The three categories of consulting work — and which ones AI should never touch
Share
It's Friday at 4:15 PM. You have a proposal due before the prospect cools off, a client deliverable that keeps slipping, and a meeting transcript from Tuesday that nobody's read yet. You open Claude. You type a paragraph. You read what comes back. You copy two sentences into the doc you're working on, close the tab, and go on with your day.
Total time saved: maybe six minutes. Total time spent: ten, including the context switch.
This is the failure mode that almost every solo consultant who uses AI runs into, and it is worth naming first because nothing about your prompts gets better until you fix it. The failure mode is using AI like a slightly faster Google. The frame is wrong, and as long as the frame is wrong, the leverage stays out of reach.
There is a different way to think about this, and it starts by accepting that not all consulting work is the same kind of work. Some of it is yours. Some of it is partly yours. Some of it shouldn't have been on your plate in the first place. AI's job is different in each case, and confusing them is what burns the hours that the productivity Twitter threads promised you back.
The three categories
A consulting week breaks into roughly three kinds of work. They want different tools, and using the same tool on all three is what produces six minutes of savings on a Friday afternoon.
Category 1 — Work that requires your judgment
Diagnosing a client's actual problem under the surface problem they described. Deciding what to put in scope and what to push back on. Reading a room. Calling out the political dynamic the client is dancing around. Writing the one sentence in the deliverable that they'll quote back to their CEO.
Time per week for the average solo consultant: 10–20 hours. AI-tractable share: roughly 0%.
This is what your bill rate buys. AI is bad at it and will sound confident anyway, which is the dangerous combination. The model doesn't have the context — it hasn't sat in the meetings, hasn't watched the body language shift when someone said "we're aligned on the priorities," doesn't know which stakeholder is the actual decider versus the named one. It will produce something fluent and unhelpful.
If you find yourself prompting an AI to do Category 1 work — "diagnose the real problem here based on these notes" — you are using AI to launder a hypothesis you should be sitting with longer. Sit with it longer.
Category 2 — Work that has structure but needs your voice
Writing the proposal. Drafting the kickoff email. Producing the strategy memo. Turning a finished engagement into a case study. Writing the section of the deliverable that synthesizes what you found. Drafting the follow-up that nudges a stalled deal back to life. Writing the newsletter post that drives the next deal.
Time per week: 8–15 hours. AI-tractable share: 60–80%.
This is where AI as first-draft engine earns its keep. You don't outsource the writing — you outsource the blank page. AI gets you to a 70% draft. You take it the last 30%, which is the part the client is paying for.
The 70% would otherwise have taken you 90 minutes. With a clear brief and a voice example, it takes 15. Compound that across a week and the math is real — three or four hours back, on the same calendar, without sacrificing anything that ships to the client.
The failure mode here is being lazy in the brief and producing AI-flavored prose that the client can sense even if they can't articulate why. Edit ruthlessly. Replace anything that reads as AI in the opening paragraph and the closing paragraph specifically — those are the two places where the smell is strongest.
Category 3 — Work that has structure and doesn't really need your voice
Pre-call research on a prospect. Cleaning up meeting notes into action items. Drafting a follow-up email that thanks the right person for the right thing. Generating four social posts from a finished case study. Writing the boilerplate sections of any deliverable. Pulling key facts out of a 90-minute Zoom transcript.
Time per week: 5–10 hours. AI-tractable share: ~95%.
This is where AI as executor lives. You give it inputs. You accept the outputs with light editing. You move on.
Most consultants do this work themselves out of habit — fifteen minutes here, twenty there — and never realize they've spent five hours on it by Friday. The five hours are the cheapest hours to fix. The work doesn't require taste; it requires structure, and structure is what AI is good at. The inversion most operators don't make is to default Category 3 to AI rather than reaching for it occasionally. The default should be: AI handles this unless you spot a specific reason it shouldn't.
The three-question screen
Before you put any task in front of Claude or ChatGPT, run three questions. If any answer is no, do the task yourself.
1. Can I describe the desired output in 100 words?
If you can't, you don't yet know what you want, and AI will helpfully invent something. Sit with the task for two minutes longer, write the brief in your head or on paper, and then come back. Most of the prompts that produce generic output do so because the brief was generic.
2. Can I tell at a glance whether the output is good?
If the output is in your domain expertise and you can spot drift, yes. If it's not — "draft the legal indemnity clause for this engagement" — no, because confident hallucination in your blind spot is how reputations die. AI will produce a clause that reads correctly to a layperson and isn't enforceable. You won't catch it until the lawsuit.
3. Is the cost of editing the output less than the cost of writing it from scratch?
Sometimes the answer is no. If editing the AI's draft would take 80% as long as writing your own, write your own. The AI was built to assist average writers, not the person who writes consulting deliverables for a living.
This is a 30-second screen. Run it before every Claude or ChatGPT tab.
What this looks like in a real consulting week
Concrete is more useful than abstract. Here's a Tuesday in a real consulting week, tagged by category. Some of these are the same task at different points in the engagement.
| Task | Time | Category | What you do |
|---|---|---|---|
| Read a prospect's website + LinkedIn before discovery call | 12 min | 3 | AI generates a 1-page dossier |
| Discovery call itself | 30 min | 1 | You. Phone in your pocket. |
| Decide whether to propose for this deal | 5 min | 1 | You. Quick judgment call. |
| Write the proposal | 90 min | 2 | AI drafts; you edit the opening, pricing rationale, and "why me" |
| Draft the kickoff email after they accept | 8 min | 3 | AI writes; you tweak two phrases |
| Draft the strategy memo for an active client | 4 hours | 2 | You outline; AI expands sections; you edit |
| Decide what to recommend in the strategy memo | (inside the 4 hours) | 1 | You. Several walks. |
| Pull action items from yesterday's client meeting | 7 min | 3 | AI from transcript; you scan for accuracy |
| Write a newsletter post about a wrapped engagement | 30 min | 2 | AI drafts from your project debrief; you punch up the lede |
| Quarterly retrospective on your pipeline | 90 min | 1 mostly | You with AI as a pattern-finder for the data, not the conclusions |
A week run like this — with the categories tagged consciously — gets you 6–10 hours of Category 3 work moved off your plate, 3–5 hours of Category 2 work compressed by ~70%, and Category 1 work fully preserved as yours.
The compound effect is whether you bill 25 hours a week or 35. Same calendar.
A note on Claude vs. ChatGPT
Both work. We use Claude as the default because it consistently produces sharper business reasoning on structured tasks — especially in Categories 2 and 3 — and it commits to a verdict when you tell it to commit. ChatGPT tends to hedge more by default and pad more in prose. Both are correctable with the right prompt; Claude requires less correction.
Where ChatGPT wins: live web search (better for Category 3 prospect research), image generation (irrelevant for most consulting work), and voice mode (occasionally useful for dictating a meeting debrief while walking).
If you only use one, get Claude. If you already pay for ChatGPT, the workflows in this article work in either — the prompts are tool-agnostic.
What to do this week
This is the part where most articles say "feel free to reach out." We don't. Concrete next step:
Pick the next three tasks on your week's calendar that you'd describe as "having structure." Tag each one mentally as Category 2 (needs your voice) or Category 3 (doesn't). Run the three-question screen on the Category 3 ones. The first one that passes the screen — give it to Claude with a 100-word brief. Time how long it takes you to edit the output. Compare to how long it would have taken to write it from scratch.
Most operators come back from this with one or two weekly tasks they'll never touch again. That's the unlock.
Going deeper
This article is the headline frame. The deeper version — eight workflow chapters that apply this model to specific parts of your consulting week, with the prompts and worked examples — is The Solo Consultant's AI Playbook. Chapter 1 is this article expanded; Chapters 2 through 8 are the workflows that drop you into a specific category-2 or category-3 task with the prompt that runs it.
If proposals are the bottleneck where this framework hits hardest in your week, the Proposal-Closer Prompt Pack is the depth-version of just that one workflow.
Either way, the thing that decides whether AI saves you any hours at all isn't the prompts. It's whether you decided which category each task belonged to before opening the tab.
— Digital Kreative