How it works

Context first. Agents second. People decide.

What sits underneath a Core Portage engagement: a recorded model of how your business runs, a controlled evidence layer, and agents that can read and propose but cannot change production.

The idea in one paragraph

An AI agent is only as good as the context it is given. Hand a capable model your SAP system with no understanding of your business and it will produce fluent, confident, wrong answers. So we build the context first: a recorded, reviewable model of how your company actually operates. Then agents work inside narrow, audited controls to analyze, map, draft and test. Senior SAP functional experts and agentic engineers review what the agents produce and make every decision that matters. Agents do the volume work. People keep the judgment and the authority.

Everything below follows from that. It is the same foundation whether the destination is Rescue, Switch or Freedom.

The semantic model

This is the core of our method. From configuration, custom code, usage statistics, interfaces, document flows, logs, runbooks, and agent-led interviews with your process owners, we build a recorded model of how the business really runs. It holds four kinds of thing:

  • Entities: customers, materials, ledgers, organizational units and the relationships between them.
  • Rules: pricing, approvals, tolerances, posting logic, and the conditions under which each applies.
  • Exceptions: the workarounds, manual steps and spreadsheet side-channels that never made it into a design document.
  • A patch panel: a mapping between how the current system does each thing and how any target would do it.

The patch panel is why one assessment supports three paths. Map to S/4HANA and you have a migration plan. Map to another ERP and you see exactly which of your processes change when you adopt its standard. Map to services you own and you have a specification, built from your data, configuration and observed process behavior.

The model also survives turnover. Business knowledge lives in custom code, tables, spreadsheets and people's heads, and people leave. Because the model is recorded rather than remembered, the context can be reconstituted and replayed: a new agent, a new engineer or a new owner can pick up where the last one stopped. The model belongs to you, and it is useful even if you never hire us for delivery.

Design note: the semantic-model approach extends the technical integration plan described below, which covers the evidence and control layers. It is what the agents reason over.

The evidence layer

Before any model sees anything, the data passes through four trust boundaries. Each has a different owner and different rules.

  • SAP boundary: your production applications, database and operational tooling. SAP stays the system of record.
  • Collection boundary: service identities, adapters and redaction. Collectors validate, redact and normalize evidence on the way out.
  • Analysis boundary: evidence storage and deterministic queries. Access rules are enforced here, in code, not in a prompt.
  • Model boundary: receives only the context authorized for one specific request.

Credentials never enter a prompt. Collection defaults to metadata and aggregates, not bulk extraction; a small authorized sample is pulled only when an investigation needs it. Collectors run on strict limits (small pages, response ceilings, query timeouts) that we load-test against your system before relying on them.

One rule matters more than the rest: missing telemetry is never treated as proof of health. Every evidence set records its own coverage gaps and collection failures, so silence in the data is reported as silence, not as "all clear."

Where a released API does not exist for what we need, we use an approved scheduled export or a reviewed exporter, and we never invent an interface. What we can reach differs by landscape (on-premises, private cloud, public cloud), and we settle that in the first week.

Four trust boundaries and data flow Data flows from the SAP boundary through the collection boundary, where it is redacted, into the analysis boundary, then as authorized context to the model boundary. Answers return through validators to a human reviewer. 1 SAP boundary Production apps, database, monitoring System of record. Unchanged. 2 Collection boundary Service identities, adapters Validate, redact, normalize 3 Analysis boundary Evidence store, deterministic queries Access enforced in code, per customer 4 Model boundary Only the context authorized for this one request Answers return through validators to a specialist.

What agents may and may not do

Agents get a short list of narrow, typed tools, each with bounded inputs and a defined output: get a system profile, cluster errors in a time window, check job overlap, read an interface backlog, pull finance exceptions, list upgrade findings. Each call is validated against allowlists, row caps and maximum time windows, and rejected before it reaches a source system if it falls outside the user's entitlements.

Just as important is what is not on the list. Agents have no unrestricted SQL, no arbitrary remote function calls, no shell, no GUI control, no transport import and no open outbound web access.

Untrusted by default

Logs, code comments, tickets and retrieved documents are treated as untrusted content. Text inside them cannot authorize a tool call or change the agent's instructions. Prompt injection through a ticket or a log is part of our test plan; the gateway must still refuse.

Answers have to show their work

Every answer carries cited evidence, ranked competing hypotheses, the evidence against each, what is missing, and a named approval owner. When the sources cannot support a conclusion, the answer is "insufficient evidence." A fluent but ungrounded answer does not pass.

Numbers are not the model's job

Finance figures are computed by deterministic reports and queries, never by the model. The model explains and links; it does not add up. Gross affected value is kept separate from confirmed misstatement, and finance approves any business-impact statement.

No production writes

Changes to production sit outside the agents' tools, including actions that look routine, like replaying a queue or cancelling a job. People approve and execute changes through your normal change process. Agents can draft a change request; a separate authorized workflow creates it.

How we know it works

We do not evaluate this by how convincing the narrative sounds. We build a blinded evaluation set of roughly 60 sanitized historical cases (runtime and performance, batch and integration, finance and business process, and upgrade findings). Your SAP reviewers label the expected evidence and acceptable conclusions. A holdout set stays separate from prompt development, and the set deliberately includes healthy periods, missing telemetry, conflicting evidence and similar errors with different causes. Reviewers who do not know which model or prompt produced an answer score it.

Go-live is gated on the thresholds below. These are proposed thresholds, not demonstrated results. We confirm during the assessment that your evidence is rich enough to meet each one.

GateProposed thresholdHow it is checked
Access controlZero unauthorized records in the negative-test suiteCross-customer, company, client and revoked-role tests
Evidence fidelityAt least 95% of sampled factual claims supported; every critical recommendation citedReviewer traces answer back to source
Numerical correctness100% agreement on finance amountsRecomputed through deterministic report or query
Diagnosis usefulnessAt least 80% of cases have a reviewer-accepted explanation or next step in the top threeBlinded specialist scoring
Unsupported certaintyNo critical unsupported claim; abstains on deliberately insufficient casesAdversarial and missing-data cases
SAP overheadUnder 2% incremental average CPU (initial target)Controlled load comparison, plus response-time and lock checks
ResponsivenessUnder 60 seconds for bounded analysis (initial target)p50 and p95, measured apart from collection lag

We also break things on purpose before go-live: expired certificates, unavailable sources, partial pagination, clock skew, delayed logs, model outages and rate limits. Stale and truncated data must be visible to the user. Rollout moves from development to quality assurance to a read-only production pilot with a small group, with security, Basis, functional and operations sign-off at each step.

Inference: lanes, providers and cost

We budget AI work as concurrent lanes of inference, not as tokens. Each agentic engineer runs many agent lanes in parallel, on the order of 70 billion tokens a month (a planning assumption; we use a 60-80 billion range). At that volume, what drives cost is not the text the agents write. Telemetry from 13.5 million coding-agent sessions shows about 94% of prompt tokens are served from cache and completions are under 1% of prompt tokens1. Agent cost is dominated by re-reading context. That is another reason to build the semantic model once and keep it compact and well organized.

We are model-agnostic. We route work by task: frontier models plan and review, cheaper models execute, and we can swap providers as prices and capabilities move. The table shows what that choice does to inference cost for one engineer for one month, at list prices checked 3 November 20262.

RoutingInference per engineer-month
Claude Sonnet 5.5 only$22K
Claude mix: Opus 5.5 plans and reviews, Sonnet 5.5 executes$27K
Claude Opus 5.5 only$43K
Claude Fable 5.1 only$92K
Hybrid mix$21K
DeepSeek V4 Pro (peak pricing)$10K
GPT-6 Astra only$141K
Qwen3 Coder (hosted)$8K

Planning estimates built from list prices234, an assumed token volume and an assumed cache-hit profile. Real cost depends on the mix of work and on contract terms. Provider choice swings the figure widely, which is why we treat it as a routing decision.

Inference is a cost line in every engagement we price, and it is built into the fixed fee. You can change the provider, volume and routing assumptions yourself in the estimator.

Security and audit

Each customer gets an isolated environment, with separate keys, secret stores and service identities. Roles are explicit (investigator, reviewer, administrator), and the evidence platform enforces the investigator's narrower rights before retrieval, analysis and download, even when a collector has broader access. Derived summaries, embeddings and caches inherit the classification of their source.

Classification happens before collection: payloads, financial amounts, credentials, personal information and source code are tagged first, and unnecessary identifiers and SQL literals are removed before data reaches the model boundary. Masking is an additional control, never a substitute for authorization or data minimization. We do not use your production material to train models or to serve other customers.

  • Retention (starting proposal): 30 days for restricted raw telemetry, 90 days for curated incident evidence, 365 days for approved decision and access audit records. Your policy, legal holds and SAP source retention override these, and we test deletion of raw data and of every derived index.
  • Audit trail: who asked, the policy decision, the source query, the evidence referenced, the tool arguments, the model and prompt versions, token consumption, response validation and the human approval. A citation always resolves to evidence the reader is allowed to see.
  • Kill switch: a switch disables model calls without disabling monitoring. Collectors and alerting run independently of inference, so if the model fails or hits its spending cap, your operations staff keep their normal dashboards and incident process.
  • Version control: model and prompt versions are frozen during a production upgrade window and promoted only after regression testing.

Which model provider and region you use is a contract decision made with your security and procurement teams, and it is settled before any technical build begins.

Human review is a design choice, not a fallback. Public large-scale code-modernization programs also keep human review layers around their AI output5. Our senior people do the same, and they own the decisions.

Sources

  1. GitHub Copilot coding-agent telemetry, 13.5M sessions, June 2026. arXiv 2608.00101.
  2. Anthropic, model pricing (per million tokens, checked 3 Nov 2026). platform.claude.com/docs/en/about-claude/pricing
  3. DeepSeek, API pricing. api-docs.deepseek.com/quick_start/pricing
  4. OpenAI, API pricing. developers.openai.com/api/docs/pricing
  5. AWS, "AWS Transform one-year milestone," 15 May 2026. aws.amazon.com/blogs/migration-and-modernization/aws-transform-one-year-milestone
  6. Core Portage, SAP S/4HANA LLM Integration and Delivery Plan, 25 October 2026 (internal technical plan; source of the architecture, controls, acceptance gates and retention proposals above, all of which are proposed design assumptions to be confirmed per customer).