, labels assumptions, cites its sources, and leaves unresolved questions visible. The test should include cases where the source is incomplete or contradictory. A useful instruction set tells the assistant what to do in those cases instead of rewarding a confident guess.

Document owners review the examples and instructions on a stated schedule and when the client changes its positioning, claims, or approval path. The owner records which examples remain approved, which have expired, and what the current instructions expect. That review keeps a brief connected to current decisions instead of a collection of files someone once thought useful.

Human-controlled workflow

From discovery to an approved brief

AI prepares a draft inside a defined assignment. People check the evidence, resolve uncertainty, and own approval.

01Discovery notesApproved source notes02Source checksFacts and assumptions03Assistant draftDraft only04Lead reviewHuman decision05Approved briefApproved brief
Text equivalent: approved discovery notes → source and assumption checks → assistant draft → account-lead review → approved brief. The assistant draft is never presented as an approved deliverable.

Approved discovery notes

Stage 1 of 5 · Account lead

Actor
Account lead
Inputs
Client interview notes, goals, constraints, and agreed terminology.
Output
A bounded set of approved facts and open questions.
Approval state
Approved by the client-facing account lead before use.

Selected stage for print: Approved discovery notes. Account lead receives Client interview notes, goals, constraints, and agreed terminology. The output is A bounded set of approved facts and open questions. Approval state: Approved by the client-facing account lead before use.

Review and approvals

We recommend a clear path from discovery notes to approved brief. The account lead first approves the source set and identifies open questions. A researcher or account lead checks material claims against their sources and labels assumptions. The assistant drafts within that bounded set. The account lead then compares the draft with the transcript and decision notes, resolves the questions that need a client answer, and records whether the brief is approved for its next use.

That approval has a scope. An assistant preparing a brief may read the approved notes and save a draft in the assigned project space. That permission does not allow it to send the brief to a client, publish a claim, commit budget, or change a client system. Each action needs an appropriate role and a separate approval from a named person. The project record should show the approver, the version reviewed, the decision, and the date. A draft and an approved brief must remain visibly different states.

Confidentiality rules apply at each step. The assistant receives only material required for the assignment, under the permissions agreed for that client. The team checks where the tool stores or processes the information, how long it retains inputs, and whether the contract permits that use. A reviewer also checks that examples and citations do not expose information from another engagement. If a source cannot be shared with the assistant, a person can perform the source check and provide only the approved finding.

A review should capture the work around generation. The team records time spent gathering and preparing information, generation, human review, correction, and any direct run cost. A comparison with the previous process uses the same assignment boundary on both sides. If an assistant creates a quick first draft that takes longer to correct, the full-job total shows that cost. The calculator below uses editable illustrative inputs to make the arithmetic visible; the values describe no Voyager assignment.

The team can use a baseline from its own records to decide whether a workflow is useful enough to continue. It should state what the comparison includes, use the same units, and retain a negative result when the assisted process costs more time or money. That result can lead to narrower instructions, a different review path, or a decision to stop. A model earns wider use through evidence from the actual workflow and the people responsible for it.

For the discovery-call brief, the account lead checks the objective, audience, deliverables, dates, claims, source references, and unresolved questions against the call and approved client records. The reviewer records each correction against its source. A missing audience, unapproved deliverable, unsupported claim, or inferred deadline has a clear place to return to. The client receives the brief only after the assigned reviewer approves that version for delivery.

Compare the full job

Count the work around the draft

Include setup, generation, human review, corrections, and run cost. Adjust every assumption; the comparison updates locally.

Time and direct cost assumptions

Illustrative example inputs. Replace them with one comparable job from your own records; these values are not Voyager results.

minutes
minutes
minutes
minutes
minutes
USD / run
USD / hour

Time per comparable job

Illustrative example
Baseline work240 minAI-assisted job118 min0240 min
Illustrative example: stacked minutes for one comparable job, not Voyager results.
  • Baseline work
  • AI preparation
  • Generation
  • Human review
  • Correction
Net time saved+122 min
Baseline time4.0 hr
Assisted time2.0 hr

Net time saved = baseline − (preparation + generation + human review + correction) = 240 min − 118 min = +122 min. A negative value means the assisted job takes longer.

Labor cost = minutes ÷ 60 × hourly rate. Net cost savings = baseline labor cost − (assisted labor cost + run cost). Negative values mean estimated net cost increases.

Net cost savings = $480.00 − ($236.00 + $4.20) = $239.80.

Comparative cost per job

Illustrative example
Baseline job$480.00AI-assisted job$240.20$0$480.00
Illustrative example: stacked US dollars per comparable job, not Voyager results. The assisted bar separates labor from direct run cost.
  • AI-assisted labor
  • Per-run tool cost
  • Baseline labor
Illustrative example — accessible values for the time chart and calculator; not Voyager results.
MeasureBaselineAI-assistedDifference
Time240 min118 min+122 min net saved
Labor cost$480.00$240.20$239.80
Run cost—$4.20Included in net savings

Revisions and new assignments

Corrections can reveal a repeated instruction problem. A reviewer might repeatedly find that drafts treat a question as a confirmed claim or move an exploratory date into the delivery schedule. The reviewer can propose a specific instruction change and explain the source of the correction. The proposal stays separate from the approved instructions while the owner tests it against examples the client and agency permit the team to reuse.

The retest should use the same review criteria as the initial test. The account lead checks whether each required detail remains, whether the brief distinguishes evidence from assumption, whether source references lead to the right material, and whether missing information remains visible. A named reviewer approves or rejects the revision. If approved, the owner records a version, date, change summary, approver, and review date. If the change makes an earlier example fail, the team can restore the previous instruction while it investigates.

Client-specific corrections stay within that client's permissions. A repeated correction can inform a general agency procedure only after the owner removes confidential details and confirms that the rule applies beyond the original assignment. The team does not copy a client brief into another client's examples to save time. Permission follows the source material, not the convenience of the next prompt.

The same method can support a project status report or a research summary, with checks suited to each task. A status report needs confirmed dates, owners, and outstanding decisions from the project team. A research summary needs a reviewer to inspect its sources, distinguish reported evidence from interpretation, and flag missing context. The assignment changes; the responsibilities, evidence trail, permissions, and approval record still need to be explicit.

Our teammates continue to speak with clients, understand what they want to accomplish, make the creative decisions, and explain our recommendations. A well-defined assistant assignment gives them a specific opportunity to spend less time reconstructing a conversation and more time applying what they know. The next person joining a Voyager project should be able to find the approved brief, see why the account lead made the important decisions, and prepare a better first question. We use that standard to decide whether an assistant is ready to help.

Controlled improvement

Improve instructions without skipping approval

Corrections inform a proposal. A proposed instruction is retested on permitted examples, then reviewed by a person before a new version is used.

01Approved brief02Reviewer corrections03Proposed instruction04Permitted-example retest05Human approval06Versioned instructions
Text equivalent: approved brief → reviewer corrections → proposed instruction change → permitted-example retest → human approval → versioned instructions. The approval gate precedes any new instruction version.
Approved brief: A human-approved brief sets the task, scope, and intended audience.

Selected loop step: Approved brief. A human-approved brief sets the task, scope, and intended audience.

Talk to a Voyager Guide

Have a brand challenge, campaign, or big idea? Let’s create something remarkable together.