Thursday morning, 8:47 AM. Maria’s coffee was already cold.
She had been staring at the spreadsheet for three minutes, which does not sound long until you consider that Maria is not a person who stares. She is a marketing manager at a mid-sized company in Bucharest, and if you have been following this series, you already know her: sharp, resourceful, self-taught on AI because nobody else was going to teach her.
In the previous article, she ran her First-Contact Protocol: a set of structured conversations with ChatGPT, Claude, and Copilot that gave her a clear map of what each tool could do, what each one could not, and which ones she could trust with company data.
Now the quarterly customer feedback report was due by Monday. Three months of survey responses, support tickets, Net Promoter Score comments, and freeform complaints, all waiting to be read, sorted, and distilled into something her leadership team could act on. Last quarter it had taken her most of a day. This quarter, she planned to hand the heavy lifting to Copilot.
She opened the feedback file. She opened Copilot. Her fingers hovered over the keyboard.
And something nagged at her.
She remembered a study from her reading the week before. The Harvard/BCG experiment we explored in The Gap, the opening article of this series: 758 elite consultants, given AI assistance on a task outside its reliable capabilities, performing 19 percentage points worse than colleagues who worked alone. Not on some exotic problem. On a business question that looked, on the surface, perfectly suited to AI. The consultants trusted the tool. The tool let them down. And because they had trusted it, they never caught the failure.
Maria put her hands in her lap.
Here is the thing about that moment. Most AI training would have told her to start typing. Every course, every tutorial, every vendor pitch begins after the decision to use AI has already been made. They teach you how to prompt. How to upload files. How to iterate. Useful skills, all of them. But they skip the step that determines whether any of it will actually help.
That skipped step turns out to explain a puzzling number. When Cedefop, the European Centre for the Development of Vocational Training, surveyed 5,342 workers across eleven EU countries, they found that 54% of people who use AI at work say the technology has not made them better at their jobs. More than half. Not people who refused to try it. People who tried it, used it regularly, and still saw no improvement.
The tools deliver extraordinary results when matched to the right task. So why is more than half the workforce shrugging? Because nobody taught them what Maria was about to do: decide whether to use the tool at all.
The Skill Nobody Teaches
That pause was, itself, a skill. In a moment saturated with AI enthusiasm, where every conference keynote and every vendor pitch insists the tool belongs on everything, Maria had done something almost nobody does. She had hesitated.
The Harvard/BCG researchers showed us why it matters. When a task sits squarely within AI’s capabilities, the gains are remarkable: consultants completed 12% more tasks, finished 25% faster, and produced work rated 40% higher in quality. The tools deliver. But when the task is poorly matched, the same tools actively degrade professional judgment. The researchers called this the “jagged technological frontier”: an unpredictable landscape where AI excels brilliantly at some surprisingly difficult tasks and fails completely at some surprisingly simple ones.
The 54% who see no improvement are not lacking tools. They are lacking a map of the frontier. They are using AI for tasks where it helps, and also for tasks where it hurts, and they have no reliable way to tell the difference.
SCOPE is the map. It is the first pillar of the SPARKS framework, and it asks a single question before anything else: should you use AI for this task? Not an acronym, just a word. As in: what is the scope of this tool’s usefulness, and does your task fall inside it?
Four Questions, One Decision
Without knowing it had a formal name, Maria was about to run through SCOPE’s Go/No-Go Checklist. Four questions. Each one filters a different dimension of risk and reward. Together, they tell you whether a task belongs in AI’s hands, in yours, or somewhere in between.
The questions themselves are not complicated. But the thinking they require is genuine, and that is precisely what makes them valuable. They force you to engage your judgment before the AI engages its pattern matching.
This is worth pausing on. SCOPE is the one step in the entire framework that AI cannot perform for you. The tool has no idea what is valuable to your team, what is risky for your company, or what your professional reputation can absorb. That context lives in your head. The AI can help you think it through, as we will see later, but the decision is yours.
Let us walk through them with Maria.
Is There a Capability Match?
Maria looked at her customer feedback report and asked herself: based on what I learned during my First-Contact Protocol, is Copilot actually good at this kind of work?
She thought back to her testing sessions. Copilot had handled text summarisation well. It had compared two documents and highlighted differences accurately. Pattern recognition in text sat solidly within its abilities. Customer feedback analysis is, at its core, reading and pattern recognition at scale: hundreds of short responses, recurring themes, sentiment shifts, illustrative examples. This looked like a match.
The key phrase is “looked like.” Maria did not assume the match based on a vague sense that “AI is good at text stuff.” She anchored her assessment to specific evidence from her own testing. She had seen this particular AI do similar work, and do it well. That distinction matters enormously. Without your First-Contact Protocol results to anchor against, every capability assessment is a guess. With them, it becomes an informed judgment.
Will This Genuinely Save Time?
This is the question that separates realistic expectations from wishful thinking. It is honest maths.
She estimated the task both ways. Manual analysis: roughly three hours. She knew this from experience. Last quarter it had taken her exactly that long, and the quarter before that, slightly more.
AI-assisted analysis: perhaps thirty minutes to craft the right prompts and upload the data, plus another sixty to ninety minutes to carefully review the output against the raw feedback. Call it two hours at the outside.
Net saving: about one hour. Maybe a bit more.
But Maria had noticed something important during her early experiments with AI. AI output is not finished product. It is a first draft that needs verification. And verification takes time. There is a hidden cost embedded in every AI-assisted task, one that the enthusiasts rarely mention and the sceptics rarely quantify: the fix-it tax. The time you spend reading through the AI’s work, checking its claims against reality, correcting its mistakes, and reshaping its output until you would actually put your name on it. The tax varies by tool, by task, and by the moment in time. Today’s models require substantial review. Tomorrow’s may require less. But the habit of budgeting for it, rather than assuming zero, is what keeps the estimate honest.
The mental maths is simple enough to carry in your head: time saved, minus time spent prompting and fixing, equals net value. If the number is clearly positive, proceed. If not, reconsider.
For a low-stakes internal memo, the fix-it tax is negligible. A quick skim, a few small edits, done. For a customer feedback report that will inform leadership decisions, the tax climbs. Maria would need to verify that the themes the AI identified actually matched the raw data. She would need to check that quoted examples were real, not fabricated or conflated. She would need to ensure that sentiment comparisons to the previous quarter held up against the numbers.
One hour saved, after accounting for the fix-it tax. On a single task, that sounds modest. But this is a quarterly report. Four times a year, every year. And customer feedback analysis is just one of Maria’s recurring tasks. Apply the same honest maths across a dozen such tasks in your workflow, and modest savings compound into something substantial. One hour here, thirty minutes there, an entire afternoon recovered somewhere else.
There was also a dimension no spreadsheet could capture. The AI would read every single response without fatigue, without the unconscious tendency to skim the twentieth complaint about the same issue. It might surface patterns Maria would have missed, because she is human, and humans get tired at response number 147. That kind of discovery does not show up in a simple time comparison. But it might be the most valuable thing the analysis produces.
Value verdict? Yes. Not because any single instance is transformative, but because the cumulative effect across tasks and quarters is real, and because the potential for genuine insight makes the exercise worth more than the hours alone suggest.
If you are curious what this kind of honest calculation looks like for your own tasks, you might open any AI tool and start a conversation that goes something like this:
I want to estimate the true time cost of using AI for a specific task, including the review and correction time that is easy to underestimate.
Here is the task: [describe the task] Here is how long it takes me to do it manually: [your estimate] Here is who will use the output and what is at stake: [brief context]
Interview me one question at a time to understand the task, then help me build an honest estimate. Consider these hidden costs: How long will it take to craft prompts that give the AI enough context? How thoroughly will I need to verify the output, and against what? What is the cost if a mistake survives my review? Are there parts of the output I will likely need to redo from scratch? How good does the output need to be before I would put my name on it?
Give me a realistic net time saving, not an optimistic one. If the fix-it tax erases the benefit, tell me directly.
The conversation that follows will force you to quantify things you would otherwise leave vague. That is the point.
What Happens If the AI Gets This Wrong?
Here is where Maria’s thinking sharpened. And here is where SCOPE’s centre of gravity lives. The first two questions tell you whether AI is useful. The risk check tells you whether useful is safe enough.
The customer feedback report was for her leadership team. Internal audience. No clients would see the raw analysis. If Copilot misidentified a theme or missed a pattern, Maria would catch it during review, and even if something slipped through, the consequences would be contained. Her team lead might ask a clarifying question. A recommendation might need refining. Uncomfortable, perhaps. Career-ending? No.
She rated the risk as medium. Not negligible, because leadership decisions could be influenced by this analysis, but manageable because she was the one presenting it and could correct errors in real time.
Compare this to a scenario she had faced the day before. A long-standing client had sent a complaint about a product defect. The email was detailed, technical, and unmistakably emotional. The client felt let down. Maria’s first instinct was to ask Copilot to help.
She paused and ran the questions.
Capability match? Copilot could certainly produce something that sounded empathetic and professional. But it would be performing empathy, not feeling it. It would not understand the eighteen months Maria had spent building a relationship with this client. It would not know that this person values directness over corporate polish, that their last conversation had ended with a joke about Romanian weather, that the trust between them was personal, specific, and earned over dozens of interactions.
Value? Here is where it got tricky. Even as a thinking partner (what the framework calls Augmentation, which we will get to in a moment), the AI would generate language. And language about a product defect carries legal weight. Every sentence in that response would need forensic review: is this an admission of fault? Could this phrase be read as a warranty commitment? Maria’s legal team would need to see the final version regardless. If she started with AI-generated text, the review overhead alone would exceed the cost of simply writing from scratch.
Risk? High. Not “embarrassing email” high. “Lose a major client and create legal exposure” high. One carelessly worded sentence about the defect, one that survived Maria’s review but not her lawyer’s, and the consequences would cascade.
An experienced AI user might reasonably ask: why not use AI as a brainstorming partner, just to think through the structure and key points, then write the actual response yourself? It is a fair question, and for many difficult communications, that approach works well. But Maria considered the specific circumstances and decided against it. The risk was not just in the final words. It was in the framing. If the AI suggested a conciliatory opening, or a particular way of acknowledging the defect, that framing might subtly influence her own writing in ways she would not fully track. When the stakes are this high, she wanted her words to come from her understanding of this client, this defect, this relationship. Unmediated.
Decision: No-Go. She wrote that response herself. It took forty-five minutes. It was exactly right. Not because AI could not have produced something competent, but because competent was not good enough. This required something no model can provide: judgment rooted in a specific human relationship.
This is the power of the risk check. It does not tell you AI is dangerous. It tells you when the consequences of a mistake outweigh the convenience of automation. For many tasks, the consequences are low and AI is a clear win. For some, the stakes are too high to delegate the thinking, no matter how fluent the machine’s output. And here is a principle worth remembering: of the four SCOPE questions, risk is the one with veto power. A task can score perfectly on the first three: strong capability match, clear time savings, obvious partnership type. A single “high” on the risk check overrides them all.
Remember Samsung’s engineers, who fed source code into ChatGPT and could never get it back. The risk assessment they should have run would have taken thirty seconds. The cleanup consumed the organisation for months.
What Kind of Help Do I Need?
Maria had decided: the customer feedback analysis was a Go. The capability was there, the value was positive, the risk was manageable. But one question remained. How should Copilot help?
This is the question of partnership type, and it matters more than most people realise. Not all AI assistance looks the same, and choosing the wrong mode for the task at hand is a reliable way to get mediocre results.
The difference between the two main modes is less about technology than about the role you want to play.
The simplest is Automation: the AI does the work, you review the finished product. Delegation with a quality check. You provide clear instructions, the AI executes, you verify before using the output. This works beautifully for high-volume, low-complexity tasks where the review overhead stays low. Reformatting meeting notes into action items, converting bullet points into prose, summarising straightforward documents where the facts are easily verified. The AI operates relatively independently, and your role is quality control.
Then there is Augmentation, which is fundamentally different. Here, you and the AI think together. It is a conversation, not a delegation. You bring your expertise and context; the AI brings its ability to process information at scale and surface perspectives you might not have considered. This mode works best for complex analysis, strategic thinking, and creative work where your judgment matters as much as the AI’s processing power. You are actively steering the work throughout, not just checking it at the end.
A third mode, Agency, exists for advanced use cases where the AI operates more independently over longer workflows. We will explore it later in the series. For the decisions most professionals face daily, the choice between Automation and Augmentation is the one that matters most.
The partnership type is not fixed. A task that demands Augmentation today, because the tool’s judgment cannot yet be trusted without oversight, may shift toward Automation as models improve. Your quarterly reassessment should revisit not just whether to use AI, but how.
For her customer feedback analysis, Maria chose Augmentation. She did not want to simply hand Copilot the data and say “analyse this,” then review whatever came back. She wanted to work with it. Start by asking it to identify themes. Then challenge those themes against her own reading of the data. Ask it to compare findings to the previous quarter. Probe areas where its analysis diverged from her intuition. Use it as a thinking partner, not a thinking replacement.
For reformatting last Wednesday’s meeting notes into action items? That was Automation. Clear input, clear expected output, low stakes. Let the AI do it, skim the result, send it to the team.
The distinction matters because it shapes everything that follows: how you prompt, how much review time you budget, what kind of output to expect. Treating an Augmentation task like Automation (just throw the data in and see what comes out) almost always produces disappointing results. Treating an Automation task like Augmentation (extensive back-and-forth conversation about reformatting meeting notes) wastes time you cannot afford.
And so Maria began her augmented analysis of the customer feedback. What happened over the next hour taught her something no training course could have.
Copilot identified seven recurring themes in the quarterly data. Six matched Maria’s own reading perfectly. The seventh made her sit up straighter: a cluster of comments from enterprise clients mentioning response times during onboarding. Not complaints exactly. More a pattern of mild friction that individually seemed unremarkable but collectively suggested a systemic issue. Maria had read every one of those comments during previous quarters. She had never connected them.
She checked the raw data. The pattern was real. Fourteen separate mentions across three months, from clients in different segments, using different words to describe the same experience. The kind of signal that hides in plain sight when you are reading your 150th survey response and your attention has started to drift. Her leadership team would want to know about this. It was, genuinely, a discovery.
And then she scrolled down one more paragraph.
“Customer satisfaction with pricing has declined 12% quarter over quarter,” the analysis stated, with the calm authority that makes AI output so seductive and so dangerous.
Maria knew the pricing data. She had built those satisfaction tracking sheets herself. She pulled up the raw numbers. The actual decline was 4%, well within normal fluctuation. The AI had compared the wrong data points, producing a figure that sounded precise, looked professional, and was quietly, consequentially wrong.
She stared at the screen. A 12% decline in pricing satisfaction would have triggered a review. Her VP would have convened a meeting. The product team would have been pulled in. Hours of executive time, weeks of analysis, all chasing a problem that did not exist. Because a machine had compared two numbers that should never have been compared, and presented the result with perfect confidence.
Had she run this as Automation, had she simply handed the data over and pasted whatever came back into her presentation, that invented number would have reached her leadership team on Monday morning. Nobody in that room would have had reason to doubt it. The figure was specific. The language was measured. It would have looked like a fact.
SCOPE had told her to use Augmentation for exactly this reason. She was in the conversation. She caught the fabrication. She kept the genuine insight. And because she had budgeted for the fix-it tax rather than wishing it away, the review that caught the error was built into the process, not bolted on as an afterthought.
That is the difference between AI making you better at your job and AI making you confidently wrong.
The Grey Zone
Not every task sorts neatly into Go or No-Go. Maria discovered this the following week.
Her manager asked her to put together a competitive analysis: three competitors, publicly available information, market positioning, pricing where available. She ran the four questions. Capability and value both looked strong. ChatGPT handled research synthesis well, and the time savings over a manual afternoon of trawling websites would be significant. But the risk question gave her pause. Competitive intelligence has a particular vulnerability: if the AI fabricates a competitor’s pricing or invents a product feature, and Maria does not catch it, her team could make strategic decisions based on fiction.
Before settling on a verdict, she tried something from her First-Contact Protocol. She described the task to ChatGPT and asked it to challenge her own assessment. Her message looked something like this:
I have a task I am considering using AI for, but I am genuinely unsure whether it is a good idea. I want you to act as a critical challenger to stress-test my thinking.
Here is the task: competitive analysis of three competitors using publicly available information, covering market positioning and pricing where available.
Here is my initial assessment: I think the capability match is strong because ChatGPT handles research synthesis well. I estimate the time saving at roughly two hours after accounting for review. The risk level feels medium because competitive intelligence is sensitive to fabrication.
Your job is to challenge me. Where am I being overconfident? What risks am I underweighting? What hidden costs might I be missing? Are there failure modes I have not considered? Could I realistically verify the output, and how would I do it? Ask me follow-up questions one at a time before giving your assessment.
The chatbot raised two points she had underweighted: outdated pricing could be presented as current with no disclaimer, and when information is scarce, AI tends to fill gaps with plausible invention rather than admitting ignorance. The decision remained hers, but the conversation had sharpened it.
Not a clean Go. Not a No-Go. It was what the framework calls a “Go with guardrails.” Proceed, but with conditions: verify every factual claim against a primary source, flag anything the AI could not cite, and mark the analysis as AI-assisted so her team would treat specific claims with appropriate scrutiny.
This is the grey zone, and you will encounter it often. SCOPE does not pretend every decision is binary. The four questions give you a framework for identifying where the risks live and putting appropriate safeguards in place. Sometimes the answer is “yes, but carefully.” That is a perfectly valid outcome.
Your Map of the Frontier
In the span of a few days, without any formal training, without a two-day workshop or a certification programme, Maria developed the skill that separates the 54% who see no improvement from those who do. She learned to look at a task, assess it against her knowledge of the tool, and make a deliberate decision about whether and how to involve AI. She learned judgment.
A pattern had emerged across her decisions. Each came down to the same four questions, and the answers sorted themselves into clear outcomes. When capability, value, and risk all aligned: Go. When the match was partial or the stakes moderate: proceed, but with specific guardrails. And when the evidence was thin or the consequences too severe: not yet. Not never. Not yet. She found herself reaching for these categories instinctively, the way a doctor reaches for a diagnostic framework. The framework does not replace clinical judgment; it ensures no dimension of the problem goes unexamined.
This is what navigating the jagged frontier looks like in practice. Not a set of rules carved in stone. Not a blanket policy that says “use AI for everything” or “use AI for nothing.” But a practiced habit of pausing before prompting, one that grows faster and more instinctive with each use.
Researchers are beginning to confirm what Maria was learning through experience. OpenAI’s GDPval benchmark tested frontier models against seasoned professionals across 1,320 tasks in 44 occupations. The results were striking: top models matched or exceeded expert quality in roughly half those tasks. In the other half, they produced work ranging from fixable drafts to outright failures. More telling still: models that scored roughly 14% in mid-2024 were scoring above 40% barely fifteen months later. The frontier moves. A task that was a clear No-Go in January might be a confident Go by September. (The full picture of GDPval and what it means for professional work is worth exploring.)
Maria thought about her No-Go decisions. The client complaint would remain one, regardless of how the tools improved. That was rooted in a human relationship, not a capability gap. No model update would change it. But other tasks she had dismissed because the tools were not yet reliable enough? Those deserved revisiting. She set a calendar reminder, three months out: “Reassess SCOPE for data analysis tasks.” The landscape shifts quickly enough that capability-based assumptions have a shelf life.
A reasonable objection presents itself here. Nobody is going to run four formal questions before asking Copilot to tidy up meeting notes. And that is exactly right. You should not. For routine, low-stakes tasks where you have already established that the AI performs well, SCOPE collapses into a one-second gut check. The questions are still there, operating below the surface, the way an experienced driver checks mirrors without consciously thinking about it.
You already know the capability is there, the value is obvious, the risk is negligible, the review is a quick skim. The formal checklist is for the tasks where the answer is not obvious: where the stakes are higher, the match is uncertain, or you are working with a tool you have not tested thoroughly. Those are the moments that determine whether AI makes you better at your job or quietly makes you worse.
The 758 consultants in the Harvard/BCG study had no version of SCOPE. They accepted AI’s help on a task it could not handle because nobody had taught them to question whether the help was appropriate. That unexamined trust is what made AI dangerous. Not the technology itself. The absence of judgment around it.
Why This Matters Beyond Productivity
Every SCOPE decision Maria makes is also compliance evidence. The EU AI Act, which we explored in the opening article, requires not just that employees can use AI but that they understand its risks and limitations. When Maria decides against AI for the client complaint, she is demonstrating exactly the judgment the regulation envisions. When she catches the fabricated pricing statistic because she chose Augmentation over Automation, she is proving she can calibrate oversight to risk. BCG’s 2025 survey of 10,635 employees found that only one third say they understand how AI tools actually function. Organisations that train employees only in how to prompt are building half a bridge. SCOPE is the other half.
Your Turn
You do not need to wait for your organisation to build a framework. You already have one. Four questions and your own professional judgment.
There is something worth trying this week, if you are curious. Think of three tasks on your calendar. Real ones, not hypothetical. Ideally a mix: one you suspect is a strong candidate for AI, one you suspect is not, and one you are genuinely unsure about.
For each, sit with the four questions for a moment. You know them by now: capability, value, risk, partnership type. It helps to write the answers down, even briefly. A note on your phone is enough. There is a difference between a vague instinct (“I guess I could use AI for this”) and a decision you can defend and learn from. Making the reasoning visible is what creates that difference.
And if one of the questions genuinely stumps you, remember what Maria discovered in the grey zone: you can use AI itself as a thinking partner for the SCOPE decision, the way she did with her competitive analysis. That conversation might begin something like this:
I need to decide whether to use AI for a specific work task. I want you to act as my decision partner by interviewing me one question at a time.
Start by asking me to describe the task in detail. Then walk me through these four assessment questions, one at a time. First: based on what I know about the AI tool I plan to use, is it actually good at this kind of work? Second: will it genuinely save me time once I account for the time I will spend prompting, reviewing, and fixing the output? Third: what is the real consequence if the AI gets something wrong, who is affected, and how badly? Fourth: if I proceed, should I delegate the task to AI and review the result, or work alongside AI as a thinking partner throughout?
After we work through all four, give me your honest assessment: Go, Go with Guardrails, or Not Yet. If guardrails are needed, tell me exactly what I should verify against primary sources, what I should flag to others, and what review process I should follow. Challenge me where my reasoning seems weak or where I may be overestimating the tool or underestimating the risk.
You are not delegating the judgment. You are using Augmentation to sharpen it.
The results tend to surprise people. Tasks that seemed perfect for AI sometimes reveal review costs that erase the savings. Tasks dismissed as “too important for AI” sometimes turn out to be strong candidates for Augmentation, where the AI’s ability to process at scale genuinely complements your expertise. You might find, for instance, that a task you assumed was straightforward, drafting a performance review for a direct report, fails the risk check not because AI would get the facts wrong but because the person would feel the difference. Some tasks carry emotional weight that no capability match can offset. And the grey-zone tasks, the ones that could go either way, tend to teach you the most about your own judgment and about the tools themselves.
Maria noticed something about her own No-Go decisions that is worth carrying with you. Some were limited by the tool’s current capability. Others were limited by the nature of the task itself. The first kind deserves an expiry date: write “revisit in three months” next to them, because the tools improve faster than intuition adjusts. The second kind, tasks that require human accountability, a specific relationship, or judgment that simply cannot be delegated, may hold indefinitely. Know the difference. Review the first kind quarterly. Respect the second.
If you want a quick reference for all four questions and how to read the results, there is a decision grid at the end of this article.
What Comes Next
Maria now has two things she did not have a week ago. A map of her tools, from the First-Contact Protocol. And a compass for deciding when to use them, from SCOPE. She caught a fabricated statistic before it reached her leadership team. She found a genuine insight hiding in hundreds of customer responses. She wrote a difficult client email herself, in her own words, because she recognised that some tasks demand a human hand. And she can explain exactly why she made each of those decisions.
That is not a casual user. That is a professional.
But here is where it gets interesting. Think back to Maria’s feedback analysis. SCOPE told her to proceed, and it told her how: as a conversation, not a delegation. The decision was sound. But the quality of what followed depended on something SCOPE does not cover. It depended on how she talked to the machine.
Which raises a question Maria herself was beginning to ask, sitting in front of a blank prompt field on the following Monday, wondering how to phrase a request she could feel but not yet articulate. How do you talk to an AI in a way that produces results worth having? How do you give it the right role, the right context, the right constraints, so that the first answer is useful rather than the fifth?
The next article covers PROMPT: the art and discipline of clear instruction. Because a good decision deserves a good conversation.
Quick Reference: The SCOPE Decision Grid
By the end of the week, Maria had distilled her experience into a simple grid. She kept it open on her phone. It was not poetry. It was a decision tool.
Question 1: Capability Match Go — Tool tested on similar work, performed well Go with Guardrails — Partial match, or gaps identified during First-Contact Protocol No / Not Yet — No evidence the tool handles this type of work
Question 2: Value Check (the Fix-it Tax) Go — Clear net time saving after review costs Go with Guardrails — Marginal time saving, but insight or scale benefits justify it No / Not Yet — Fix-it tax erases or exceeds the saving
Question 3: Risk Check Go — Low stakes, errors easily corrected Go with Guardrails — Moderate stakes, specific safeguards needed No / Not Yet — High stakes, errors could cause serious harm
Question 4: Partnership Type Go — Either mode, matched to task complexity Go with Guardrails — Augmentation preferred: stay in the loop throughout No / Not Yet — Task requires unmediated human judgment
All four questions must pass for a clean Go. Risk is not one of four equal factors: it is the one with veto power. A single “No” on the risk check overrides the other three. Capability-limited “Not Yet” decisions have an expiry date: revisit quarterly as tools improve. Principle-limited “No” decisions, those rooted in human accountability or relationships rather than tool limitations, may hold indefinitely.
This is Article 3 of 10 in “GenAI School: AI Literacy for Everybody,” a complete, free course on working effectively with AI.
Missed the beginning? Start with The Gap.
Follow for the next article, or share this with someone who needs it.
Next: PROMPT