Nick QuirkScoping is where agency margin is won or lost. A scope written in twenty rushed minutes before a proposal deadline becomes a twelve-week project that runs 40% over hours. A scope written properly takes three to five hours of senior time — which is why it usually doesn't get written properly.
AI project scoping closes that gap. Used well, it turns discovery notes into a structured phase plan, task list, and draft timeline in about 20 minutes, leaving your senior people to do the part machines are bad at: judgement, risk, and pricing. Used badly, it produces a beautiful document full of invented hour estimates that your team can't hit.
Here's the workflow that separates the two.
What AI is genuinely good at in scoping — and what it isn't
Be honest about the division of labour before you build a process around it.
AI is reliably good at:
- Turning unstructured discovery notes into structured deliverables
- Decomposing a deliverable ("technical audit") into 15–25 discrete tasks
- Spotting gaps — the deliverable you forgot to include, the dependency you didn't name
- Producing three scope variants (lean / standard / premium) from one input
- Rewriting internal task language into client-facing scope language
AI is unreliable at:
- Hour estimates, unless you supply your own historical data
- Knowing that this particular client takes 11 days to approve anything
- Pricing, risk premiums, and where to draw the "out of scope" line
- Understanding that the client's dev team is a single contractor with a 3-week queue
The rule I use: AI drafts structure, humans supply constraints and numbers. Every time an agency skips the second half, the scope becomes a liability.
Build a context stack before you write a single prompt
Scope quality tracks input quality almost linearly. Before prompting, assemble a "context stack" — one document containing:
- Discovery raw material. Call transcript or notes, the RFP, the client's stated goals in their own words.
- Technical reality. Site size (pages indexed), CMS, hosting, dev resource availability, current traffic and revenue baseline.
- Data you already have. Crawl summary, Core Web Vitals, keyword gap export, backlink profile snapshot.
- Commercial constraints. Budget or retainer hours, contract length, roles available and their weekly capacity.
- Two or three comparable past projects with their actual logged hours by phase — not the estimates.
Point five is the one everyone skips and the one that matters most. If your last three ecommerce technical audits took 34, 41, and 38 hours, AI now has a defensible range. Without it, it will confidently say "16 hours" because that's the median number in its training data.
If your time tracking lives in the same system as your projects, this takes two minutes to pull. If it's scattered across timesheets and spreadsheets, fixing that is a higher-leverage project than any prompt engineering. Agencies comparing systems on exactly this point usually start with a roundup of PM tools built for agency work rather than generic task managers.
A three-pass prompt framework
Don't ask for the whole scope in one shot. Three passes produce dramatically better output.
Pass 1: The discovery digest
Ask AI to summarise, not create. Something like:
"From the attached discovery notes and data exports, extract: (a) the client's three stated business objectives, (b) every deliverable they explicitly requested, (c) every deliverable they implied but did not request, (d) every constraint mentioned (budget, timing, internal resource, legal), (e) every open question we still need answered before scoping. Do not add recommendations."
That last instruction matters. Section (e) is the money — it typically surfaces four to six things nobody wrote down, like "who owns the staging environment" or "is the product feed managed in-house."
Pass 2: The phase skeleton
Now ask for structure with your constraints baked in:
"Using the digest above, draft a 6-month SEO engagement in phases. Constraint: 40 retainer hours/month, split across one SEO strategist (16h), one technical specialist (12h), one content lead (12h). Each phase needs a name, an objective, deliverables, and an exit criterion. Front-load work that unblocks later phases. Flag anything that cannot fit in the hour budget."
The "exit criterion" requirement is what turns a task list into a real plan. "Phase 2 ends when all P1 crawl errors are resolved and validated in Search Console" is testable. "Phase 2: Technical fixes" is not.
If you're scoping a competitive engagement, the same logic applies to research work — we broke down how to structure that in turning competitor data into project phases.
Pass 3: Task decomposition with estimates
Finally, go granular — and supply your actuals:
"Break Phase 1 into tasks of 1–4 hours each. Assign each to a role. For estimates, use these historical actuals from comparable projects: [paste]. Where you have no comparable data, mark the estimate as ASSUMPTION and explain your basis. Output as a table."
The ASSUMPTION flag is your review queue. On a typical 6-month scope you'll get 8–15 flagged items; those are the ones your senior strategist checks in ten minutes rather than re-estimating everything.
Turning the scope into a timeline that survives contact with clients
A task list with hours is not a timeline. The conversion needs three things AI won't infer.
Client latency. Give it a number. "Client approvals take 5 business days; dev deployments happen fortnightly on Thursdays." Without this, AI will chain tasks end-to-end and produce a timeline that's 30–40% too optimistic. In my experience, approval and deployment latency accounts for more slippage than execution ever does.
Real capacity. Nobody delivers 40 billable hours in a 40-hour week. Tell it to plan at 65–75% utilisation. On a 40-hour retainer that means roughly 28 hours of planned work with headroom for the algorithm update, the emergency migration, or the client's "quick question" that eats a Tuesday.
Dependency direction. Explicitly state what blocks what: crawl before fix list, fix list before content brief, brief before publish, publish before measurement window. Then ask for the critical path and the float on everything else. AI is genuinely good at this once told the rules.
One prompt that consistently earns its keep: "List every point in this timeline where a delay of 5+ days would push the final milestone. For each, suggest one mitigation." That output goes straight into your risk section — and often into the SOW as a client responsibility clause.
Milestones clients actually care about
AI defaults to activity milestones ("audit delivered"). Rewrite them as outcome milestones ("all P1 technical issues resolved and validated," "12 priority pages live and indexed"). Where the engagement is performance-based, anchor milestones to search data rather than deliverable dates — the approach outlined in
