Finding high-value AI use cases

Video lecture

15 chapters · about 9 min · full transcript

Coming soon

Chapter 1 of 15

Finding high-value AI use cases

Problems first, models second
A repeatable discovery method
Scoring and choosing

The narrated lecture is in production

Every chapter is scripted and ready. Browse the chapters and read the full transcript now — the video will appear here when it’s published.

Chapters

Transcript of the narration, chapter by chapter.

0:00 Finding high-value AI use cases

Most AI projects don't fail because the model is weak. They fail because someone picked the wrong problem. A chatbot for a question nobody asks. An impressive demo for a task that happens twice a year. In this lesson you'll learn a repeatable way to find the work where AI creates real value, score it honestly, and choose your first projects with confidence, whether you run a five-person agency or a department in a large company.

0:33 Analogy: choosing a shop location

Here's an analogy. Choosing AI use cases is like choosing where to open a new shop. You don't pick the street with the prettiest buildings. You count footfall, look at what people already buy, and check the rent. Frequency is footfall. Time spent is what people already pay in effort. Feasibility is the rent. The prettiest idea, a flashy chatbot, often sits on an empty street.

1:02 Where AI creates value

Start with problems, not models. The best candidates are work that's valuable, frequent and painful. Then ask whether AI is the right tool. Language and multimodal models are strongest with unstructured information like emails, documents, tickets and calls; with transforming language, such as drafting, summarising, translating and classifying; with judgement at scale where a human can review; with search and synthesis across sources; and with repetitive steps that have too much variation for old rules-based automation. They're weaker when you need guaranteed exactness without checking, when there's little data or context, or when stakes are high and there's no room for review.

1:46 Discovery method

Here's the method. Map workflows, not departments. Walk through a real process step by step, like handling a supplier invoice or answering a sales enquiry, and note volume, time, errors and frustrations. For each step, ask whether it's language-heavy, repetitive, rules-with-exceptions or search-heavy. Estimate value: time saved times volume, faster response, better quality, revenue or reduced risk. Estimate feasibility: data availability, integration difficulty, the accuracy you need and the risk if it's wrong.

2:18 Value–feasibility grid

Now plot candidates on a value and feasibility grid. High value and high feasibility are quick wins: start there, because they build credibility and teach your team. High value but low feasibility are strategic bets: plan carefully and de-risk with pilots. Low value and high feasibility are nice-to-haves, only worth it if nearly free. And low value, low feasibility, you avoid. To find the numbers, ask simple questions. How often does this happen? How long does it take, and who does it? What happens when it's done badly? Would a good draft that a human edits save most of the time?

3:02 Simple example: 'send me last year's return'

A simple example first. A three-person accounting firm notices every client email asking for a copy of last year's return takes about ten minutes: find the file, check identity, attach, reply. It happens around twenty times a week. That's over three hours weekly on a frequent, clear, low-judgement task, where a drafted reply with the right file attached, checked by a person, saves most of the time. Not glamorous, but an ideal first use case.

3:35 Worked example: a three-country agency

Let's see it work. A digital agency serving e-commerce clients in the UK, UAE and Pakistan maps three workflows. Monthly client reports eat hours of analyst time, and AI can draft commentary from exported data for analysts to review: a quick win. New business proposals are bespoke, but AI can draft sections from past proposals: medium value, medium feasibility. Automated media buying could be valuable, but it's high risk and dependent on platform data: a strategic bet for later, with approval gates. They start with reports, measure hours saved and client feedback, and use the result to fund the next project.

4:19 Business example (illustrative)

Let's add illustrative numbers to the agency. Monthly reports take analysts about six hours per client across twenty clients, around a hundred and twenty hours a month. If AI drafts the commentary and analysts review for ninety minutes each, that's about thirty hours, saving roughly ninety analyst hours a month. The proposal assistant might save a third of that. Media buying could be worth more, but only after months of data work. The grid made the order obvious.

4:53 Red flags

Watch for red flags. AI-for-everything initiatives with no named process owner. Use cases chosen because they're impressive rather than frequent. No way to measure success. Tasks where errors are costly and outputs can't be checked. And automating a broken process instead of fixing it first. The best ideas usually come from the people who do the work, so run short sessions where they show you their most repetitive, frustrating tasks with real examples. You'll find better candidates, spot edge cases early, and build the goodwill that adoption depends on.

5:32 Hands-on in the lesson

In the lesson's hands-on section you'll find an interview prompt that turns rough notes into a structured task list, quoting evidence and writing unknown instead of guessing. Then a short Python script scores a use-case backlog from a spreadsheet, combining hours per week with agreed scores for value, data readiness, integration ease, error tolerance and risk, and prints a ranked list with an illustrative monthly value. Agree the weights with leadership, keep the file shared, and re-score every quarter, because model capabilities and costs change fast.

6:09 More mistakes

A few more mistakes worth naming. Asking leadership for use cases instead of the people doing the work. Counting a task's total time without noticing that most of it is waiting, not working. Scoring every idea as high value because nobody wants to disappoint colleagues. And forgetting data readiness: an idea that needs data you don't have is a data project first and an AI project second.

6:38 How you'll know you picked well

How will you know you picked well? Within the first few weeks of a pilot, the people doing the work say it saves them time, unprompted. The volume assumptions in your scoring turn out roughly right. And the measured time saved is in the same range as your estimate. If the estimate was wildly off, revisit your scoring method before picking the next use case.

7:06 Watch me do it: from notes to ranked backlog

Watch me do it. First, I paste twenty minutes of interview notes from our operations lead into the interview prompt. The output is a table: task, trigger, frequency, minutes, who, inputs, output, pain and sensitivity. Two numbers say unknown, so I ask a follow-up question and fill them in. Next, I save three rows into usecases dot csv and score value, data readiness, integration ease, error tolerance and risk from one to five with the team. Then I run the script. It computes hours per week from frequency times minutes, normalises that to a one to five time score, flips risk so low risk scores high, and blends everything with the weights. It prints a ranked list: client report commentary on top, about six hours a week, then invoice chasing, then proposal drafts. Finally, I check the top item against the red flags, name the head of client services as owner, and book a scoping session.

8:14 Recap

To recap: start from valuable, frequent, painful work. AI is strongest with unstructured information, language, reviewable judgement and search. Map workflows, estimate value and feasibility, plot them on the grid, begin with quick wins and stage the strategic bets. Your next step is to map one workflow in your organisation step by step, score three candidate steps, and place them on the grid. In the next lesson, we'll turn your top candidate into a tight scope with success metrics.

8:48 Try this now (45 minutes)

Try this now. Book a twenty-minute conversation with someone whose work you know is repetitive. Ask them to show you, not tell you, their most frequent task, and time it. Then run the interview prompt on your notes and put three candidate steps into the scoring script or sheet. By the end, you'll have your first scored backlog, grounded in real work instead of assumptions.

Start with problems, not models

Many AI initiatives start with "we should use AI for something". They produce demos that impress in a meeting and die in production. Successful teams start from work that is valuable, frequent and painful, then ask whether AI is the right tool.

Where AI tends to create value

AI (particularly language and multimodal models) is strongest where work involves:

Unstructured information: emails, documents, tickets, calls, images.
Language transformation: drafting, summarising, translating, rewriting, classifying.
Judgement at scale with tolerance for review: triage, first drafts, recommendations that humans approve.
Search and synthesis: finding and combining information across sources.
Repetitive steps with variation: processes too irregular for traditional rules-based automation.

It is weaker where work requires guaranteed exactness without verification, has little data or context available, or involves high-stakes decisions with no room for review.

A discovery method

Map workflows, not departments. Walk through a process step by step (for example "handling a supplier invoice" or "responding to a sales enquiry"). Note time spent, volume, error rates and frustrations.
Identify candidate steps. For each step: is it language-heavy, repetitive, rules-with-exceptions, or search-heavy?
Estimate value. Time saved multiplied by volume, faster response, higher quality, revenue impact, risk reduction.
Estimate feasibility. Data availability, integration difficulty, accuracy needed, risk if wrong.
Prioritise on a value-feasibility grid.

The value-feasibility grid

                 Low feasibility          High feasibility
High value   |  Strategic bets:        |  Quick wins:             |
             |  plan carefully,        |  start here              |
             |  de-risk with pilots    |                          |
Low value    |  Avoid                  |  Nice-to-have: only if    |
             |                         |  nearly free             |

Quick wins build credibility and learning. Strategic bets need executive sponsorship and staged investment.

Questions that reveal real value

How many times a week does this happen?
How long does it take, and who does it (and what does their time cost)?
What happens when it's done badly or slowly?
Would a good-enough draft that a human edits save most of the time?
What would the team do with the time saved?

Worked example: a small agency

A digital agency serving e-commerce clients in the UK, UAE and Pakistan maps three workflows:

Monthly client reports: analysts spend many hours collating platform data and writing commentary. AI can draft commentary from exported data; analysts review. High value, high feasibility: a quick win.
New business proposals: each is bespoke; AI can draft sections from a library of past proposals and case studies. Medium value, medium feasibility.
Automated media buying decisions: high potential value, but high risk and dependence on platform APIs and data quality. A strategic bet for later, with human approval gates.

They start with reports, measure hours saved and client feedback, and use the result to fund the proposal assistant.

Red flags in use-case selection

"AI for everything" initiatives with no named process owner.
Use cases chosen for impressiveness, not frequency.
No way to measure success.
Tasks where errors are costly and outputs cannot be checked.
Automating a broken process instead of fixing it first.

Involve the people who do the work

The best use-case ideas usually come from front-line staff, not strategy decks. Run short sessions where people describe their most repetitive, frustrating tasks and show you real examples. You will find better candidates, understand edge cases early, and build the goodwill that adoption later depends on.

Hands-on: a scored use-case backlog in 30 minutes

Step 1: collect candidates with an interview prompt. Paste this into any capable AI assistant (with approved, non-sensitive notes) after a 20-minute conversation with a team, or use it live as a structured interview guide:

You are helping me discover AI use cases in one team's workflow.
From the interview notes below, list each distinct recurring task as a row with:
task | trigger | frequency per week | minutes per occurrence | who does it |
inputs (documents, systems) | output | pain or error described | data sensitivity (public/internal/confidential/restricted)
Rules: only use facts from the notes; write "unknown" rather than guessing; quote the note that supports each number.
<notes>
{{PASTE INTERVIEW NOTES}}
</notes>

Step 2: score and rank. Save the rows as usecases.csv with columns name,freq_per_week,minutes,value_other,data_ready,integration_ease,error_tolerance,risk, where the last five are 1 to 5 scores agreed with the team. Then run:

import csv

WEIGHTS = {"time_value": 0.35, "value_other": 0.15, "data_ready": 0.15,
           "integration_ease": 0.15, "error_tolerance": 0.10, "risk": 0.10}
LOADED_COST_PER_HOUR = 30.0   # illustrative; use your own loaded hourly cost

rows = list(csv.DictReader(open("usecases.csv", encoding="utf-8")))
max_hours = max(float(r["freq_per_week"]) * float(r["minutes"]) / 60 for r in rows) or 1
for r in rows:
    hours = float(r["freq_per_week"]) * float(r["minutes"]) / 60
    r["hours_per_week"] = round(hours, 1)
    r["time_value"] = 1 + 4 * hours / max_hours            # normalise to 1-5
    r["risk_inv"] = 6 - float(r["risk"])                    # low risk scores high
    r["score"] = round(
        WEIGHTS["time_value"] * r["time_value"] + WEIGHTS["value_other"] * float(r["value_other"]) +
        WEIGHTS["data_ready"] * float(r["data_ready"]) + WEIGHTS["integration_ease"] * float(r["integration_ease"]) +
        WEIGHTS["error_tolerance"] * float(r["error_tolerance"]) + WEIGHTS["risk"] * r["risk_inv"], 2)
    r["est_value_per_month"] = round(hours * 4.3 * LOADED_COST_PER_HOUR)
for r in sorted(rows, key=lambda x: -x["score"]):
    print(f'{r["score"]:>5}  {r["name"]:40} {r["hours_per_week"]:>5} h/wk  ~{r["est_value_per_month"]}/month')

The weights are a starting point; agree them with leadership so the ranking reflects strategy. Keep the file in a shared drive and re-score quarterly.

Step 3: sanity-check the top three against the red flags above, and name a process owner for each before anything is built.

Going further

Maintain a use-case backlog with consistent scoring (value, feasibility, risk, strategic fit) and revisit it quarterly. Model capabilities and costs change quickly; a use case that was infeasible six months ago may now be a quick win, and vice versa if costs or regulations shift.

Start from valuable, frequent, painful work, not from the technology.
AI is strongest with unstructured information, language transformation, reviewable judgement at scale and search-and-synthesis.
Map workflows step by step, estimate value and feasibility, and prioritise on a grid.
Begin with quick wins, stage strategic bets, and avoid impressive-but-rare or unmeasurable use cases.

Map one workflow in your organisation step by step. Score three candidate steps on value and feasibility and place them on the grid.

Enrol for free to save your progress

Reading is always free. Enrol to keep your place, take the final assessment and earn a verifiable certificate.

Finding high-value AI use cases