Chapter 3
Separating the Brittle from the Unbreakable
Not All Pain Is Fixable
Sam's merge scored 9 pain signals—the screaming weak link across all three Media Buyers at Apex.
Jordan's error review scored 5—a major weak link.
Morgan's review wait scored 5—another major weak link.
But here's the question that matters: Can you actually decouple these links with AI?
Just because something is painful doesn't mean automation can fix it.
Client relationship building is painful. Requires empathy, reading subtle interpersonal cues, building trust over months. AI can't replace that.
Negotiating contract terms is stressful. Requires understanding leverage, reading the room, knowing when to push and when to compromise. AI can't navigate those dynamics.
Some links are painful but unbreakable (with current AI capabilities).
This chapter shows you how to test which links are brittle—automatable—and which should stay human.
By the end, you'll have an RDS Assessment showing exactly which weak links can be decoupled and at what level (full automation, AI assists, or Keep Human-Led).
What AI Actually Is (and Isn't)
The AI hype says: "AI can do anything! Automate your entire workflow!"
That's not true.
AI is not magic. It's not sentient. It's not "intelligent" in the way humans are.
AI is fundamentally three things:
A Pattern Matcher
Learns from examples.
AI doesn't understand WHY campaigns with certain characteristics perform better. It recognizes patterns: "In the last 1,000 campaigns, when X characteristics were present, ROAS averaged Y."
Needs repeated patterns to recognize and replicate.
If you show AI 100 examples of successful campaign structures, it can recognize similar patterns in new campaigns. If every campaign is completely unique, AI has nothing to learn from
A Guided Instruction Follower
There's a critical distinction between traditional automation and modern AI:
Traditional automation:
Follows explicit "if this, then that" logic perfectly
Cannot deviate from the script
Cannot infer anything unstated
Breaks when encountering exceptions
Example: Traditional script says "Match on Campaign ID field" → If ID doesn't exist, script fails completely.
Modern AI:
Can follow loosely guided instructions
Can infer SOME unstated rules within defined boundaries
Can handle variations and edge cases
Example: "Match campaign names across platforms, accounting for spacing and naming variations"
→ AI handles "Brand_Search_Q1" matching to "Brand Search - Q1"
→ Learns patterns from examples
→ Can generalize to new variations it hasn't seen
But modern AI still can't:
Make nuanced decisions with incomplete information
Say "this feels wrong for reasons I can't articulate"
Apply deep contextual knowledge it doesn't have
Navigate truly novel situations with no similar historical patterns
The line: AI can handle structured variation (campaign naming differences). AI can't handle unstructured judgment (is this client relationship deteriorating?).
A Probability Engine
Makes predictions based on training data.
AI gives you likely answers with confidence scores, not guaranteed correct answers.
When AI merges campaign data, it might be 98% confident "Brand Search Q1" in Meta matches "Brand_Search_Q1" in Google. That 2% uncertainty is why human review matters.
Therefore:
AI is great at:
Repeatable tasks (same pattern every time, or predictable variations)
Definable tasks (clear rules for "correct" that can be checked)
Safe-to-delegate tasks (errors are obvious and easy to fix before external impact)
AI struggles with:
Novel situations (no historical pattern to match against)
Subjective judgment (no objective "right answer," context-dependent)
High-stakes errors (mistakes are catastrophic, hard to detect, or have external consequences)
The RDS Assessment tests three dimensions to separate brittle links (can be decoupled) from unbreakable ones (Keep Human-Led).
Important note about AI evolution
AI capabilities evolve rapidly. What was impossible 18 months ago is routine now.
This means:
Design for adaptability:
Build automation systems that can incorporate better AI models over time. Don't hard-code today's limitations into your architecture.
Reassess failed pilots:
If you tried to automate something 2 years ago and it didn't reach desired accuracy (scored RDS = 5), test it again with current AI models. The weak link might still be relevant, and AI might now be capable.
Keep learning:
What's unbreakable today might become brittle tomorrow. Revisit your "Keep Human-Led" decisions annually.
But be realistic about current capabilities when scoring. Score based on what AI can do TODAY, not what it might do in 3 years.
The Three RDS Dimensions
For each weak link, score 1-3 on:
R - REPEATABLE (Pattern Strength)
D - DEFINABLE (Criteria Clarity)
S - SAFE (Error Tolerance)
Each dimension scored 1-3. Minimum possible score: 3 (if every dimension scores 1). Maximum: 9 (if every dimension scores 3).
Here's what each dimension tests:
R - REPEATABLE: Can AI Learn the Pattern?
The core question:
"If you gave someone 10 past examples of this task, could they write instructions that would work for the next 10 instances?"
Why this matters:AI learns from patterns in examples. If every instance follows a repeatable structure (even with variations), AI can learn the pattern. If every instance is completely unique, AI has nothing to learn from.
Score 3 — Almost identical every time, or variations follow clear rules
What this means:
Same steps in same order, OR variations are systematic and predictable
Observable test:
Last 5 times you did this task look essentially the same
Changes follow learnable patterns (Platform A always uses format X, Platform B always uses format Y)
You could write a 1-page checklist that works every time
Example:
Data extraction from standard sources (same fields, same format every month—or predictable platform-specific variations like "Google Ads always uses alphanumeric IDs, Meta always uses numeric IDs")
When to score 3:
Can you create a step-by-step checklist that works every time?
Could you train a new hire to do this with a 1-page instruction doc?
Do you do the exact same steps in the exact same order?
OR: Do variations follow predictable rules that can be written down?
Score 2 — Follows loose structure with predictable variations
What this means:
There's a template or pattern you follow, but you adapt based on context
Observable test:
There's a recognizable structure, but details vary by situation
You need judgment to decide how to adapt the template
Similar, but not identical, each time
Example:
Writing client email updates (always same sections—progress, issues, next steps—but specific content varies by project. You're following a template but customizing based on what happened this week and client's communication style)
When to score 2:
Is there a template or structure you follow, but with judgment calls?
Do you adapt the approach based on circumstances?
Could you teach this, but it would require examples and coaching, not just a checklist?
Score 1 — Every instance unique, no learnable pattern
What this means:
Every situation is fundamentally different; you "figure it out fresh" each time
Observable test:
Last 5 times sound completely different from each other
No template or pattern connects them
Requires deep contextual knowledge and judgment
Example:
Contract negotiation (each client's leverage, needs, alternatives, and relationship dynamics are unique—what worked with Client A won't work with Client B)
When to score 1:
Is every situation fundamentally different?
Do you "figure it out fresh" each time?
Would teaching this require months of apprenticeship and contextual knowledge?
Could you even create a template, or is each instance too unique?
D - DEFINABLE: Can You Write the Rules?
The core question:
"Can you write down the rules for when this task is 'done correctly'?"
Why this matters:AI needs explicit success criteria. If "correct" is objective and checklistable, AI can validate its own output. If "correct" is based on taste, judgment, or contextual nuance, AI can't know if it succeeded.
Score 3 — Clear objective checklist, pass/fail criteria
What this means:
"Done correctly" can be defined with factual yes/no criteria—no interpretation needed
Observable test:
Two people independently checking would always agree
All success criteria are measurable (count, compare, verify presence)
No room for debate about whether it's "correct"
Example:
"Are all 119 campaigns present in the merged dataset?" (count expected vs actual, yes/no answer. Either all 119 are there or they're not—no interpretation required)
When to score 3:
Can you create a checklist where every item is yes/no?
Would two people independently checking always reach the same conclusion?
Are the criteria purely factual? (numbers match, fields present, format correct)
Score 2 — Guidelines exist but require interpretation
What this means:
There are standards, but applying them needs judgment about quality or completeness
Observable test:
General guidelines exist, but "good enough" is somewhat fuzzy
People sometimes disagree on whether something meets the standard
Part objective (format), part subjective (quality)
Example:
"Does this project update match our communication standards?" (format is definable—includes timeline, next steps, blockers—but "is it complete and clear?" requires judgment. One reviewer might want more detail, another thinks it's sufficient)
When to score 2:
Are there guidelines, but they require interpretation?
Do people sometimes disagree on whether something is "done correctly"?
Is there a "good enough" threshold that's somewhat subjective?
Score 1 — Highly subjective, gut feeling only
What this means:
Success is based on taste, feel, or contextual judgment—"I know it when I see it"
Observable test:
No objective criteria can be written down
Experts often disagree on what's correct
Quality depends on unstated context or aesthetic judgment
Example:
"Is this email tone persuasive to this specific client?" (opinion-based, requires relationship knowledge. What feels persuasive to one person might feel pushy to another. Context about client personality and relationship history determines success)
When to score 1:
Is quality based on aesthetic judgment or "feel"?
Do experts often disagree on what's correct?
Is success context-dependent with no universal criteria?
S - SAFE: Can Errors Be Caught Safely?
The core question:
"If AI did this task wrong, could you catch the error and fix it in under 5 minutes—before it causes real damage?"
Why this matters:
AI WILL make mistakes. Always. The question isn't "will it make errors?" but "are those errors obvious, trivial to fix, and caught before they cause damage?"
This dimension measures: How safe is it to delegate this task to AI with human oversight?
Score 3 — Obvious errors, trivial to fix before any delivery
What this means:
You'd spot the error in 30 seconds by glancing at output, fix in 2 minutes, and no external stakeholders see it
Observable test:
Errors are visually obvious (missing data, totals don't add up)
This is internal-only (you review and approve before anything goes external)
Quick to fix (re-run automation, 2-minute manual correction)
Validation checks exist (expected counts, required fields, totals match)
Example:
Campaign count wrong in merge (see immediately: "Expected 119, got 117"—two campaigns missing. Re-run the automation or manually add the missing campaigns in 90 seconds. Nothing has been sent to client yet)
When to score 3:
Are errors visually obvious? (numbers don't add up, missing data jumps out)
Is this internal-only? (no external stakeholders see it until you review and approve)
Can you fix it in under 5 minutes? (re-run automation, quick manual correction)
Are there validation checks? (totals must match, required fields present)
Score 2 — Internal rework needed, time-consuming but fixable
What this means:
Error requires re-doing work internally, but caught before external stakeholders see it
Observable test:
Errors aren't immediately obvious—need investigation to find
Rework is time-consuming (30-90 minutes) but doable
Consequences are internal (team frustrated, deadlines slip, but no client impact)
Caught in internal review before going external
Example:
Internal report has wrong numbers (analysis shows Campaign X drove 40% of conversions, but actual was 25%. Team needs to redo the entire analysis with correct data—takes 90 minutes. Frustrating and delays the project, but client doesn't see the error because it's caught in internal QA)
When to score 2:
Do errors require investigation to find? (not immediately obvious, need to dig)
Is rework time-consuming? (takes 30-90 minutes to fix, but doable)
Are consequences internal? (team frustrated, deadlines slip, but no client impact)
Score 1 — Serious external damage possible
What this means:
Mistake reaches clients/customers and causes significant damage (revenue loss, relationship harm, legal issues, safety problems)
Observable test:
Errors are hard to detect (look plausible, require deep expertise to spot)
External stakeholders see mistakes before you catch them
Damage is significant (lose contract, harm relationship, compliance violation, financial loss)
Example:
Sending pricing proposal to client with wrong numbers (shows $250K instead of $350K due to formula error. Client accepts the lower price. You've just lost $100K in revenue and can't backtrack without damaging trust. The $800K annual contract relationship is now at risk)
When to score 1:
Are errors hard to detect? (look plausible, require deep domain knowledge to spot)
Are consequences external? (clients, partners, regulators see this before you catch it)
Is damage significant? (revenue loss, relationship damage, legal/compliance issues)
The RDS Scoring Scale
Add the three scores. Minimum possible: 3 (if every dimension scores 1). Maximum: 9 (if every dimension scores 3).
Score
Automation Level
Human Role
9
Full Automation
Validator: AI does it, human spot-checks 2-5 min
7 - 8
High Automation
Reviewer: AI does it, human reviews every output and sensitive actions 5-10 min
5 - 6
AI Drafts
Editor: Human finishes and refines AI's first draft (saves 60% of original time)
4
AI Assists
Operator: Human does it, AI helps with pieces (saves 30% of original time)
3
Keep Human-Led
Owner: Human owns it, AI can't meaningfully help
RDS Across Industries
Here's how different types of work score on RDS:
High RDS scores (8-9)
Construction
"Consolidate 15 subcontractor invoices from AIA G702/G703 forms"
R=3 (same standard forms monthly, predictable variations by sub)
D=3 (line items match source PDFs, totals validate, all subs present)
S=3 (internal, Principal catches errors in review before client sees)
RDS = 9
E-commerce
"Export inventory levels from Shopify and calculate reorder points"
R=3 (same process weekly, formula-based)
D=3 (all SKUs present, formula validates: velocity × lead time + safety stock)
S=3 (internal, obvious if quantities wrong)
RDS = 9
SaaS Company
"Export product usage scores from analytics dashboard and aggregate with Salesforce data"
R=3 (same export weekly, predictable account name variations)
D=3 (all accounts present, data format correct, numbers validate)
S=3 (internal, easy to verify by spot-checking totals)
RDS = 9
Marketing Agency
"Export ad platform data and merge into master spreadsheet"
R=3 (same steps monthly—or predictable variations like different ID formats that AI can learn)
D=3 (all campaigns present, totals match ±2%, no formula errors)
S=3 (errors obvious in 30-second scan, re-run in 2 min, internal-only)
RDS = 9
Pattern: Extracting data from standard sources (even with format variations AI can learn) and consolidating = high automation potential
Medium RDS scores (5-6)
E-commerce
"Write supplier email explaining why reorder quantities changed significantly"
R=2 (template exists, but reasons vary: seasonality, promotion, trend shift)
D=2 (professional tone is somewhat subjective, clarity varies)
S=2 (supplier-facing, affects relationship)
RDS = 6
SaaS Company
"Write executive summary of at-risk accounts (why at risk, recommended actions)"
R=2 (same format—account, risk factors, actions—but context varies)
D=2 (good recommendations require judgment about customer context)
S=2 (influences exec decisions, needs accuracy)
RDS = 6
Marketing Agency
"Write performance narrative for client (which campaigns won/lost, why, what to do next)"
R=2 (same structure—winners/losers/recommendations—but insights vary by client's specific data and goals)
D=2 (good narrative is somewhat subjective: "Is this insight valuable?" requires judgment)
S=2 (client-facing, needs accuracy and strategic soundness)
RDS = 6
Pattern: AI can draft first pass using the pattern/structure, human adds strategic insight, contextual knowledge, and refinement.
Low RDS score (4)
These are barely automatable. AI can gather supporting information, but human owns the work.
E-commerce
"Decide which additional products to add to hit supplier MOQ (based on sales trends, seasonality, inventory space, gut feel)"
R=1 (different every time based on current market conditions)
D=1 (based on feel for what will sell)
S=2 (wrong choice = dead inventory, but internal consequence)
RDS = 4
SaaS Company
"Decide which at-risk customer to prioritize for outreach (based on relationship history, contract terms, competitive alternatives)"
R=1 (context-specific to each account's unique situation)
D=2 (prioritization criteria exist but fuzzy: strategic value, churn likelihood, save-ability)
S=1 (wrong call = customer churns, revenue lost)
RDS = 4
Marketing Agency
"Review client deck for strategic alignment with their specific business goals"
R=1 (every client's goals and context unique)
D=2 (some guidelines: addresses their priorities, recommendations are actionable)
S=1 (client-facing, wrong strategic direction damages $800K annual relationship)
RDS = 4
Pattern: Keep Human-Led. AI can surface relevant data (past sales, contract terms), but human makes the judgment call.
Minimum RDS score (3): Relationship, strategy, novel problems
Don't try to automate these.
Any Industry
"Build relationship with key client stakeholder"
R=1 (each person unique, relationship dynamics unpredictable)
D=1 (relationship quality subjective, context-dependent)
S=1 (lose relationship = lose account)
RDS = 3 → Keep Human-Led
Marketing Agency
"Negotiate media buying rates with ad platform rep"
R=1 (every negotiation's leverage and context unique)
D=1 ("good deal" is contextual: depends on client budget, relationship, alternatives)
S=1 (client's $10M annual spend at stake)
RDS = 3 → Keep Human-Led entirely
Pattern: These are the jobs humans excel at. Don't automate the valuable judgment work.
RDS Scoring for Apex Weak Links
This section demonstrates scoring each weak link from Chapter 2 using the RDS Assessment. Apply the same process to your weak links.
Remember, from Chapter 2 we revealed:
Sam merges three CSVs (9 pain signals)
Jordan reviews for errors (5 pain signals)
Morgan reviews deck (5 pain signals)
Sam fixes formula errors (3 pain signals)
Now you'll test: which of these are actually automatable?
Sam merges three CSVs into Excel template
Merge three CSVs into master
Excel template
90 Min
This scored 9 pain signals (highest). But can AI actually decouple it?
R - REPEATABLE: Can you write instructions from 10 examples?
The Director asks Sam: "Walk me through the last 5 times you did this merge for TechVantage. How similar were they?"
Sam: "Literally identical structure. Every month:
Open the master Excel template for TechVantage
Import Google Ads CSV (67 campaigns, always same columns: Campaign Name, Spend, Conversions, CPA, ROAS)
Import Meta CSV (34 campaigns, different column names but same metrics)
Import LinkedIn CSV (18 campaigns, account ID prefix issue)
Match campaigns across platforms (same campaign list monthly—Google's 'Brand_Search_Q1' always maps to Meta's 'Brand Search Q1')
Populate master tabs with metrics
The structure is identical monthly. Campaign names stay the same. The only things that change are the metric values—spend went up or down, ROAS changed. But the framework is exactly the same."
The Director asks: "What about the variations—like different ID formats across platforms?"
Sam: "Those variations are consistent and predictable:
Google ALWAYS uses alphanumeric IDs like '12345ABCD'
Meta ALWAYS uses numeric IDs like '98765432'
LinkedIn ALWAYS prepends the account ID like '506849291_CampaignName'
It's not random. An AI could learn 'Google uses this ID format, Meta uses that ID format, match them by campaign name instead since IDs don't align.'"
The Director asks: "Could you write a step-by-step instruction document that would work every month?"
Sam: "Absolutely. In fact, I trained our intern on this last year. Gave her a 2-page checklist:
Step 1: Download these 3 CSVs
Step 2: Open master template
Step 3: Match campaigns using name (not ID), here are the naming pattern variations
Step 4: Copy these specific columns from each source
Step 5: Validate totals match
She followed it perfectly. If an intern can learn it from a checklist, AI definitely can."
→ Score: 3 (Highly repeatable—same structure monthly, variations follow learnable patterns)
D - DEFINABLE: Can you write rules for 'correct'?
The Director asks Sam: "How do you know when the merge is 'done correctly'?"
Sam: "I have a validation checklist:
All 67 Google campaigns present in master ✓ (count = 67)
All 34 Meta campaigns present ✓ (count = 34)
All 18 LinkedIn campaigns present ✓ (count = 18)
Total campaign count = 119 ✓
Each campaign's Spend metric matches source CSV ✓ (spot-check 5 campaigns)
Total spend across all platforms = sum of individual campaign spends ✓ (math check: $480K + $300K + $53K = $833K)
No blank cells in required metric columns ✓ (visual scan)
No #VALUE! or #DIV/0! or #N/A formula errors ✓ (visual scan)
Blended ROAS calculates correctly ✓ (formula check)
Those are all objective yes/no checks. Two people validating would get the same answer."
The Director asks: "Is 'correct' based on judgment or facts?"
Sam: "Pure facts. Either Campaign X is present or it isn't. Either the total is $833K or it isn't. Either there are formula errors or there aren't. No interpretation needed."
→ Score: 3 (Fully definable with objective checklist)
S - SAFE: Can you catch and fix mistakes in under 5 minutes?
The Director asks Sam: "If AI did this merge and made a mistake, how would you know?"
Sam: "I'd spot-check—would take about 3 minutes:
Check campaign count: Expected 119 total (67 + 34 + 18). If I see 117, something's missing—immediately obvious.
Check total spend: Should be around $833K for TechVantage. If it shows $650K, huge red flag.
Scan for blank cells: Visual check, takes 15 seconds.
Look for formula errors: #N/A or #VALUE! jumps out visually.
Errors would be obvious within 2-3 minutes of opening the file."
The Director asks: "If you found an error, how long to fix?"
Sam: "Depends on the error:
Campaign missing: Re-run the automation, or manually add one row from source CSV—30 seconds
Wrong metric value: Check source CSV, correct the cell—1 minute
Formula error: Copy formula from adjacent cell—15 seconds
Under 5 minutes to fix any typical error. And worst case, I can always fall back to my manual process if AI completely fails—would take 4 hours, but it's there as emergency backup."
The Director asks: "Who sees this if it's wrong?"
Sam: "Jordan sees it next in his error review. It's internal at that point—client doesn't see anything until Jordan creates the final presentation deck after validating the data. So mistakes get caught in our internal review, not in front of the client. Safe to delegate with review."
→ Score: 3 (Safe—errors obvious, easy to fix, internal-only, multiple review gates before client sees anything)
RDS TOTAL: 9/9 → Full Automation Potential
What this means:
AI can handle the merge end-to-end:
Extract from 3 platform APIs (or CSVs)
Learn the campaign naming variations (underscores vs spaces vs ID prefixes)
Match campaigns across platforms despite ID format differences
Populate master Excel with validated data
Sam's role transforms to Validator. reviews the output for 3-5 minutes (not 4.5 hours):
Read QA summary email
Spot-check 3 campaigns against source
Validate totals
Approve if clean
Expected time savings: 4.5 hours → 5 minutes per client (98% reduction)
Scaled: 67.5 hours/month → 1.25 hours/month for all Media Buyers (66.25 hours freed)
This is a jackpot candidate for automation.
And it's not just Sam: Jeremy and Steven score this link identically (9/9). All three Media Buyers can be freed from this weak link.
Jordan reviews consolidated data for errors
Account Manager
Review consolidated data for obvious errors
30 Min
This scored 5 pain signals (major weak link). Can AI help?
R - REPEATABLE:
The Director asks Jordan: "Walk me through how you review Sam's data. Same process every time?"
Jordan: "Sort of. I have two types of checks:
Objective checks (same every time):
Campaign count correct? (Should be 119 for TechVantage)
Total spend adds up? (Platform totals should equal grand total)
Any blank cells? (red flag)
Any campaigns with $0 spend? (usually means data error)
Contextual checks (vary by what's happening with the client):
Campaign spend way up or down from last month? (But 'way off' depends—did we increase budget intentionally? Are they testing new creative?)
ROAS significantly worse? (Could be real underperformance, or could be we're in test phase)
New campaigns appearing? (Expected if we launched new ads, or error if it's wrong client data)
The objective checks are repeatable—AI could do those. The contextual checks require me knowing what we're doing with TechVantage this month."
→ Score: 2 (Partially repeatable—some checks are objective and automatable, some require context)
D - DEFINABLE:
The Director asks Jordan: "Can you write rules for what's an 'error' vs. 'correct'?"
Jordan: "For objective errors, yes:
Campaign that was present last month is missing this month (and we didn't intentionally pause it) → Error
Total spend doesn't equal sum of platform spends → Error
Any blank cells in Spend, Conversions, CPA, ROAS columns → Error
Any campaign with exactly $0 spend (usually means data didn't pull) → Likely error
For judgment-based checks, harder:
Campaign X's ROAS dropped 30% → Is that a data error or real underperformance? Depends on whether we changed creative, adjusted targeting, shifted budget.
Meta spend doubled → Did we increase budget strategically, or is this Sam pasting wrong client data? Need to cross-reference with our media plans.
I need context to decide if changes are errors or real performance shifts."
→ Score: 2 (Some errors are definable objectively, others require contextual judgment)
S - SAFE:
The Director asks Jordan: "If AI flagged something as 'possibly wrong,' how would you verify?"
Jordan: "I'd check the source data or ask Sam. If it's truly an error, we fix it before the client sees anything. Takes 30 minutes to investigate and correct. Internal rework, annoying but not catastrophic."
The Director asks: "What if AI missed an error and you missed it too?"
Jordan: "Depends on the error:
If totals don't add up and we both missed it: I'd catch it when writing the narrative because numbers wouldn't make sense. Would realize total spend is wrong.
If there's a real data error (wrong client's campaign data): Client might notice in our presentation. 'Wait, we don't run a campaign called X.' Embarrassing. We'd correct it live and send updated deck. Damages credibility a bit but not catastrophic.
Worst case: Client loses confidence in our data quality. Could affect renewal if it happens repeatedly."
→ Score: 2 (Moderately safe—internal rework if caught early, relationship risk if errors reach client, but multiple review gates reduce probability)
RDS TOTAL: 6/9 → AI Assists (Flags Issues for Jordan's Review)
What this means:
AI can flag objective errors automatically:
Missing campaigns (expected 119, found 117 → flag)
Totals don't match (platform sum ≠ grand total → flag)
Blank cells detected → flag
Spend = $0 for active campaign → flag
AI can flag anomalies for Jordan's contextual investigation:
Spend variance >50% vs last month → flag with context: "Review recommended: Meta spend up 65%"
ROAS change >0.5 → flag with context: "5 campaigns show ROAS drop >30%"
Jordan reviews flagged items only (instead of all 119 campaigns line-by-line)
Expected time savings: 30 min → 10-15 min per client (50% reduction)
Not full automation, but helpful assistance.
Note: If we decouple #1 (the merge), #2's error rate will likely drop. Jordan catches errors 25% of time now (mostly Sam's merge mistakes). If AI merge is 98% accurate, Jordan might only find errors 5% of time. This weak link's pain signals could drop from 5 to 2-3.
Morgan reviews deck
Client Success Director
Review deck for narrative quality/strategic alignment and approve
30 Min
This scored 5 pain signals. Can AI help?
R - REPEATABLE:
The Director asks Morgan: "What are you checking when you review these decks? Same process for all 15 clients?"
Morgan: "No, every client is different. I'm checking:
For TechVantage specifically:
Does the narrative address their stated goal of scaling enterprise lead gen? (They told me in our last QBR they want to double enterprise pipeline)
Are the recommendations aligned with their risk tolerance? (They're conservative, don't like big swings—Jordan's recommendations should reflect that)
Does this set up the conversation I need to have with their CMO? (I know she's getting pressure from the board about CAC—does this deck help or hurt that conversation?)
For Client B (different context):
They're aggressive testers, want bold recommendations
They care about brand awareness metrics, not just direct response
Different tone needed entirely
Every client review requires me applying unique contextual knowledge about their business, their stakeholders, their strategic priorities. Can't checklist it."
→ Score: 1 (Every review is contextually unique, no repeatable pattern AI could learn)
D - DEFINABLE:
The Director asks Morgan: "Can you write rules for what makes a 'good deck'?"
Morgan: "I can write guidelines:
Narrative addresses client's stated strategic priorities (but those priorities are different for every client and evolve monthly)
Recommendations are specific and actionable (but 'actionable' is subjective—what TechVantage can execute vs what Client B can execute is different)
Data supports the narrative (that's objective)
Tone matches client relationship maturity (formal for new clients, casual for long-term partners—judgment call)
The data validation is definable. The strategic alignment and tone are fuzzy. Two directors reviewing the same deck might disagree on whether the tone is right or recommendations are strategic enough."
→ Score: 2 (Some guidelines exist, but most criteria require interpretation and contextual judgment)
S - SAFE:
The Director asks Morgan: "If AI approved a deck without your review and Jordan presented it to TechVantage, what could go wrong?"
Morgan: "We could recommend something that's technically data-driven but strategically tone-deaf. Like last month, the data said 'pause the brand awareness campaign—low ROAS.' But I know from conversations with their CMO that the board is pushing them to build brand recognition in the enterprise space. If we recommend pausing brand campaigns without that context, we look like we don't understand their business.
That damages the relationship. They're paying us $800K annually not just for data, but for strategic partnership. If we seem to miss the bigger picture, they'll question the value."
The Director asks: "How bad is that damage?"
Morgan: "Could lose the contract at renewal. These are multi-million dollar relationships. One tone-deaf presentation might not kill it, but a pattern of missing the strategic context would."
→ Score: 1 (Not safe to delegate—$800K annual contract at risk if strategic misalignment damages client confidence)
RDS TOTAL: 4/9 → Keep Human-Led (AI Can't Replace Strategic Review)
What this means:
Morgan's review is valuable judgment work that requires contextual knowledge about each client's unique business situation, stakeholder dynamics, and strategic priorities
AI could help at the margins:
Pre-populate a summary: "Deck recommends pausing Brand Campaign X. Client's stated Q1 goal: increase brand awareness. Potential misalignment—review recommended."
Surface relevant context: "Last QBR notes: CMO mentioned board pressure on brand metrics"
But Morgan still needs to read the deck, apply her client relationship knowledge, and make the judgment call
Expected time savings: Maybe 5-10 minutes from AI-generated context summary, but 20-25 minutes of thoughtful review still needed
Don't try to automate this. It's real judgment work that protects $800K relationships.
Note: The 24-36 hour wait is a separate issue (Morgan's calendar availability, not the review itself). Address that by clearing Morgan's calendar or delegating some reviews—not by automating the strategic judgment.
Sam fixes formula errors and validates data
Fix formula errors and validate data
45 Min
This scored 3 pain signals. Can AI help?
Important context: If we automate Weak Link #1 (the merge), this link might disappear entirely. The formula errors are caused by manual merging—copy-paste mistakes, format inconsistencies. If AI does the merge correctly, there shouldn't be formula errors to fix.
You'll score it anyway in case some errors persist:
R - REPEATABLE:
Sam: "The errors are always the same types:
#N/A errors from VLOOKUP failures (happens because IDs don't match)
#VALUE! errors from text/number format mismatches
#DIV/0! errors from blank cells in denominators
Formulas that didn't copy down to all rows
Same error types monthly. Same fixes: remove VLOOKUPs that failed, reformat cells as numbers, copy formulas down, fill blanks."
→ Score: 3 (Same error types, same systematic fixes)
D - DEFINABLE:
Sam: "Yes, very definable:
Scan all cells for #N/A → If found, review VLOOKUP logic (usually needs to match on name not ID)
Scan for #VALUE! → If found, check if source cell is text vs number → reformat as number
Scan for #DIV/0! → If found, check denominator → if blank, fill with 1 or flag as missing data
Check formula coverage → If rows 1-67 have formula but rows 68-119 are blank → copy formula down
Objective error detection and correction rules. Two people would apply the same fixes."
→ Score: 3 (Clear error detection and correction rules, no judgment needed)
S - SAFE:
Sam: "If AI missed a formula error, Jordan would probably catch it when he reviews—totals wouldn't add up, or ROAS would look absurd like 500.0. And it's internal—client doesn't see this until after Jordan validates. Safe enough."
→ Score: 3 (Safe—caught in downstream review, internal-only, easy to fix)
RDS TOTAL: 9/9 → Could Be Fully Automated
But: If we automate Weak Link #1 (the merge), this problem likely goes away. AI doesn't create the same formula errors that manual copy-paste does. AI won't paste text when it should be a number. AI won't forget to copy formulas down.
Decision: Focus on decoupling #1. This weak link should resolve automatically as a side benefit.
The Complete RDS Assessment
Weak Link
R
D
S
Total
Level
Sam merges three CSVs
3
3
3
9
Full
Jordan reviews for errors
2
2
2
6
AI Drafts
Morgan reviews deck
1
2
1
4
AI Assists
Sam fixes formula errors
3
3
3
9
Full
Insight
If we decouple #1 (the merge), we solve:
Sam's biggest pain point (9 signals, 4.5 hrs → 5 min)
Sam transforms: Manual executor → Validator → Freed for optimization analysis
The root cause of #4 (formula errors disappear)
~50% of #2's pain (Jordan catches fewer errors when merge is clean)
ONE decoupling fixes 2.5 weak links across the chain.
Your Action Plan
Create your RDS Assessment using the three-dimension scoring process.
Process:
For each weak link from Chapter 2, score 1-3 on three dimensions:
R - REPEATABLE: "Could you write instructions from 10 examples?"
Score 1: Every instance unique, no learnable pattern
Score 2: Follows loose structure with predictable variations
Score 3: Almost identical every time, or variations follow clear rules
D - DEFINABLE: "Can you write a checklist for 'correct'?"
Score 1: Highly subjective, gut feeling only
Score 2: Guidelines exist but require interpretation
Score 3: Clear objective checklist, pass/fail criteria
S - SAFE: "Can errors be caught and fixed in <5 minutes before external damage?"
Score 1: Serious external damage possible (client loss, revenue risk)
Score 2: Internal rework needed, time-consuming but fixable
Score 3: Obvious errors, trivial to fix, internal-only
Add the three scores (minimum 3, maximum 9)
When in doubt between two scores, go lower (conservative protects from failed builds)
What you'll have:
Each weak link scored on R, D, S (1-3 per dimension)
Total RDS score per link (3-9)
Automation level determined:
9 = Full automation (AI handles, human spot-checks)
7-8 = High automation (AI does it, human reviews)
5-6 = AI drafts (AI first pass, human refines)
4 = AI assists (human does it, AI helps)
3 = Keep Human-Led (AI can't meaningfully help)
What this gives you:
Directors get:
"I can defend automation decisions with objective criteria to the CFO. The merge scored 9/9—clearly automatable. Morgan's strategic review scored 4/9—we should keep that human. No one can argue with the RDS scores. And I can explain to the board: 'We're automating tedious data work (RDS 9), not replacing strategic judgment (RDS 4).'"
Team members get:
"Leadership understands why some of my work SHOULD be automated (the tedious VLOOKUP-failure merge) and why some should stay with me (Jordan's contextual error review, Morgan's strategic judgment). I'm not being replaced—tedious chains are being broken so I can focus on real work."
Everyone gets:
A reality check. Not all pain is automatable. Not all weak links are brittle. Focus on decoupling the links that are both painful AND breakable.
You know which weak links are brittle enough to decouple. Next: prioritize which to decouple first and spec exactly how.
