Perspectives

Where platform decisions get made before the RFP.

Practitioner perspectives on AI, data, and the enterprise platforms that run on them — written by the people who deliver the work, not a content team.

Content pillars

Four lenses on the enterprise stack

Filter by pillar to read what practitioners are seeing across live engagements — before the RFP is written.

Get one perspective a month. No sales pitch — just what we're seeing across live engagements.

Subscribe to ACSA Insights

Ready to publish

Agentic Enterprise & AI ROI

From AI Pilots to P&L Impact

A pragmatic operating model for scaling AI beyond the pilot — who owns it, how it's funded, how it's measured.

The pilot trap

Most enterprises don't have an AI pilot problem. They have a second-pilot problem: the first proof of concept works, gets a warm reception in a steering committee, and a second team requests budget to build something adjacent — with its own data pipeline, its own vendor, its own definition of success. Eighteen months later, an organization can have a dozen AI initiatives and no shared operating model connecting any of them to profit and loss. The fix isn't more governance. It's fewer, better-owned initiatives with a funding model that matches how the value actually shows up.

Who owns it

Assign AI initiatives to the P&L line they're meant to move, not to a central AI team. A model that automates warranty claims adjudication belongs to the service organization's cost line, not to a shared "AI Center of Excellence" with no accountability for the outcome. Central teams still matter — reusable infrastructure, model governance, a shared knowledge graph — but ownership of the business case sits with whoever answers for the number it's supposed to move.

How it's funded

Fund AI in the same increments the business already funds anything else: a business case with a payback period, reviewed at the same cadence as capital requests. The pilots that stall are almost always the ones funded outside that discipline — through innovation budgets with no expectation of return, which means no one is accountable when the number doesn't move.

How it's measured

Three metrics, not thirty: time-to-value from deployment to first measurable business outcome, cost-to-serve on the process the model touches, and adoption — the share of eligible transactions actually flowing through the new workflow rather than being routed around it. Adoption is the metric most programs skip, and the one that kills them quietest; a model with 92% accuracy that a third of the workforce quietly avoids is not a success.

What changes when this works

Organizations that get this right stop measuring "AI maturity" and start measuring which cost lines moved. That's a smaller, less exciting story than "we deployed twelve agents this year" — and it's the only version of the story that survives a budget review.

Platform Modernization

The Clean-Core Advantage

Why extension discipline in S/4HANA determines whether your next upgrade takes months or years.

The upgrade that never gets easier

Every enterprise that customized ECC the same way learns the same lesson at RISE with SAP time: the upgrade isn't hard because S/4HANA is hard. It's hard because a decade of custom code, modified standard objects, and side-door integrations turned the core into something no vendor patch was ever tested against. Clean core isn't a purity exercise — it's the difference between an 18-month migration and a 6-month one.

What clean core actually means

Clean core means every extension lives outside the core: in the SAP BTP extension layer, via released APIs and CDS views, never through direct table access or core object modification. It's a discipline, not a product — SAP will sell you the tools, but no tool enforces the discipline for you.

The three-question test

Before approving any customization request, three questions decide whether it belongs in the core or the extension layer: Does this run on every transaction, or only for an exception? Would we lose this capability entirely if we didn't build it, or is it a convenience? Can it be built on a released, upgrade-stable API? If the answer to the third question is no, the request needs an architecture review before it needs a developer.

The part nobody budgets for

Clean-core discipline has a real cost: it's slower to ship the first version of anything, because the sanctioned path is rarely the fastest path. Enterprises that skip this step save three weeks on the first release and lose eight months on the next upgrade. The math only works if leadership protects the discipline when a business stakeholder is impatient for a shortcut — which is the actual failure point in most clean-core programs, not the technology.

What good looks like

An enterprise doing this well can name every extension to its core, knows which SAP release cycle each one depends on, and treats its next S/4HANA or RISE upgrade as a scheduled event rather than a program. That's the real advantage — not a cleaner architecture diagram, but a modernization roadmap the business can put a date on.

Industry Playbooks

Warranty as a Revenue Line

How machinery OEMs turn field data and warranty automation into recurring aftersales revenue.

The installed base is the business

For a heavy equipment or industrial machinery OEM, the sale is the smallest transaction in a twenty-year relationship. Parts, service contracts and warranty claims routinely generate more lifetime revenue than the original unit — yet most OEMs run this side of the business on the systems with the least investment: spreadsheets for claims, phone calls for parts pricing, and a warranty process built to catch fraud rather than generate revenue.

Where the revenue actually is

Three places aftersales revenue hides in plain sight: proactive service triggered by telemetry before a failure happens, not after; parts quoting fast enough that a dealer doesn't source the part from a third party while waiting; and warranty adjudication fast enough that dealers don't quietly absorb costs rather than file a claim through a slow process.

Warranty as a discipline, not a cost center

Treat warranty automation as a revenue function, not a claims-processing function. Every claim that resolves in hours instead of weeks is a dealer relationship that trusts the OEM's aftersales program enough to route more business through official channels instead of gray-market parts. Every fault code connected automatically to a parts recommendation is a service call converted to a sale instead of a support cost.

What the technology needs to do

None of this works without one thing: a single knowledge graph connecting product configuration, service history and warranty terms, so a service quote, a warranty adjudication and a parts recommendation all draw from the same source of truth instead of three disconnected systems that disagree with each other. OEMs that consolidate onto one platform for CPQ, warranty and field service typically find the biggest win isn't speed — it's that dealers stop routing around the OEM's own systems.

The number that matters

Track one metric above all others: the share of the installed base under an active service or warranty relationship. Every equipment manufacturer already knows how many units it sold. Few know how many of those units they're still in a revenue relationship with — and that gap is usually where the next several points of margin are sitting.

Data & AI Governance

Lineage Before Language Models

Why data lineage, not model choice, is the real blocker to trustworthy enterprise AI.

The wrong question

Most enterprise AI governance conversations start with "which model" and "how do we prevent hallucination." Both are the wrong first question. An agent that retrieves the wrong version of a pricing table and states it confidently isn't hallucinating — it's accurately reporting bad data. No model choice fixes a lineage problem.

What lineage actually means here

Data lineage, in the context of agentic AI, means being able to answer three questions for any output an agent produces: which system did this fact come from, when was it last updated, and who is accountable for its accuracy. Most enterprises can answer this for financial reporting, because auditors require it. Almost none can answer it for the data an AI agent pulls into a customer-facing answer.

Why this gets skipped

Lineage work is unglamorous and doesn't demo well. A retrieval-augmented generation pilot can look impressive in a steering committee with three weeks of clean, hand-picked source documents. It falls apart in production against the actual mess of duplicate folders, three versions of the same policy document, and a data warehouse where "customer segment" means something different in two divisions.

The governance model that scales

Model risk management borrowed from financial services is a better governance template for agentic AI than most AI-specific frameworks being pitched today: a model inventory, a defined owner per model, a documented data source per output, and a review cadence tied to how much authority the model has — an internal drafting assistant needs less oversight than an agent that can adjudicate a warranty claim or issue a customer-facing quote.

Where to start

Before scaling any agentic AI initiative past a pilot, build the lineage map for the three data sources it depends on most. If that map can't be built in under two weeks, the initiative isn't ready to scale — it's ready for a data cleanup project that should happen first, and will make every AI initiative after it faster to trust.

In the pipeline

Next four pieces

What’s coming next from ACSA practitioners — subscribe to get them when they publish.

Platform Modernization

The Lakehouse Decision

A consolidation framework for moving off legacy warehouses onto Databricks without a two-year migration tax.

Agentic Enterprise & AI ROI

The Agent That Didn't Ship

What derails agentic AI projects between demo and production, and the three gates that catch it early.

Industry Playbooks

Clienteling, Rebuilt for Retail

What a modern Salesforce-led clienteling stack looks like for CPG and specialty retail.

Data & AI Governance

Model Risk Isn't Just a Bank Problem

A practical model-risk checklist for any enterprise putting agentic AI into a regulated workflow.

Talk to an ACSA practitioner