Listen instead with enhanced depth on any podcast platform.

What “bad AI” really means

Key idea: AI is “free” the way a puppy is free. The model isn’t the cost. The mess is the cost.

Bad AI in this context is the same. Its not about the model, but the mess it causes.

Here, bad AI comes in 3 different flavors:

1. Wrong fit: You delegate judgment-heavy decisions to an LLM with no guardrails. Example: letting AI set discount levels in wholesale outreach.

2. Wrong goal: You optimize for speed (more emails) when the expensive thing is accuracy For example, the AI sends one wrong discount to a specific segments that it was supposed to get it which will nuke your margins.

3. Wrong setup: You launch with no baseline metrics, no off-switch, and no review step. You only find problems when customers do.

So the question here is:

When to have a human in the loop?

The answer is The 100:1 Rule

If a single AI error can erase the value of 100 correct actions, a human gatekeeper isn’t optional, it’s a financial necessity. Use human-in-the-loop wherever mistakes are costly, irreversible, or brand-damaging.

The 9 hidden costs that drain your margins

These costs rarely sit on one line of the P&L, which is why they’re easy to miss.

Here’s a quick map you can scan and share.

When automation misfires: real-world cautionary tales

1. Air Canada chatbot ruling (2022): A tribunal ruled the airline responsible for incorrect info provided by its chatbot to one of their customers who purchased based on the bot's response.

   Lesson: your AI’s words are your words. Accountability (and cost) sits with you, not the model.

   
   

2. Consulting report hallucinations: This happened just recently with Deloitte. Reports drafted with generative AI have hit headlines for fabricated citations and quotes, triggering refunds and reputational damage.

   
   

   Lesson: If high-priority outputs aren’t reviewed, AI can turn expert work into expensive rework.

The napkin math: sales outreach in D2C

Scenario: You’re emailing potential retail partners.

Baseline (your store now with no AI, hypothetical scenario:

• 20,000 emails/month

• 1.5% positive replies = 300 replies

• 20% of replies become meetings = 60

• 30% of meetings become first orders = 18

• Margin contribution per first order = $600

• Monthly margin from outreach ≈ $10,800

Now you get happy after such results. So you decide to add AI to max your results.

So you hypothesize an experiment:

• Volume +30% with same team = 26,000 emails

• Reply rate +20% relative (to 1.8%) = 468 replies

• Meetings ≈ 94; Orders ≈ 28; Margin ≈ $16,800

• Cost of AI/Infra/tokens/seats ≈ $4,000/month

Now, lets say just 3 things go wrong here.

• Deliverability dip from a send spike: 1-month 0.4% reply-rate drag → fewer meetings/orders ≈ $3,700 margin loss

• Over-discounting on 0.2% of emails (52 incidents × $150 margin hit) ≈ $7,800

• Oversight/QC: sample 10% at 30s each ≈ $1,300

Net for month one

• Upside ≈ +$6,000

• New costs ≈ $12,800

• Net ≈ −$6,800

And this is what we mean when we say ''pricing the downside''.

Most founders are so stuck in future ROI projections that they forgot to factor in the downside if something goes wrong.

So, how do you engineer out that downside? How do you go from 20k in unforeseen damages to positive returns and profit right away?

Well, you Price the downside first.

So Here's the playbook for doing just that without burning your cash.

The playbook: deploy AI without lighting money on fire

1. Start in low-regret zones

• Use cases: internal drafts, support macros, subject line variations, call summaries, routing/triage.

• Keep humans over pricing, claims, segmentation, and anything irreversible.

2. Instrument everything

• Track on a weekly or monthly basis the replies (positive/neutral/negative), meetings booked, contribution margin per order

• Unsubscribes, spam complaints, domain reputation signals

• Discount usage and exceptions

• Rework time (“the time your team spends to clean up AI mess”) and reopened tickets
  Rule: if any critical metric dips, slow or stop and fix before scaling.

3. Human-in-the-loop with clear thresholds

• AI proposes, humans approve priced offers, large sends, and sensitive segments.

• Confidence thresholds: auto-send only for low-risk actions with high confidence.

4. Guardrails that protect margin

• Allowlists/denylists: use only pre-approved claims and offer language.

• Retrieval hygiene: strip untrusted external text to avoid prompt injection.

• Rate limits: warm domains, stagger volume, and cap daily sends.

• Safety checks: PII scans, tone checks, forbidden-claim filters.

5. Continuous evaluation (no set-and-forget)

• A/B test AI vs a human control every week.

• Use cost-weighted evaluation: penalize wrong discounts far more than missed personalization.

• Roll back on drift, schedule regular evaluations.

6. Budget guardrails (token hygiene)

• Short prompts. Reuse variables. Don’t paste your brand bible every time you want the LLM model to do something for you (i,e, write me a follow-up email to x).

• Cache common fragments (intros, product blurbs).

• Right-size models: small/fast for summaries, larger for complex reasoning.

• Review token and seat spend weekly, renegotiate quarterly.

7. Data governance (protect trust)

• Redact PII before sending to external services.

• Use enterprise endpoints that don’t train on your data.

• Train teams on what not to paste into the LLMs and keep audit logs.

• Reference: NIST AI RMF, OWASP Top 10 for LLM apps. its practical, founder-friendly guardrails.

8. Vendor due diligence (with teeth)

• Must-haves: SSO, SOC 2/ISO attestation, data residency options, retention controls, training opt-out.

• Product truths: exportability, model choice, off-switch, incident response.

• Contract terms: quality SLOs, data ownership, portability, 30-day termination for underperformance.

Red flags your AI is burning cash

Check the following notes if you want to see the impact of any LLM you deployed:

• Reply rate flat or down while volume is up

• Unsubscribes and spam complaints climbing

• Discounts/order rising but contribution margin falling

• Top performers spend more time QA’ing AI than selling

• “We can’t explain why it sent that” type of chat across your teams.

• Prompts getting longer every week to patch edge cases

• Teams quietly using different tools to “work around” the process

Start this week: a simple rollout plan

• Pick one SKU/collection and one retailer segment.

• Create three approved blocks: brand story, product proof, no-price offer.

• Let AI generate intros and transitions only, humans assemble and approve.

• Send conservatively. Track replies, meetings, unsubscribes, and margin/order.

• If AI beats the human control two weeks in a row, expand by 20%. If not, pause and fix.

Key takeaway

The biggest cost of bad AI isn’t the software, it’s the mess. Price the downside first, keep humans over costly decisions, and scale only what’s boringly reliable.

That’s how you turn AI from a hype expense into a margin engine for your ecommerce brand.