Ask AI for a first draft of almost anything and you’ll have it in seconds. It will be well formatted and confident, and it may even be accurate to the facts. Whether it’s ready to send to a client or key stakeholder is a different question, and answering it is where most of the work now happens.
A year ago, most of the conversation about AI at work was about prompting: how to ask for the right thing. Today, good prompts still matter, but the people getting the most from AI do more than send a quick ask. They treat editing as more than a final pass. It starts with how they set up the work and ends with deciding what’s ready to send.
What Can Go Wrong With Accurate Output
In a recent internal session on human-led, AI-enabled consulting, our team reviewed a real AI-generated client deliverable: a tool and data inventory synthesized from discovery interviews. Every quote came from a transcript. Nothing was invented. By any accuracy check, it passed, but it still wasn’t close to client ready. The team caught these problems in internal review, before anything went to the client.
The problems were judgment problems. Editorial language had crept into what should have been a factual column. Direct quotes exposed disagreement between stakeholders, which was true but pointed at individuals instead of the alignment gap the client needed to solve. One person’s frustration was presented as the view of the whole organization. The cells were so dense that no one could scan them, and nothing told the reader what to do with the information.
None of those issues would show up in a fact check. With a human draft, mistakes tend to look like mistakes. With AI, the draft reads smoothly and the facts hold up, so the problems hide in what was emphasized and who it was written for.
Why Polish Makes It Harder
AI output looks finished, and a good template makes it look even better. That polish is easy to mistake for good thinking.
Readers have also gotten better at spotting AI. Em dashes became the best-known tell—sadly—and new ones keep appearing: the word “quietly” in every other paragraph, stacks of adjectives, three-item lists where only one item matters, and closing lines that restate the point. Once a reader notices one, they read everything else with less trust. In client work or when enrolling key partners, losing that trust costs far more than the time AI saved on the draft.
Editing Starts Before the First Prompt
The most effective editing happens before the AI writes anything. Our team uses a framework for this: Think, Direct, Build, Verify. The first three steps shape the draft before it exists. The last decides whether it’s ready to send.
Think: The Problem Before the Tool
Think means starting with the problem, not the tool. Before opening a chat, answer three questions. Who is the audience? What decision do we need from them? What do we believe the answer is? Even a rough outline written from your own head changes the output more than any prompt technique. It gives the AI a point of view to work toward and gives you a standard to check the draft against later. Without it, the AI defaults to summarizing everything it was given, which is how a factual inventory ends up with editorial commentary and no clear next step.
Direct: The Context and the Rules
Direct means giving the AI the context you would give a teammate taking over the work: the background, the storyline, what to include and leave out, and an example of what good looks like. It also means setting rules. Most people fix the same problems in every AI draft. The better approach is to fix them once.
The first thing we recommend is a standing set of instructions, saved in the AI’s memory or at the top of every prompt, that covers the edits you would otherwise make by hand. Start with rules like these. Lead with the answer, then the detail. Use plain English and remove unnecessary context or conversational notes. No em dashes. Remove unnecessary adjectives. Avoid multiple one-line hooks and hyperboles. For quick requests, make a reasonable assumption and state it instead of asking five questions. Flag anything that’s a guess. Match my tone when I provide an example.
The habit underneath matters more than any single rule: any fix you make twice becomes a rule. For work that repeats every week, like meeting recaps and status updates, we go a step further and save the full workflow so it runs the same way each time. A good set of rules is also far easier to share across a team than a good prompt.
Build: One Section at a Time
Build means working section by section and asking the AI to raise clarifying questions before it starts. Left alone, it will run ahead on assumptions. Remind it to ask questions before it starts, and hit “stop” if it starts running before you’re done explaining or have cleared your mind of inputs. Review each section before moving to the next. A wrong assumption caught in the first section takes minutes to fix. Caught in the final draft, it can mean starting over.
A good rule of thumb: if a full deliverable comes back in under five minutes, it probably skipped steps. If AI could build a complete client deliverable on its own, most of us would be out of a job. Good work still takes real people, real thinking, and real time with the problem.
Verify: The Questions a Good Editor Asks
Verify means checking more than facts. Check whether the framing fits the audience and whether the output answers the question that was asked. A few questions do most of the work.
Can I defend it? If a senior leader asks “how do you know that?” or “what’s the rationale for recommending that?” the person presenting should be able to answer without going back to the chat. If they can’t, the work isn’t ready, no matter how good it looks.
What’s missing? AI summarizes what it was given. It won’t be able to read between the lines. In discovery work, the gap a client doesn’t recognize as a problem is often the most useful insight.
Who is this for? An internal prep document can be blunt. The client version needs different framing. AI won’t make that distinction unless someone tells it to.
What would a critic say? AI is useful here in a different role. Instead of asking it to write, ask it to do what a skeptical executive would do: test the assumptions against the raw data, find the strongest reasons the recommendation is wrong, and flag the evidence that’s missing. The executive questions may come in the form of “Test against a counter point of view to determine anything to add or prepare in advance” or “Test against an everyday marketer’s language for ease of understanding.” Just don’t let it have the final word on its own work. The last check should come from outside the chat, whether that’s the source data, a colleague, or the client’s own numbers.
The Honest Limits
Editing takes time, and it can eat into the time AI was supposed to save. A first draft in seconds followed by an hour of rework is often still faster than starting from a blank page, but only if the thinking happened up front. Without it, teams end up reworking the same draft over and over.
Rules also need upkeep. They can overcorrect: tell an AI to never use em dashes and it will strip out the one that was grammatically right. Models change and new tells appear, so a rule set from six months ago will miss things. No set of instructions replaces knowing the client. The context that makes the work useful still lives with the people on the engagement.
The Takeaway
AI has changed where the work happens. Less time goes into writing and more goes into judgment about what’s true and what the reader needs to do next. Teams that build these steps into how they use AI from the start spend less time fixing what it produces and more time solving the problem in front of them.
Being a boutique consultancy gives us an advantage here. Our team has access to leading AI tools, the freedom to experiment with them on real client work, and the builders to create what doesn’t exist yet. We also bring the domain knowledge and subject matter expertise that give us credibility when using AI. We pair that speed with the judgment to know when work is ready for a client, and that’s the standard we bring to every engagement. If you’re struggling with how to get your team using AI productively, or worried they will lose out on building the skills needed for the future, let’s talk.
Ali Barnard, Analyst
