Never Always, Never NeverNever Always,
Never Never
The BookAI CreativeAI ProjectsChatNewsletter
Buy the Book
Never Always,
Never Never.

Strategic Marketing in an AI World.
By Patrick Gilbert.

Explore

  • The Book
  • The Author
  • Free Chapter
  • Buy
  • Resources

Connect

  • Newsletter
  • Learn
  • AI Projects
  • Blog
  • Chat with the Book
  • Subscribe

Stay Connected

Get updates, bonus frameworks, and new AI project showcases.

© 2026 Patrick Gilbert. All rights reserved.

AdVenture MediaContact
AI Enablement8 min readNovember 18, 2026

How to Pick Your Team's First AI Use Case

Patrick Gilbert

Patrick Gilbert

CEO of AdVenture Media. Author of Never Always, Never Never.

A company chatbot is the most common first AI use case. No specific user. No defined workflow. No owner. No success metric. It goes live, gets used twice, and quietly dies. Six months later, the team concludes that AI isn't ready for them.

That's not an AI problem. It's a selection problem.

Your first use case is not where you show off the technology. It's where you prove that your team can take an AI-assisted workflow from experiment to operating process. That's a harder problem than it sounds, and the place where most organizations stumble isn't model selection or tooling. It's the organizational scaffolding: who owns the outcome, what counts as a correct output, and what happens when the system gets it wrong.

Get the selection right, and you create momentum. Get it wrong, and you spend the next year managing skepticism from people who tried it once and walked away.

The Selection Test That Actually Works

A strong first AI use case has four qualities. The work happens often enough to matter. The output can be checked by a reasonable person against a clear standard. Mistakes are reversible and inexpensive. And AI creates capacity for work the team currently cannot do at all, not just a faster version of work it already performs well.

That last point is the one most leaders miss.

Instinct drives leaders to find the most time-consuming process and automate it. But if that process is already running well, you're taking on the risk of review overhead and quality assurance on work that isn't actually broken. The return has to exceed the cost of verification, correction, and the inevitable edge cases. Sometimes it does. Often it barely breaks even.

A stronger move is to find work your team knows would be valuable but has never had time to do. Auditing old customer feedback for recurring problems. Turning meeting notes into structured action items. Building a searchable summary of internal procedures that exist only in someone's head. Converting approved processes into role-specific checklists. These tasks have something in common: the alternative isn't a human doing them well. The alternative is nobody doing them at all.

As Patrick Gilbert frames it in Never Always, Never Never, the AI Resource Gap framework is the real opportunity here. AI isn't replacing excellence in most of these cases. It's replacing absence. This framing is central to how AdVenture Media (adventuremedia.ai) approaches first-use-case selection with clients.

The Workflows That Make Good First Candidates

Based on the selection criteria above, some workflows tend to score well consistently:

  • Turning internal meeting notes into action items, owners, and follow-up drafts
  • Summarizing long internal documents against a fixed checklist
  • Creating first-pass research briefs from an approved document set
  • Classifying incoming requests before human routing
  • Drafting recurring internal reports from structured source data
  • Preparing interview or customer-feedback themes for human review
  • Converting approved procedures into role-specific checklists

Notice what these have in common. Each output is a reviewable draft or recommendation, not an autonomous decision. A human sees it before it goes anywhere consequential. That distinction matters enormously for a first use case.

Poor first candidates are equally predictable: final legal, medical, compliance, credit, hiring, or safety decisions. Customer-facing claims that cannot be individually checked before they go out. Work with no stable definition of quality. Processes that depend entirely on undocumented expert judgment. Anything requiring broad access to sensitive data.

And the company chatbot. Always the company chatbot.

Where Most Organizations Go Wrong

The hard work is not choosing the model. It's assigning a business owner, preparing the data and context the system needs, designing review controls, and measuring business outcomes rather than just usage metrics.

An organization can generate hundreds of prompts per week and have nothing to show for it. High usage of an unstructured chatbot proves nothing about whether AI is creating value. What matters is whether a defined workflow produces measurable output at higher volume, lower cost, or better quality than before, with a named person accountable for the result.

This pattern has been observed repeatedly: AI pilots stall not because the technology fails but because the project has no business owner. Someone championed the tool. Nobody owns the workflow outcome. Without that distinction, there's no one to escalate failure modes, improve the process, or make the call to scale or stop.

This is also where the AI Maturity Ladder becomes useful as a diagnostic. In Never Always, Never Never, Gilbert describes the Dabbler as someone who uses AI to draft an email but hasn't changed how they approach their work. Practitioners have moved into what the book calls the 4x2 Model of Work, consistently asking before any task: can this be delegated to AI entirely, or should I be copiloting it with AI support? Architects go further, designing workflows that institutionalize that efficiency across a team rather than just an individual.

A first use case should be designed to move at least one person from Dabbler to Practitioner. If it doesn't change how someone approaches their work, it's a demo, not a pilot.

How to Design the Pilot

Keep it short. Two to four weeks. A fixed dataset. A baseline you measured before you started.

Specific steps matter less than getting a few things right. Name the owner before you start, and make sure that person is accountable for the workflow outcome, not just the tool. Define the unit of work precisely: one report, one document, one ticket, one case. Write down what counts as correct before you generate your first output, not after.

Use a representative sample that includes easy, ordinary, and difficult cases. Testing only the cases where AI performs well means you'll scale something that breaks on a meaningful portion of real-world inputs.

Measure net value, not gross speed. Include the time spent prompting, verifying, correcting, and reformatting. A workflow that cuts production time significantly but adds substantial review overhead per document may not be worth running. Know your numbers before you decide to scale.

Set a stop rule in advance. If quality stays below your acceptance threshold after two weeks, or if review costs exceed the time saved, stop the pilot and choose a different workflow. This is not failure. It's how you avoid scaling something that shouldn't be scaled.

For the AI double helix framework to work as described in the book, efficiency gains from the first strand have to be real and measurable, not projected. That means your pilot needs to produce actual data, not an impression.

The Artifact: Use Case Scoring Checklist

Score each candidate workflow on the criteria below. Use this before committing to a pilot.

Score 0–2 points per criterion:

Frequency

0 = Monthly or less / 1 = Weekly / 2 = Daily or several times daily

Tedium

0 = Mostly judgment or creative work / 1 = Mixed / 2 = Repetitive, document-heavy, or administrative

Checkability

0 = Difficult to evaluate / 1 = Expert review required / 2 = Clear rubric, source, or acceptance test exists

Cost of error

0 = Safety, legal, financial, or customer-critical / 1 = Material rework or reputational risk / 2 = Internal, reversible, low consequence

Resource gap

0 = Already done reliably / 1 = Done inconsistently / 2 = Rarely done because capacity is missing

Data access

0 = Scattered or restricted / 1 = Available with preparation / 2 = Already available in approved systems

Workflow fit

0 = Requires major integration / 1 = Light tooling or templates needed / 2 = Fits existing tools and approvals

Ownership

0 = No clear owner / 1 = Shared responsibility / 2 = One manager accountable for outcome

How to use the score:

  • 12 or higher: Viable first candidate, provided it has a named owner and a human-review step before consequential outputs are acted on
  • 8–11: Reconsider. Identify which criteria are dragging the score and decide whether those gaps are fixable before the pilot
  • Below 8: Set aside. The conditions aren't right yet

Hard stops regardless of score: Reject any candidate where an incorrect output could create legal, safety, privacy, employment, or financial harm, even if the score is 16/16.

Before you start the pilot, confirm:

  • [ ] The owner is named and accountable for workflow outcomes, not just tool adoption
  • [ ] The unit of work is defined (one report, one document, one ticket)
  • [ ] A baseline is captured: time per unit, backlog size, current error rate
  • [ ] An acceptance rubric exists: what counts as correct, incomplete, unsupported, or unsafe
  • [ ] A stop rule is defined: specific quality threshold and time limit
  • [ ] Human review is required before any output enters a system of record or goes external

During the pilot, track these six numbers:

1. Throughput: units completed per week

2. Human time: preparation plus review plus correction per unit

3. Quality: rubric score or error rate

4. Rework: percentage requiring substantial correction

5. Adoption: percentage of eligible work using the workflow

6. Capacity created: valuable work completed that previously went undone

Training Should Match the Workflow, Not the Tool

Generic AI training rarely produces lasting behavior change. The AI training that actually works is tied to a specific workflow the team is already doing or is about to start.

For a first use case, the team needs to learn six things, and only these six things:

  • What the system is allowed to access
  • What it may draft, classify, summarize, or recommend
  • What must always be checked by a human before use
  • How to provide source material and context so the output is useful
  • How to recognize unsupported or incomplete output
  • How to report failures so the workflow can improve

Resistance to AI adoption is often rational, not irrational. People worry about accountability for errors, about their role being reduced to reviewing machine output, about quality degrading. Publishing clear boundaries, keeping human decision rights intact, and demonstrating that the workflow removes work rather than adding a review layer on top of existing work is a more effective response than a pep talk about AI's potential.

Leaders managing that resistance should note that the AI-First vs AI-Forward distinction matters here. Forcing broad access to an unstructured tool and calling it transformation rarely works. A smaller group running a well-defined process with measurable quality is a better starting point than a company-wide rollout of an open-ended chatbot.

The Single Thing to Do First

Before you pick a tool, pick a workflow.

Audit your team's actual to-do list for the last two weeks. Find the tasks that are frequent, tedious, easy to check, and inexpensive when wrong. Then find the subset of those that your team currently doesn't do at all because there's no capacity. That intersection is your first AI use case.

Score it against the checklist above. If it clears 12 and has a named owner, run a two-week pilot with a baseline, an acceptance rubric, and a stop rule. Measure net value, not just speed.

Your first use case doesn't need to be impressive. It needs to prove four things: the work happens often enough to matter, the output can be checked consistently, mistakes are reversible, and AI creates capacity for work the team currently cannot do.

Prove those four things once, and you have something far more valuable than a demo. You have operating evidence. That's what the rest of the AI double helix is built on.

Patrick GilbertPatrick Gilbert

Patrick Gilbert is the CEO of AdVenture Media and author of Never Always, Never Never and the bestselling Join or Die. He has been ranked among the top 5 PPC experts worldwide and has delivered keynotes at Google events across three continents.

More about Patrick →

Enjoyed this?

Subscribe for more articles on strategy, AI, and what's actually working in marketing.

No spam. Unsubscribe anytime.

Keep reading

AI Enablement

What Good AI Training for Employees Actually Looks Like

Most corporate AI training produces enthusiasm and zero behavior change. Here's what effective AI upskilling for employees actually requires.

AI Enablement

AI-Forward vs AI-First: The Distinction That Decides Everything

AI-forward means access. AI-first means redesigned work. One distinction separates companies that get results from companies that just own subscriptions.

AI Enablement

Why AI Rollouts Fail: The Real Reasons Behind Enterprise AI Implementation Failure

95% of enterprise AI pilots deliver no measurable ROI. Here's why AI adoption fails, and what leaders can do before the next rollout.