Field Notes: Three AI Projects I'd Refuse to Take in 2026

Table of Contents

• Introduction: Why Some AI Projects Are Doomed Before They Start
• The Three AI Engagements I’d Walk Away From in 2026
  • 1. The Blind Bolt-On: “Just Add AI”
  • 2. Compliance Theater Disguised as Innovation
  • 3. The Over-Engineered Agentic Swarm That Solves Nothing
• Spotting the Red Flags Early
• What to Build Instead: AI That Actually Moves the Needle
• How PADISO Evaluates AI Engagements
• Conclusion and Next Steps

Introduction: Why Some AI Projects Are Doomed Before They Start

AI is breaking things—in the best possible way. But after a decade of building technology for mid-market brands, scale-ups, and private-equity portfolios, I’ve learned that not every AI project deserves a green light. In fact, the most dangerous ones often look shiny on the surface: a charismatic founder with a “vision,” a board mandate to “do AI,” or an RFP from a PE firm itching to show portfolio uptick. As the founder of PADISO and a fractional CTO who’s been in the trenches with operators across the US, Canada, and Australia, I’ve seen patterns. I’m writing this field note to share three AI engagements I’d refuse in 2026—and, more importantly, the reasoning you can use to avoid them. This isn’t a theoretical exercise. I’ve said “no” to these projects more times than I’ve said “yes,” and the data bears out: companies that skip the strategic footing waste hundreds of thousands on shelfware, kill momentum, and erode trust.

By the end of this piece, you’ll have a clear framework for spotting the time bombs before they detonate. You’ll also see what we build instead—the kind of AI that has helped 50+ businesses generate $100M+ in revenue through strategic AI implementation and technology leadership. Let’s dive in.

The Three AI Engagements I’d Walk Away From in 2026

1. The Blind Bolt-On: “Just Add AI”

Picture this: A mid-market hardware manufacturer calls, eager to “infuse AI” into its 20-year-old ERP. The board read a Gartner report and now insists on a generative layer that will “predict procurement needs” and “automate vendor negotiation.” The problem? Nobody has defined the business outcome, mapped the data flow, or even asked frontline buyers what they need. This is the blind bolt-on, and it’s the fastest way to a $200K write-off.

Blind bolt-ons share a telltale DNA: a technology-first approach with no measurable revenue or EBITDA target. The assumption is that AI, by its mere presence, will elevate the product or process. That’s wrong. Even with frontier models like Claude Opus 4.8, Sonnet 4.6, or Haiku 4.5, the model itself is only a tiny fraction of a successful system. The heavy lifting is in data integration, change management, and continuous evaluation—none of which appears in a bolt-on budget.

I’ve seen this play out when firms attempt to staple a chatbot front-end onto a legacy ERP, only to discover that the underlying data is siloed, inconsistent, and lacks the semantic context an LLM needs. The result is a hallucination-prone demo that impresses no one. Worse, it creates technical debt that an incoming engineering team must unwind. At PADISO, we’ve built a discipline around Venture Architecture & Transformation specifically to avoid this trap. Before a single model is fine-tuned, we force the question: “What business metric must move, and by how much, to justify the project?” If the answer is fuzzy, the engagement is dead on arrival.

For a more grounded starting point, take our AI Readiness Test. In under two minutes, it surfaces whether your organization has the data maturity, leadership alignment, and problem clarity to make an AI investment pay off. No test will guarantee success, but it’s a cheap filter for the inbound requests that should never become a statement of work.

2. Compliance Theater Disguised as Innovation

Some of the scariest projects come dressed in a three-piece suit. A holding company needs to show ESG or technology auditors that it’s “embraced AI,” so it commissions a project that checks an ethical-AI box without delivering real capability. The deliverable is often a glossy report, a lightweight bias-detection wrapper around a non-critical process, or a token “AI governance framework” that gathers dust on SharePoint. I call this compliance theater, and I won’t touch it.

In 2026, with the EU AI Act imposing strict requirements on high-risk systems and growing enforcement around algorithmic transparency, the pressure to look compliant is real. But window-dressing fails. Regulators and sophisticated customers can smell a fabrication from a mile away. The International AI Safety Report 2026 underscores that AI-generated content can already influence beliefs at scale, and laboratory studies show measurable persuasion effects. If your governance program doesn’t address that head-on, it’s a mirage.

True compliance—what frameworks such as UNESCO’s Recommendation on the Ethics of Artificial Intelligence advocate—requires human oversight, ethical review boards, and AI ethics by design. At PADISO, we don’t perform theater. Our Security Audit (SOC 2 / ISO 27001) service, powered by Vanta, brings organizations to genuine audit-readiness, not a rubber stamp. And when we layer on AI Strategy & Readiness, we embed bias testing, explainability, and ongoing model monitoring into the operational fabric. That’s how you build trust with auditors, customers, and private-equity operating partners—not by slapping a fairness filter onto a high-risk HR model and calling it a day.

If your board is asking for “AI governance” in a vacuum, redirect the conversation. The Top 8 AI Ethics & Governance Considerations for Project Managers in 2026 provide a practical starting point: algorithmic bias, data privacy, transparency, and accountability must be operational, not decorative. For enterprises navigating these waters, the Ethical AI Implementation: 2026 Compliance Strategy & Risk Guide offers a step-by-step path to aligning with the EU AI Act, including risk classification and governance pillars. Use such resources to build muscle, then call us when you’re ready for real implementation.

3. The Over-Engineered Agentic Swarm That Solves Nothing

Every quarter, a new flavor of over-engineering appears. Lately, it’s the “agentic swarm.” A team gets captivated by the potential of multi-agent frameworks—orchestrating a dozen autonomous agents to handle expense reporting—and spends six months building an intricate system that a single GPT-5.6 (Sol) or Claude Opus call with a well-crafted prompt could have handled. The allure is understandable: models like Fable 5 demonstrate striking planning capabilities, and the open-weight ecosystem keeps marching forward. But complexity for its own sake is a luxury no mid-market firm can afford.

I saw a Series A startup in New York design a seven-agent system to parse and categorize customer feedback. The agents passed JSON blobs among themselves, each introducing latency, prompt token cost, and a failure point. The total monthly inference bill hit five figures before anyone asked, “Does this move retention?” It didn’t. The same outcome could have been achieved with a single Sonnet 4.6 call and a retrospective dashboard—built in a week, not six months—at a fraction of the cost.

The root cause is often an engineering team that’s been told to “go deep on agents” without a business-side sparring partner. That’s where a fractional CTO acts as the reality check. Our Platform Design & Engineering practice starts with the simplest viable architecture: one prompt, one LLM, one feedback loop. Only when performance bottlenecks or reliability demands require it do we introduce agentic decomposition, and even then, we stress-test the economics.

If you’re tempted to build a swarm, ask yourself: “What single metric am I improving, and what’s the simpler alternative?” Often, a dash of AI & Agents Automation with a single well-orchestrated agent delivers 90% of the value at 10% of the complexity. The rest is noise—and an invitation to technical debt that a future acquirer will penalize during diligence.

Spotting the Red Flags Early

After participating in dozens of AI evaluations for PE roll-ups and mid-market boards, I’ve distilled a short checklist that I run mentally before every call. These are the patterns that, in combination, almost always spell trouble:

• The business case is described in adjectives, not numbers (“more intelligent,” “better experience”).
• The project scope includes a technology shopping list before the problem is defined.
• No one can name the end-user persona or their job-to-be-done.
• The sponsor treats AI as a cost center initiative rather than a growth lever.
• The timeline is driven by a board meeting, not by technical readiness.
• There’s zero mention of data engineering, evaluation datasets, or drift monitoring.

Any single flag might be manageable. When three or more appear together, I decline. To help teams standardize these gut checks, I often guide them through a decision tree similar to the one below. Use it the next time you face an internal pitch:

graph TD
    A[AI Project Pitch] --> B{Clear Business Outcome?}
    B -- No --> C[Red Flag: Refuse or Pivot]
    B -- Yes --> D{Non-AI Solution Exists?}
    D -- Yes --> E[Implement Simpler Path]
    D -- No --> F{AI Maturity Sufficient?}
    F -- No --> G[Revisit Later or Pilot]
    F -- Yes --> H{In-House Expertise Available?}
    H -- No --> I[Engage Fractional CTO/Partner]
    H -- Yes --> J[Proceed with Guardrails]
    C --> K[Document Learnings]
    I --> J

The decision steps PADISO applies before green-lighting any AI engagement. If a pitch can’t clear the first two nodes, it’s almost never worth pursuing.

What to Build Instead: AI That Actually Moves the Needle

When we say “no” to the foregoing, what gets a “yes”? At PADISO, we look for projects that tie directly to a hard business driver: revenue growth, gross margin expansion, or a measurable step-level reduction in a controllable cost. Here are three archetypes that consistently deliver ROI for our clients:

1. Process Orchestration with Clear Unit-Cost Metrics. Instead of layering AI on everything, identify the manual workflow that consumes the most labor hours. For a logistics operator in Brisbane, for example, our team automated the freight-booking reconciliation that was eating 400 person-hours a month. We combined an extraction model (Haiku 4.5 for speed) with a deterministic business-rules engine. The project paid itself back in 11 weeks. That’s the kind of outcome our Platform Development in Brisbane engagements target—not a science experiment.

2. Compliance-Enabled Revenue Streams. When a fintech in Sydney needed to launch a new product that required APRA CPS 234 and ASIC RG 271 compliance, they didn’t go for theater. They needed a proper AI risk assessment, bias audit, and model card documentation embedded in the platform. Our AI for Financial Services team delivered a system that passed the regulatory review and let the firm capture a new market segment. Compliance wasn’t the goal—it was the enabler for a revenue line.

3. Turnkey Data Platforms for Portfolio Companies. Private-equity firms often ask us to stand up a shared analytics layer across a roll-up. This isn’t flashy AI, but it’s foundational. By deploying Apache Superset with ClickHouse on AWS—through our Platform Development in the United States practice—we give an operating partner a single pane of glass across three acquired businesses. That visibility drives EBITDA lift, full stop. Later, we layer in predictive models once the data quality is hardened.

These aren’t hypotheticals. Our Case Studies page details real engagements, from a health-tech scale-up that halved its cloud spend to a PE portfolio achieving SOC 2 readiness in under 90 days. The common thread? A tight coupling between AI investment and a metric the CEO already cares about.

How PADISO Evaluates AI Engagements

I founded PADISO with a simple ethos: we only take on work where our C-suite-level partnership can drive an outcome that’s visible on a board slide. That means our initial call isn’t a discovery pitch—it’s a mutual qualification. We ask the hard questions early, and we’re transparent when we don’t see a fit.

Our engagement model varies by need. For private-equity firms pursuing roll-ups, we often step in as the fractional CTO for the platform, bringing a team to handle Platform Design & Engineering and AI Strategy & Readiness across the portfolio. For growth-stage startups, our Venture Studio & Co-Build arm embeds a technical co-founder for 12-18 months while shipping an initial product. And for mid-market enterprises with an existing engineering team, we run a targeted AI & Agents Automation sprint that proves value in 8-12 weeks.

Geographically, our footprint aligns with where the hardest problems live. We have boots on the ground in Sydney, Melbourne, New York, and San Francisco, with deep domain knowledge in financial services, logistics, and health. We also serve clients in Darwin for edge-compute and sovereign hosting, and the Gold Coast for tourism and SMB automation. Each region has its own regulatory and talent profile, but the principle is identical: no project begins without a clear, quantified line of sight to value.

Across all these engagements, the tech stack is a conversation, not a religion. We build heavily on AWS, Azure, and Google Cloud, leveraging their native AI services where it makes architectural sense. The model layer is chosen based on the task: Claude Opus 4.8 for complex reasoning, Sonnet 4.6 for high-volume enterprise workflows, Haiku 4.5 for real-time extraction, and open-weight alternatives when data sovereignty demands on-premise inference. Yes, the frontier models are impressive, but the differentiator is almost always the data pipeline, the evaluation harness, and the incentive structure the business wraps around the AI. Without those, even a GPT-5.6 Terra is just an expensive token generator.

Conclusion and Next Steps

I wrote this field note because I’m tired of cleaning up after projects that should never have been started. The blind bolt-on, the compliance theater, and the over-engineered agentic swarm are not just wastes of capital—they poison the organization’s appetite for the AI that could actually change its trajectory.

If you’re a CEO, private-equity operating partner, or head of engineering who wants to avoid these traps, here’s what I recommend:

1. Take ten minutes to do the hard upfront work. Our AI Readiness Test is a quick diagnostic that often exposes whether you have a real problem to solve or just FOMO.
2. Bring us your worst idea. Before you commit to a $200K sprint, book a one-hour session with our team. We’ll stress-test the concept against the decision framework above. If we see a better path, we’ll say so.
3. Join our upcoming AIR Bootcamps. We run executive-level sessions that teach how to spot value-destroying AI projects before they land—drawn from our About story and the $100M+ in revenue our clients have unlocked.

The AI landscape will keep accelerating, but the fundamentals of good business have not changed. Solve a real problem, measure the outcome, and don’t build complexity before it’s earned. That’s how we ship at PADISO, and it’s the filter I’ll keep applying to every project that comes my way.

Let’s talk about what you’re building—and what you should probably stop building.