From Punch Cards to Agentic AI: Christopher Carter on Modernizing SAP (Without Throwing Away the Past)

Guest: Christopher Carter, Founder & CEO, Approyo
Show: Interchain Me – Old Tech
Listen to the Podcast

Chris Carter straddles two worlds: green-screen terminals and generative AI. He grew up on punch cards, tapes, and AS/400s—and today helps large SAP shops wrap modern AI around decades of core processes without breaking what already works. His playbook favors small language models (SLMs) grounded in a company’s single source of truth over sprawling internet-trained systems.

Why this matters now

• Legacy isn’t “old”—it’s valuable. Mainframes, AS/400s, and ERP cores (SAP) still run mission-critical finance, logistics, and manufacturing. Rewrites are risky; augmentation is pragmatic.

• Enterprise data is the new arms race. Your private contracts, parts catalogs, service tickets, and telemetry—not the open web—determine your AI advantage.

• Right-sized AI wins. In an SAP estate with one canonical database, SLMs are faster, cheaper, safer, and often better than generic LLMs.

What Chris is building

• Overwatch (internal tool). A layer that blends AI + predictive analytics to monitor SAP estates, surface Z-code and transition hotspots, and de-risk cloud migrations.

• SLM-first patterns for SAP. Retrieval over your enterprise DB; code assist with validation; guardrailed agents to automate checks, drafts, and routine triage.

• Cloud migration accelerators. Move x86-era SAP to the cloud while wiring in AI for testing, documentation, and change impact analysis.

“I don’t need to boil the ocean. My customer has one enterprise database. A small language model on that beats a giant model on the internet.”

Efficiency and effectiveness (not job loss)

Carter’s lens is pragmatic: AI won’t gut experienced SAP teams; it will amplify them. The near-term disruption is at the junior tier—newcomers must level up with AI copilots, code validation, and data-aware workflows. Veterans who pair domain expertise with AI become irreplaceable.

• SAP talent skew: ~57% are 40+; only ~13% are 20–30.

• Translation: teach AI to the next cohort, and teach the next cohort to use AI.

The near future: speed, then a plateau

Expect a burst—agentic AI, GenAI, SLMs—followed by a capacity pause:

• Power & grids aren’t ready everywhere.

• Data center supply is constrained.

• Skilled staff shortages will bottleneck delivery.

This is a good forcing function: fewer giant models scraping everything, more tailored systems over owned data.

Data is the moat

Law firms with 200 years of briefs, cities with centuries of maintenance logs, hospitals with longitudinal cohorts—these are AI gold mines. Carter sees customers:

• OCR-ing archives at scale into blob storage.

• Building private RAG over curated corpora.

• Shifting from “web search” to answer engines that return decisions, not links.

“Tesla isn’t just a car company. It’s a data company. That’s the real game.”

Agentic AI: power and guardrails

Carter embraces agents that act—drafting tickets, reconciling data, scheduling, even kicking off jobs—with permissions today, possibly fewer prompts tomorrow. The mandate:

• Human-in-the-loop for sensitive actions.

• Policy, identity, and audit from day one.

• Least-privilege access to systems and data.

Risks he’s candid about

• Fraud & deepfakes: synthetic IDs, e-sign misuse, prescription mills.

• Information authenticity: photos, voices, and “news” are trivially faked.

• Over-attachment to AIs: human tendency to anthropomorphize tools—design with ethics and transparency.

Practical wins already here

• Search becomes answers: “Hey Siri/Copilot—give me the recipe,” not ten links.

• Autonomy on the ground: think John Deere’s GPS-guided tractors freeing farmers for higher-value work.

• Legacy UX upgrades: conversational wrappers over SAP transactions (without replatforming).

How to start (Carter’s playbook)

1. Inventory your crown jewels. Systems, schemas, and the real “single source of truth.”

2. Pick 2–3 high-value use cases. Change impact, code/docs generation, finance close support, or supply chain exceptions.

3. Favor SLMs + RAG over your data. Faster, cheaper, safer.

4. Stand up a governed data layer. Access control, lineage, PII handling, redaction.

5. Pilot with a human in the loop. Automate drafts and checks; keep approvals human.

6. Measure what matters. Cycle time, error rate, rework, and user NPS—not model metrics.

7. Train the people, not just the models. Pair domain experts with AI fluency.

Pull quotes

“Legacy doesn’t need a eulogy—it needs interfaces.”
“A well-tuned small model on your ERP beats a giant model on everyone’s internet.”
“Data will bring you to the top—or to your knees.”

My take

Chris is the bridge many enterprises need: respect for systems that already work, coupled with a clear-eyed plan to layer AI responsibly. If your SAP runs the business, you don’t need a moonshot—you need a measured path to better decisions, faster cycles, and happier users.

Credits

Recorded for Interchain Me. Host: Roberto Capodieci. Guest: Christopher Carter (Approyo—fun fact: the name riffed off pitcher Bronson Arroyo).

Connect

Want to wrap AI around SAP or other legacy cores? Chris welcomes conversations: Approyo (approyo.com) and Christopher Carter on LinkedIn.