AI discovery deliverable

AI Opportunity Map Example: From Three Workflows to One Pilot

An opportunity map compares possible workflows and recommends the next decision. It helps an owner see why one candidate can move into a bounded test while another needs preparation, a narrower scope, manual handling, or a stop.

The map is a planning output. It does not prove that a pilot will work, produce a return, or be ready for production.

Inspect the example Score one workflow

A BaristaLabs constructed opportunity-map diagram shows three workflow candidates entering a qualitative comparison. One route continues into a bounded pilot area with a retained human-review checkpoint, while two routes remain visible as deferred work. — BaristaLabs constructed example: three candidate workflows enter a qualitative comparison; one becomes a conditional bounded-pilot recommendation with human review retained, while the other work remains visible for preparation, narrowing, manual handling, or deferral. This is a planning model, not client proof or a performance result.

Prepare the input

Start with three real workflow candidates

A useful comparison begins with work the business can describe, not a list of AI features. For each candidate, capture enough current-process detail to expose ownership, consequence, review, and a reversal path. The public AI consulting page asks buyers to bring three candidates; the readiness assessment can help you inspect one of them in more detail.

Accountable owner: Who explains the current work and owns the next decision.
Trigger and current pain: What starts the work, where it stalls, and why that matters.
Systems and sources: Where inputs originate, where work goes, and which records are authoritative.
Data sensitivity: Which data classes appear and which must stay outside the first boundary.
Consequence: What a mistake could do to a customer, member of staff, record, payment, or public claim.
Review point: What a person sees and can edit, reject, escalate, or keep manual.
Reversibility: How proposed work is discarded or a completed action is restored or cleaned up.
Stop condition: The condition that pauses the test before more examples or access are used.

For the engagement boundary behind those fields, compare AI discovery and a pilot. If the team is still choosing between a product, internal build, contractor, or specialist, use the AI implementation path decision matrix after the workflow is clear.

Constructed example

Compare three workflows with qualitative evidence

The fictional business below is considering support-request triage, invoice-exception preparation, and public content drafting. These are ordinary teaching examples, not a client account. The comparison uses no composite score because the available facts support a reasoned decision, not pseudo-precision.

Conditional pilot candidate

Support-request triage

Advance only as a draft-and-route test. The support lead remains responsible for category, priority, and assignment.

Owner: Support lead
Trigger and pain: A new request enters a shared inbox; staff repeatedly read, label, and route routine messages.
Systems and allowed sources: A privacy-safe set of historical request text, the current category guide, and the team roster.
Data boundary: Exclude credentials, payment details, medical details, attachments, and records outside the prepared sample.
Staff or customer consequence: A wrong label can delay a response or send work to the wrong person.
Proposed AI role: Suggest a category, priority, destination, and short rationale. Do not send a reply or update the system of record.
Retained human review: The support lead confirms or changes every suggestion before routing.
Reviewability and reversibility: The source message and rationale stay beside the suggestion; rejected drafts make no external change.
Stop condition: Pause when the source is missing, the category guide conflicts, sensitive data appears, or the reviewer cannot explain the route.
Qualitative view: Useful repeated work; modest integration effort; bounded but meaningful risk; current categories and ownership are clear enough to test.

Prepare before a pilot

Invoice-exception preparation

Document exception rules and resolve source ownership before any AI-assisted preparation touches financial work.

Owner: Finance manager
Trigger and pain: An invoice differs from a purchase record or lacks required information, and staff assemble context for review.
Systems and allowed sources: Invoice records, purchase records, vendor terms, and approval history would be relevant, but their ownership is not yet settled.
Data boundary: Financial and vendor records require a field-level access and retention review before a representative sample is approved.
Staff or customer consequence: Incorrect preparation can influence payment handling, vendor communication, or an accounting record.
Proposed AI role: Undecided until the exception taxonomy and authoritative sources are stable.
Retained human review: A finance reviewer would retain every payment and record decision, but the evidence they need is not yet consistent.
Reviewability and reversibility: A draft packet could be discarded, yet the current process does not reliably show which source justified each exception.
Stop condition: Do not start a pilot until exception types, source authority, sample approval, and reviewer evidence are documented.
Qualitative view: Potentially valuable; higher source and integration effort; higher consequence; current-process clarity is too weak for a bounded test.

Narrow or keep manual

Public content drafting

Choose one low-consequence content lane and establish approved sources and claim review before testing drafts.

Owner: Marketing owner
Trigger and pain: A recurring public update needs a first draft, but requests arrive with uneven source material.
Systems and allowed sources: Approved public product facts, a current brief, and brand guidance could be allowed after an owner marks them current.
Data boundary: Exclude private customer material, unpublished commercial terms, unapproved testimonials, and claims without a named source.
Staff or customer consequence: An unsupported or stale statement can mislead readers and create correction work.
Proposed AI role: Draft from an approved packet only; no autonomous publishing.
Retained human review: The marketing owner verifies every claim, edit, link, and destination before publication.
Reviewability and reversibility: Drafts are easy to discard and published copy can be corrected, but source provenance is not consistently carried into review.
Stop condition: Keep manual when the source packet is incomplete, the claim owner is absent, or rights and approval status are unclear.
Qualitative view: Moderate value and low build effort; public-claim risk remains; narrow process clarity must improve before a pilot.

Ranking reason

Why support-request triage advances conditionally

This candidate has the clearest inputs, accountable owner, visible review point, reversal path, and bounded data and action surface. The AI role can stop at a suggestion. A support lead can compare that suggestion with the source request and current guide before any route changes. Rejected work can disappear without contacting a customer or changing a financial record.

Invoice-exception preparation is deferred until the business documents exception rules, source authority, and approved financial-data access. Public content drafting should narrow to one source-owned lane or remain manual until claim review and provenance are consistent. A different business can reach a different decision because its owners, sources, consequences, and controls will differ.

Selected workflow boundary

Define the first pilot boundary before choosing the tool

The first boundary keeps the test small enough to inspect. Allowed evidence sits beside explicit exclusions, retained authority, stopping, cleanup, and untested conditions.

Allowed sources: A privacy-safe set of historical support requests, the current category guide, and an approved team-routing list.
Excluded data: Credentials, payment details, medical details, attachments, unrestricted mailbox access, and any record outside the approved sample.
Proposed AI role: Prepare a category, priority, routing suggestion, and short source-based rationale.
Retained human decision: The support lead decides the final category, priority, assignee, escalation, and whether the request remains manual.
Destination action: Write nothing automatically. The reviewer copies or confirms an approved route in a contained test destination.
Representative examples: Ordinary requests, ambiguous wording, missing context, conflicting categories, sensitive-data cases, and requests that should stay manual.
Stop condition: Pause on excluded data, missing evidence, conflicting guidance, an unexplained suggestion, or an unavailable reviewer.
Rollback or cleanup: Discard unapproved suggestions, remove the contained test output, reconcile any reviewer-confirmed route, and revoke test access.
Evidence to collect: Source class, suggestion, rationale, reviewer decision, edit or rejection reason, final route, exception class, and cleanup state.
Untested conditions: Live mailbox access, attachments, multilingual requests, automated replies, autonomous routing, peak volume, and production reliability.

Use the AI workflow security review worksheet to inspect fields, access, retention, vendors, and removal. Use the AI workflow controls guide to define review, evidence, escalation, monitoring, and rollback before any workflow receives more permission.

Decision record

Record risk, architecture direction, and deferred work separately

A map becomes harder to audit when observed facts, assumptions, and unresolved questions blend together. Tool and architecture direction should remain conditional on the workflow, source, access, review, maintenance, and evidence requirements.

Constructed opportunity map decision record
Source facts	The three candidates have accountable business owners; support categories and review ownership are documented; the first test can use a privacy-safe prepared sample.
Assumptions to test	The category guide is current enough for reviewers; suggestions can be evaluated without live mailbox access; a draft-only test will reveal useful exception classes.
Unresolved questions	Which request fields may be retained, how long test records should remain, how conflicts in the category guide are resolved, and what reviewer capacity is available.
Conditional architecture direction	Start with a contained evaluation surface that shows source evidence beside each suggestion. Choose a model, vendor, or integration only after data, retention, access, and maintenance fit are reviewed.
Deferred work	Live inbox integration, automated assignment, response drafting, multilingual handling, attachment processing, finance exception rules, and public-content source provenance.

The deliverable

The map ends with a decision

A pilot is one valid output. The map can also narrow the workflow, identify preparation, point to an existing tool, retain manual work, or stop. If a candidate advances, define the evidence its test must leave behind with the AI pilot proof guide.

Pilot

Test one bounded workflow with representative examples and retained review.

Narrow

Reduce the users, sources, actions, or exception classes before testing.

Prepare

Clarify the process, owner, data, examples, or review evidence first.

Choose an existing tool

Use a product when it fits the workflow and control boundary without custom work.

Keep manual

Retain the current process when consequence or ambiguity outweighs the likely value.

Stop

End the idea when the business cannot establish a useful, owned, reversible boundary.

The pricing and engagement guide explains how discovery, a pilot, and optional support should end with explicit scope and ownership. Current case studies are adjacent delivery proof; they are not direct proof of an AI opportunity assessment or bounded AI pilot.

Limits of this example

This is a constructed teaching example. It is not a case study, client quote, assessment result, pilot result, performance claim, price, schedule, savings estimate, return estimate, accuracy statement, reliability statement, production outcome, or source approval. BaristaLabs has no approved public case for an AI opportunity assessment or bounded AI pilot.

Prepare your candidates

Bring one workflow to the score, or three to the discussion

Use broad workflow descriptions. Do not send credentials, customer records, financial details, health information, or other sensitive data through the public form.

Score one workflow Discuss three candidate workflows

BaristaLabs uses strategic AI consulting to help owners rank candidates and define the next decision.