Blockchain in 2026: from crypto hype to critical infrastructure for agentic AI and regulatory compliance

In December 2017 we published an introduction to blockchain on this blog. Back then, Bitcoin was flirting with $20,000 and Satoshi Nakamoto’s whitepaper was the only mandatory reference. Eight years later, the term ‘blockchain’ has disappeared from mainstream headlines. However, the technology —now under acronyms like DLT, tokenization or Web3 infrastructure— remains key. For example, it underpins $1.2 trillion in investment decisions on tokenized assets (BCG, 2024). Moreover, it is a sine qua non condition for deploying autonomous AI agents with legal accountability.

At BAOSS we have accompanied six clients —banking, logistics, energy and public administration—. All have moved from ‘innovation theater‘ pilots to production with contractual SLAs. This article sums up what we have learned. It covers the real 2026-2027 problem and the architecture that works. It also includes the metrics that justify CAPEX before the management committee.

The real problem 2026-2027: three bottlenecks holding agentic AI back in production

1. Provenance and auditing of training data. EU Regulation 2024/1689 (AI Act) requires full traceability of the datasets used in high-risk models. A Spanish bank fined €4.2M in Q1 2025 could not prove that its scoring model excluded protected variables. Indeed, its centralized data lake offered neither immutability nor non-repudiation.
2. Payments and settlement between autonomous agents. Agent-to-agent flows (MCP, A2A protocol) require instant, programmable and auditable settlement. SEPA Instant transfers (10 s) don’t cover micropayments < €0.01 nor smart contracts conditioned on off-chain oracles.
3. Governance of federated models and continuous fine-tuning. Data science teams use LoRA/QLoRA on Llama-3.1-70B or Nemotron-3-Ultra in vLLM/Ollama clusters. Without an immutable record of which checkpoint served which client, legal liability falls on the provider.

The BAOSS 2026 reference architecture: DLT as a trust layer, not as a database

The key lesson: don’t replace PostgreSQL with Hyperledger Besu. Use DLT only where you need trust minimization between parties that don’t trust each other. Our production-proven pattern:

• First, operational data layer: PostgreSQL + TimescaleDB (time series) + pgvector (RAG embeddings).
• Then, evidence and audit layer: Hyperledger Besu (permissioned network, IBFT 2.0) or Polygon CDK (validium), anchoring daily Merkle roots of critical tables.
• Then, AI orchestration layer: LangGraph + CrewAI for multi-agent flows; MCP (Model Context Protocol) for on-chain/off-chain tool exposure.
• Identity and payments layer: DID/VC (W3C) + ERC-4337 account abstraction for gasless user ops. Then, Circle USDC / EURC for fiat on-ramp settlement.

Case 1: Wholesale banking — Scoring traceability under the AI Act (anonymized)

Context: A Spanish financial institution (assets > €80B) needed to certify one specific thing to the supervisor: that its SME scoring engine (XGBoost + Llama-3.1-8B fine-tuned for risk narrative) did not use gender/ethnicity variables implicitly embedded in the embeddings.

BAOSS solution (12 weeks, 4 engineers):

• CI/CD pipeline with GitLab + DVC: every training run generates a signed artifact (cosign + sigstore). Then, the hash is anchored in Besu (permissioned network Banco de España + 3 institutions).
• RAG over AI Act regulation (PDF + EUR-Lex) with a BGE-M3 reranker. Then, a LangGraph agent validates SHAP feature importance against the prohibited list before promoting to staging.
• Auditor dashboard: immutable lineage query (dataset → checkpoint → inference) via SQL view + on-chain Merkle proof.

Metrics:

• 68% reduction in supervisor request response time (from 22 days to 7).
• Also, zero findings in the EBA Q3 2025 inspection.
• Also, DLT infrastructure cost: €0.003 per anchoring transaction (vs €0.12/tx on Ethereum mainnet).
• In short, 3.2x ROI in 6 months (fine savings + compliance efficiency).

Case 2: Perishables logistics — Agent-to-agent automated settlement with IoT oracles

Context: Fruit exporters’ consortium (Spain → Nordics, 12M pallets/year). Disputes over temperature deviation in transit: 18% of invoices disputed, 45-day average collection time.

BAOSS solution (16 weeks, 5 engineers + 2 domain experts):

• BLE 5.3 sensors (Nordic Semiconductor) sign readings every 15 min with an ECDSA secp256k1 key. Then, the edge gateway (Raspberry Pi 5 + Ollama Llama-3.2-1B) aggregates and sends an hourly Merkle root to Polygon CDK validium.
• A CrewAI ‘Quality Oracle’ agent evaluates the time-temperature integral against the contract. If there is a breach, it fires an ERC-4337 user op. This settles the penalty in USDC to the carrier and credits the difference to the exporter (gasless for the end user).
• WhatsApp Business + MCP interface so operators check pallet status in natural language (‘Did pallet ES-2026-044512 keep the cold chain?’).

Metrics:

• Disputed invoices: 18% → 2.3% (Q4 2025).
• Also, average collection days: 45 → 1.2 (on-chain T+0 settlement).
• Also, cost of the full transaction (sensor → settlement): €0.008.
• Moreover, 40% reduction in deployment time vs the 2023 pilot (reuse of BAOSS IoT-DLT Bridge modules).

Case 3: Public administration — Immutable registry of fine-tuning for open LLM models

Context: A state agency (AI budget €12M/year) deploys 14 specialized models (procedures, grants, inspection) on a Llama-3.1/Nemotron base. It requires full traceability for parliamentary accountability.

BAOSS solution (10 weeks, 3 engineers):

• Mandatory ‘Model Card’ registration at every merge to main. It includes dataset hash, LoRA hyperparameters, eval metrics (MMLU, hallucination rate with RAGAS) and human signatory.
• Daily Merkle root anchoring of the model_registry table on a Besu network (nodes: agency, CNI, Court of Auditors, university).
• An AutoGen ‘Compliance Watcher’ agent monitors production drift (embedding shift > 0.15 cosine). Then, it opens a Jira ticket, notifies Slack and proposes a scheduled re-train.

Metrics:

• First, 100% traceability of production models (0 ‘shadow models’ detected in the 2025 audit).
• Also, audit-ready Model Card generation time: 4 min (automated).
• Also, on-chain cost: €12/month (Besu gas + IPFS/Filecoin storage).
• Finally, 55% reduction in MLOps team hours on compliance tasks.

Tools and the 2026-2027 stack we use at BAOSS

Mistakes we keep seeing (and how to avoid them)

1. Using public mainnet for sensitive data. Even with ZK-proofs, the metadata (frequency, addresses) leaks competitive information. Therefore, the solution: permissioned validium / L2 + periodic commitment to mainnet if public verifiability is needed.
2. Ignoring account abstraction (ERC-4337). Also, forcing end users to manage keys/seed phrases slows adoption. Smart accounts + paymasters + passkeys (WebAuthn) = Web2 UX with Web3 security.
3. Underestimating oracle governance. A manipulated price feed or a compromised IoT sensor invalidates the whole chain of trust. Therefore, it requires multi-source + TEE (Intel TDX / AMD SEV-SNP) + on-chain slashing.
4. 
   

   Prefer to talk to us directly? Contact the BAOSS team and tell us about your challenge.

   Related articles: Blockchain in the enterprise: beyond cryptocurrencies / Interview with La Voz Empresarial, El Mundo Newspaper.