Nine Systems That Break Before an AI Agent Goes Live | Alex Lieberman
Alex Lieberman Founder & Podcast Host AboutArticlesContact Get in touch
Home / Articles / AI AI
Nine Systems That Break Before an AI Agent Goes Live
Alex Lieberman October 5, 2026 · 4 min read
IN THIS ARTICLE
01 Scope Without Ambiguity Kills Most Projects Early 02 Bad Source Data Becomes Confident Bad Output 03 Why Evals Must Be Infrastructure, Not Afterthought 04 Deterministic Scaffolding Around the LLM 05 Least Privilege and Prompt Injection Defenses 06 Governance Gets Considered From the Jump 07 Step-Level Tracing of Every Decision 08 Embed Where People Already Work 09 Measure ROI Against the Baseline Set Before Launch 10 Raw Intelligence Won't Save a Broken Machine
I'm not at all surprised so many enterprise AI deployments fail. So many things have to go right for an agent to make it to production. If one cog in the machine breaks, timelines are pushed back in the best case scenario. In the worst case, leadership's trust is eroded. Until you've truly spent time in the guts of an enterprise, you don't realize just how hard meaningful change is and just how much has to happen beyond raw frontier intelligence.
Scope Without Ambiguity Kills Most Projects Early
One workflow with a named business owner, a baseline metric, and a dollar outcome. That's where you start. Autonomy has to be matched to blast radius: suggest first, then act with approval, then act autonomously. You earn each step with data. Kill criteria get defined up front. And there has to be clarity around the current way of working and what the desired future state looks like.
Bad Source Data Becomes Confident Bad Output
Data and system access have to be configured and trustworthy. That means read/write access to systems of record, legacy APIs, and unstructured sources. It means service identities, SSO, and permission models that work for a non-human actor. Data quality and freshness matter. Bad source data becomes confident bad output.
Why Evals Must Be Infrastructure, Not Afterthought
Evals must be a core part of infrastructure. Build a golden set from real historical cases, including edge cases and adversarial ones. Measure success against the human baseline, not against perfection. Run a regression suite on every prompt, model, or tool change.
Deterministic Scaffolding Around the LLM
Architecture and reliability have to be sound. Use deterministic scaffolding around the LLM. Use code for anything that can be code. Design tools well, with timeouts and fallbacks. Build model abstraction so you can swap providers. Set latency and cost budgets per task, plus rate-limit handling at scale.
Least Privilege and Prompt Injection Defenses
Security has to be buttoned up. That means least-privilege agent identity and scoped tool permissions. Prompt injection and exfiltration defenses, especially with untrusted inputs like emails, documents, and web content. Full audit trail, tenant and data isolation, residency, and vendor DPAs.
Governance Gets Considered From the Jump
Governance and risk have to be considered from the start. IT, legal, compliance, and finance sign-off. Regulatory fit with SOX, HIPAA, GDPR, EU AI Act, and industry-specific rules. Human-in-the-loop thresholds, a kill switch, and one accountable owner for agent behavior. If you're working through building an AI roadmap, this is where most teams underestimate the work.
Step-Level Tracing of Every Decision
Observability and operational rigor matter. That means step-level tracing of every decision and tool call. Monitoring for cost, quality drift, and failure clusters. Versioning and rollback for prompts, models, and tools. And a feedback loop from human corrections back into evals.
Embed Where People Already Work
The conditions for adoption have to be ripe. You need an executive sponsor with budget authority, plus the frontline process owners involved from the start. The agent should be embedded where people already work, whether that's Slack, Salesforce, or Outlook. Not a new destination. Clear escalation paths and honest handling of job-displacement anxiety. And incentives aligned so the team benefits from the agent working.
Measure ROI Against the Baseline Set Before Launch
Economics have to be underwritten and measured. That means fully loaded cost per completed task, covering inference, infrastructure, review labor, and maintenance, against the human cost. ROI measured on the baseline set before launch, not estimated after. And understanding whether it's part of an experimental, efficiency, innovation, or infrastructure budget. When you're ready to move forward, following a structured framework for turning business goals into AI pilots helps avoid the most common pitfalls.
Raw Intelligence Won't Save a Broken Machine
Nine systems have to work in concert before an AI agent reaches production. Scope, data access, evals, architecture, security, governance, observability, adoption conditions, and economics. Miss any one and you're pushing timelines or losing trust.
This is why enterprise AI feels so much harder than the demos suggest. The frontier model is the easy part. Everything around it is where deployments actually fail.
Start with one workflow, one owner, and one metric before you touch any of the rest.
WRITTEN BY
Alex Lieberman
Founder & Podcast Host
Alex Lieberman is the co-founder of Morning Brew, the business media company he led as CEO from 2016 to 2021 and grew to more than 4 million subscribers before its $75 million sale to Insider in 2020. He is now Managing Partner of Tenex, an AI engineering and transformation firm, and co-founder of storyarb and GrowthPair. He hosts The Founder's Journal podcast and was named to the Forbes 30 Under 30 Media list in 2018. He holds a BBA from the University of Michigan's Stephen M. Ross School of Business.
X
Instagram
LinkedIn
TikTok
← All articles
Share on X LinkedIn
Alex Lieberman
MEDIA & STARTUPS
Helping companies go from AI-absent to AI-native.
X Instagram LinkedIn
RECENT POSTS
How to Run a Five-Day AI Agent Sprint for Your Enterprise OCT 8, 2026 The Year After Selling Morning Brew Nearly Broke Me OCT 7, 2026 Nine Systems That Break Before an AI Agent Goes Live OCT 5, 2026 Six Steps to Turn a Business Goal Into an AI Pilot SEP 20, 2026
BEST POSTS
A 32-Year Overnight Success READ · 3 MIN What I Have Learned Building in AI READ · 3 MIN Don't Try This: Running 5K While Explaining How I Built and Sold Morning Brew READ · 4 MIN
PAGES
AboutArticlesContactSitemap
© 2026 Alex Lieberman New York City