Practice 03 · Agentic & Applied AI
Governed agentic AI on SAP BTP: MCP tool catalogs, Joule and A2A, SAP AI Core, LangGraph and RAG, under a control model where every write is read before, confirmed by a person and verified after. Agents explain. Humans decide. SAP stays the system of record. Underneath the agents sit methods our team has built and run: retrieval-augmented generation, metaheuristic optimisation, probabilistic models and constraint engines, most published with source code.
A 45-minute call with an architect, not a salesperson. Replies within one business day. How we run engagements
A read-only MCP tool catalog against your landscape first. Then confirm-to-act operations under the control model. Joule and A2A where SAP provides the surface; LangGraph and SAP AI Core where it does not.
Deliverablestool catalog with policies, control-model implementation, evidence pack for InfoSec, runbook.
The control modelThe controls a CIO can hand to any AI vendor, implemented rather than promised: read before write, confirm to act, verify after, and an audit trail per case. Workshops with finance, InfoSec and internal audit.
Deliverablescontrol model, policy versions, review UI patterns, audit design.
Get the checklistFaster on repeatable objects: mapping specs, integration flow scaffolds, tests, runbooks, extension boilerplate. Never transport approval, never production changes, never arithmetic inside an LLM.
Deliverablesaccelerated build with architect review on every artefact.
See SAP Intelligence SuiteApplied research
Agentic AI is the visible layer. Underneath it are methods our team has built and written up: probabilistic models, constraint engines, metaheuristic optimisers and retrieval systems. Most come with source code. The optimisers come with seeded results that reproduce to the cent.
Order-1 and order-2 chains with context back-off and temperature-controlled sampling, written from scratch in the standard library. Transitions are learned by counting, so every probability can be read and audited.
Applied tosequence generation from reference data, next-state prediction, process and clickstream simulation.
Seeded weighted random walks over hand-authored weight profiles, with each value projected onto an allowed set. 1 seed reproduces a run, in Python and in a JavaScript port for workflow automation.
Applied toprofile-driven sequence generation, scenario and what-if sampling, synthetic test data.
Several candidates are sampled per step and the lowest-cost one is kept under hard and soft rules: range, step size, forbidden pairings, learned likelihood. A repair pass then finds remaining violations and fixes them locally.
Applied torule-conformant multi-stream generation, scheduling and allocation under rules, configuration validation with auto-repair.
Single- and multi-objective annealing with a Pareto archive and an 11-point weight sweep, so a decision-maker sees the whole trade-off frontier and a knee point. Our book documents it with a working prototype for transport mode selection against cost and CO2.
Applied totransport mode selection, cost-versus-emissions planning, multi-objective scheduling.
A genetic algorithm whose mutation and crossover rates adapt online to population diversity and stagnation, with capacity and demand limits as soft penalties. Our book documents it with a working prototype for 3-tier supplier, plant and distribution-centre assignment.
Applied tosupply-network design, plant and route selection, capacitated assignment.
Particle swarm and a hybrid swarm-genetic method for continuous decisions, benchmarked against the closed-form EOQ. Runs are repeated across 5 seeds and the mean is reported, not only the best run.
Applied toinventory policy and order quantities, continuous parameter tuning.
Hybrid dense and keyword retrieval with cross-encoder re-ranking over SAP documentation and BPMN process exports. Every answer cites its document and page, and the model must say so when the context does not hold the answer.
Applied toSAP functional questions, process documentation, private document libraries that stay on your infrastructure.
A reference design for A2A agents on SAP BTP: agent cards, LangGraph orchestration and SAP AI Core models. The agent calls S/4HANA Cloud APIs as typed business tools from Joule or an external client. MCP tool catalogs sit under the same control model in Value Lens.
Applied togoverned agents on SAP BTP, Joule and agent-to-agent scenarios, custom Python jobs on SAP AI Core.
Numbers still come from deterministic code. These methods generate, rank, retrieve and explain; a person decides.
New product · Launching 12 November 2026
Our MCP production assistant turns a written brief into an editable Logic Pro session, with tracks, MIDI, song sections and arrangement structure. The same discipline as our enterprise agents: defined tools, verified actions and a person who makes the creative decisions.
Governed AI on SAP BTP
The same agent that can raise a purchase order can raise the wrong one. Our agents explain; your people decide. The numbers never come from the model.
The agent reads the current SAP state before it proposes anything.
A person reviews and confirms every create and update.
SAP is read back after every action, and the result goes into the audit trail.
Numbers come from deterministic code, never from the model. Every case carries its sources, its calculation and its decisions.
0
numbers produced by a model
Deterministic code only.
1
reviewer on every write
5
algorithms in our book, each with a working prototype
Adaptive genetic optimisation, simulated annealing, particle swarm, hybrid PSO-GA and Bessel-Fourier classification. Seeded, so every figure in the book reproduces.
25+
years of SAP delivery experience
Across our key architects.
AI-assisted engineering
We use our own AI-assisted tooling for the repeatable parts of integration and extension work. Every generated artefact is reviewed by a certified architect before it reaches your landscape.
Checklist
One page. The three controls (read before write, confirm to act, verify after) and the audit trail a CIO can hand to any AI vendor and ask: which of these do you implement?
No. There is no path in the catalog that writes without a person confirming, and the confirmation expires.
From deterministic code under a versioned policy, in integer cents. The model explains and proposes hypotheses, labelled as hypotheses.
Where SAP exposes the surface you need, yes; A2A for agent-to-agent. Where it does not, LangGraph on SAP AI Core or your cloud, with the same control model.
Tool catalog and policies, sample audit trails, read-back results, and the control-model document your InfoSec can review.
Value Lens is our product for order-to-cash margin leakage, in private beta. Client agents use the same control model and catalog approach against your own processes.
Deterministic code for every number. That rule comes first. Around it, methods our team has built and run. Markov chain and stochastic models for sequences and simulation. Constraint-based generation with repair passes. Metaheuristic optimisation: adaptive genetic, simulated annealing, particle swarm and hybrid PSO-GA. Bessel-Fourier descriptors with a linear SVM for image classification. Retrieval-augmented generation with hybrid retrieval, re-ranking and citations. A2A agents with LangGraph on SAP AI Core, and MCP tool catalogs in Value Lens. Most are published with source code; the rest are in our book with working prototypes.
Next step
A 45-minute call with an architect, not a salesperson. Replies within one business day.