Editor’s note: The following is a guest post by Zaré Baghdasarian, co-CEO at Amberd.ai.
A pattern is playing out in customer accounts across the channel right now that most partners will recognize.
An enterprise customer — let’s say a mid-market manufacturer or a regional financial services firm — asks a managed service provider or systems integrator to help them explore AI. Presales gets involved. A data analyst spends several weeks mapping data sources. A pilot gets scoped, usually around a Copilot layered onto an existing business platform or a generative AI interface bolted onto SharePoint. The demo looks impressive. The customer’s IT leadership nods enthusiastically.
And then nothing ships. The pilot lives in a slide deck. The engineering hours don’t convert into a packaged service. Six months later, the customer is asking to “explore AI” again.
This is the pilot trap, and it is quietly becoming one of the most expensive problems in the channel.
The problem is a structural mismatch between piloting AI and getting value from the technology.
Consider the channel’s history with a different technology: BI software. Selling a BI platform follows a well-understood playbook. You assess the environment, configure the software, train the users, and go live. The outcome is largely defined by the vendor’s feature set — Power BI, Tableau, Qlik, ThoughtSpot. The partner’s differentiation is in the speed and quality of the deployment. The channel has refined this model for decades, and it works because the product itself defines what’s possible.
AI is different. An AI solution that delivers genuine business value must be architected around a specific customer need, not a capability the vendor’s platform happens to offer. That requires structured discovery across multiple levels of the client organization. MSPs and SIs need to talk with stakeholders at varying levels of seniority to map existing workflows, identify friction points, and pinpoint where AI can impact business outcomes. When partners skip that work and lead with the tool instead of the decision, they end up building something technically functional that nobody uses.
The downstream consequence is real and quantifiable. When consultants and senior engineers spend weeks on a pilot that stalls, that capacity isn’t being used to build repeatable, billable service offerings.
The data readiness myth
One of the most reliable ways to predict whether a customer AI initiative will stall is to listen for a specific phrase in the early discovery calls: “We just need to get our data house in order first.”
Partners hear this constantly, and the instinct to validate it is understandable. Clean, well-governed data matters. But believing a customer must first build a perfect centralized data environment — a fully governed data lake, a unified semantic layer, a single source of truth across all systems — before any AI deployment can proceed is one of the most effective ways to delay an AI initiative indefinitely.
Experienced applied AI teams encounter a different situation in the field. Most enterprise customers already have the data required to support a meaningful AI application. It exists in ERP transaction logs, CRM pipeline records, financial spreadsheets, company policies, business objectives, mission statements, institutional knowledge and industry reference data. It’s not perfectly clean. It’s not centralized. But it is sufficient to support a model working against a specific, bounded decision.
Data readiness work happens in parallel with a defined business objective and is a prerequisite to finding one.
Copilots are not decision support
The BI model — resell, implement, train and repeat — is how most channel partners think about AI deployments. The instinct is to layer a generative AI interface onto an existing BI stack: a copilot that answers questions in natural language instead of requiring a query. The user interface may look better, but the decision architecture remains unchanged and still rests on the same incomplete information. For many resellers and integrators, that means treating AI as a UX add-on to existing BI deals rather than as a new layer of decision infrastructure they can design, implement and manage.
Decision-grade AI is different in kind, not just degree. Consider what’s happening right now to midmarket logistics and customs brokerage firms. They are scrambling to file tariff refund claims through a process that presents a serious data challenge. It requires detailed entry-level data, strict eligibility rules and phased deadlines that evolve week to week.
One customs brokerage we worked with, R.L. Jones, a major U.S.–Mexico border firm, faced this exact pattern: hundreds of affected customers, tens of thousands of historical entries, and no system capable of handling complex reimbursement logic at scale. The engagement started with structured discovery sessions across operations, compliance and leadership to map workflows, identify the precise compliance decisions that had to be made and define what “production-ready” meant in practice.
Only after those decision parameters were clear did the technical work begin. A private, in-tenant LLM platform was deployed in the brokerage’s cloud environment, with data ingested on a controlled schedule so the firm retained data ownership and regulatory posture. Operating directly against that in-place data, the system handled tariff logic, refund calculations and reconciliation work that analysts had been doing manually, turning hours of per-customer spreadsheet work into minutes, with a defensible, line-by-line audit trail that could scale across hundreds of client accounts without adding analyst headcount.
For channel partners, the lesson is clear: decision-grade AI starts with decision design and domain workflows, then moves to secure, embedded AI infrastructure that can be offered as an ongoing decision-support service rather than a one-off project.
3 key moves
The practical question for IT service providers is how to sort real AI opportunities from conversations that will end in another stalled pilot. The simplest test you can run in any executive strategy session is to ask the customer to name a decision they make every week or month without complete information. If they can name one, there is a viable AI engagement; if all they can articulate is a desire to “have better AI” or “see what’s possible,” you are looking at a science project, not a production deployment.
The vetting process comes down to three moves:
- Redefine the motion: No out-of-the-box solution fits every customer, decision environment or regulatory context. AI that creates business value has to be engineered around a specific customer’s objectives, data environment and decision workflows, with the platform as an ingredient rather than the answer, and the partner operating as an architect rather than just an installer.
- Build an applied AI function: The partners who are converting AI conversations into production deployments have an applied AI Engineering capability inside their practice that goes beyond a pre-sales or vendor-funded PoC team. Their mandate is to sit with customers, map operational workflows, identify the decisions where AI can change the outcome and architect solutions using whatever mix of models, data pipelines and integration layers the problem requires.
- Make the structural shift: Channel partners who want durable AI practices need to invest in people who understand how work gets done. They should replace conversations about exploring AI with an evaluation of what decisions a client is routinely getting wrong or making too slowly,
Real margin in AI comes from the engineering judgment required to make AI work for a specific customer in the real world.
