COO anti-patterns with AI — automating broken processes

The single biggest mistake I see SMB COOs make with AI is automating a broken process at full speed. The tool works. The process doesn't. The output looks productive and is actively destructive.

Why COO mistakes hit hardest

Operations is where AI value is supposed to land. If the COO gets it wrong, the CFO sees it in the cost line, the CMO sees it in customer complaints, and the CEO sees it in board questions. Stanford's 51-deployment study showed escalation-routing yields ~71% productivity gain vs ~30% for approval-routing — the difference between those two outcomes is almost always a COO design decision.

Anti-pattern 1 — Automating a broken process

What it looks like: Your invoice approval has six steps, three of which are duplicated work between AP and the requesting manager. The COO automates all six with an agent. Now the duplication runs at machine speed and errors compound.

Why it happens: AI procurement timelines pressure the COO to "show automation". Fixing the process first feels slower; it doesn't ship a vendor logo.

Visible damage: Output volume up, error rate up faster. The Microsoft Copilot rollout (>80% drop in 3 weeks) is partly this story — automating workflows that were never tightened first. The AI Tax of ~37% rework eats most of the saved time, sometimes all of it.

What to do instead: Map the process by hours-spent and step count first. Eliminate redundant steps before introducing AI. The classic rule: "don't automate what you should eliminate."

Definition: Process amplification — when AI executes an existing workflow faster, including its bugs. The faster the tool, the bigger the amplification.

Anti-pattern 2 — RPA-thinking applied to LLM agents

What it looks like: The COO treats an LLM agent like an RPA bot — deterministic, scripted, brittle. They expect zero variance and panic when the model deviates. Or worse, they let it deviate without guardrails because "it's smart."

Why it happens: Many COOs come from RPA-era automation thinking, where the bot does exactly what you scripted, and you debug by re-scripting. LLMs are probabilistic; the mindset doesn't transfer cleanly.

Visible damage: Either the agent is wrapped in so many rules it loses its advantage, or it runs unchecked and ships hallucinated outputs to customers. Klarna's 2025 walkback is the public version of the second failure mode.

What to do instead: AI agents need probabilistic-thinking guardrails — confidence thresholds, escalation rules, sample-based human review, not deterministic scripts. The COO who gets this right designs the escalation path before the happy path.

Anti-pattern 3 — Ignoring shadow AI

What it looks like: The COO is rolling out an enterprise AI platform. Meanwhile, half the ops team is using free ChatGPT on their phones, with customer data, off-policy, no logging.

Why it happens: Shadow AI is invisible by definition. Stanford's "77% rule" — most AI work in orgs is unofficial — applies hardest to ops, where individual contributors are creative under pressure.

Visible damage: Compliance exposure (46% of employees have uploaded confidential data to public AI tools), inconsistent quality, and a sanctioned tool with low adoption because the unsanctioned one already works. EU AI Act fines reach €35M or 7% global turnover; Replika's €5M Italy fine and Clearview's €30.5M Dutch fine are public reminders.

What to do instead: Survey before procurement. Ask every ops team: "What AI tools do you use right now, sanctioned or not?" The honest answers redesign your roadmap. Shadow usage is a free signal — the things people already use are the things that work.

Definition: Shadow AI — unofficial AI tool usage by employees, often using free consumer products with company data, invisible to IT and procurement.

Anti-pattern 4 — Single-vendor stack lock-in

What it looks like: The COO buys one platform that promises "the entire AI ops stack" — agents, RPA, analytics, document processing. Twelve months later, the platform handles 60% of needs adequately and 40% poorly, but the contract makes adding a second tool look like failure.

Why it happens: Single-vendor stacks are operationally simpler to procure. One contract, one integration, one neck to choke.

Visible damage: Your worst-performing 40% of workflows stay broken because moving them off the platform means admitting the platform was the wrong choice. Builder.ai's $1.3B collapse echoes this in extreme — over-promising platforms that under-deliver in long tails.

What to do instead: A best-of-breed approach with clear integration contracts (APIs, webhooks, shared identity) wins over 24 months. Pick the platform that is best at your three highest-volume workflows. Add specialists for the rest. Don't be afraid of three vendors.

Anti-pattern 5 — Skipping the diagnostic phase

What it looks like: The COO walks into an AI tool decision with a gut sense of which workflows hurt most. They're 60% right and 40% wrong. The 40% wrong is invisible until the tool is deployed and the wrong workflow gets the budget.

Why it happens: Diagnostics feel like delay. The COO's job is to ship operational results, not commission internal studies.

Visible damage: Three to six months of wrong-priority work. MIT's 95% pilot-fail rate clusters here — pilots that succeeded technically but on workflows that didn't matter at scale.

What to do instead: A 7-day Plan vs Fact vs Gap diagnostic produces a ranked list of where the actual hours go. The list is rarely the one the COO expected. Skip this and you're betting your AI roadmap on intuition.

Manager scan (2-minute digest example)

• Plan: Roll out invoice automation to AP team. Target: 60% time saved.
• Fact: Tool live week 3. AP queue down 22%. Error rate up 11%.
• Gap: Process never debugged before automation. Three duplicate-approval steps still running.
• Plan: Customer onboarding agent live for ops team.
• Fact: Adoption 40%. Avg handle time down 18%.
• Gap: No escalation path for high-value accounts; two churned in week 6.
• Plan: AI document classifier for inbound contracts.
• Fact: 73% accuracy on training set; 51% in production.
• Gap: Production data shape differs from training. No human-review gate above 30% confidence.
• Plan: Standardise team on enterprise GPT.
• Fact: 12 of 40 logged in.
• Gap: 22 of 40 still using personal ChatGPT — shadow AI, customer data exposure.

Tool tip (AIAdvisoryBoard.me): AI Advisory Board's Plan → Fact → Gap engine is built for exactly this COO problem — you can see at a glance whether the process and the tool are aligned, or whether one is racing ahead of the other. It pulls from your real ops systems (tickets, invoices, CRM, project trackers) and produces a daily 2-minute morning digest. The diagnostic alone usually pays for itself within the first vendor decision it changes. See it: https://aiadvisoryboard.me/?lang=en

Micro-case (what changes after 7-14 days)

A 220-person logistics company had spent six months negotiating an enterprise AI ops platform contract — high six figures annually. The COO ran a 7-day diagnostic before signing. Two surprises emerged: the highest-volume workflow (delivery exception handling) was already 78% solved by a free tool the dispatch team had quietly adopted; the workflow the platform was supposed to fix (invoice routing) had three structural duplications that no AI tool would help with. Total budget reallocation: ~$400K saved on the platform that didn't get bought, ~30 person-hours/week saved by fixing the duplication.

Note on this case: This example is illustrative — based on typical patterns we observe with companies of 30-500 employees, not a single named client. Specific numbers are rounded approximations of common ranges, not guarantees.

Tool tip (AIAdvisoryBoard.me): Before automating any operational process, run a Plan → Fact → Gap diagnostic across the team for one week. The output: a ranked list of where hours actually go, where the process amplifies errors, and which workflows are already shadow-automated. AI Advisory Board does this without IT integration projects — start here: https://aiadvisoryboard.me/?lang=en. It's a fraction of the cost of one bad RPA contract.

FAQ

Q: Isn't process improvement the BPO/Lean team's job, not the COO's AI scope? In 30-500-person companies the COO usually owns both. Splitting "process" from "AI" creates the exact handoff gap that lets these anti-patterns thrive. Treat them as one workstream.

Q: How do I balance speed-to-deploy with diagnostic depth? A 7-day diagnostic is not a delay — it's a parallel track. Procurement conversations and integrations can run during it. The diagnostic just changes which contract you sign at the end of week one.

Q: What about RPA — should we still use it? Yes, for stable structured-data workflows. RPA + LLM agents are complementary: RPA handles the deterministic spine, agents handle the variance. The mistake is using one when you need the other.

Q: How do I surface shadow AI without scaring the team? Ask "what tools help you most" not "what unauthorised tools are you using". Frame it as procurement input. The honest answers will surface, and they're operational gold — every shadow tool that works is a sanctioned roadmap candidate.

Conclusion

The COO's AI job is not to ship the most automation. It's to ship the right automation in the right order on processes that are tight enough to handle machine speed. The five anti-patterns above all share the same fix: see the work clearly, in real time, before deciding what to automate.

If you want a system that surfaces the Plan → Fact → Gap automatically — every day, across operations — see how the 7-day diagnostic works: https://aiadvisoryboard.me/?lang=en