Enterprise AI Agents in 2026: How to Run Automation on the Systems You Already Have
António Calçada
Enterprise AI agents moved from research topic to board-level line item in under three years. Deployment has not kept pace. Roughly 31% of enterprises run one in production, and the reason is rarely the model itself: it is the integration and governance work sitting underneath.
This guide covers what enterprise AI agents are, why the majority never reach production, the architecture that separates the ones that do, and the deployment sequence, governance controls and metrics that regulated organisations use to run autonomous operations at scale.
Key Takeaways (TL;DR)
Enterprise AI agents are execution systems. The test is whether the case closes. Reading an intake, querying SAP and three other systems, applying policy and writing the outcome back counts as a completed operation; producing a draft for someone to action does not.
The bottleneck is infrastructure. Getting AI into production inside an enterprise is roughly 90% an infrastructure, integration and process design problem, and 10% an AI problem.
The production gap is measurable. Gartner forecasts that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from under 5% in 2025, while Deloitte's 2026 survey of 3,235 leaders found only 25% had moved 40% or more of their agent pilots into production, and only 21% report mature agentic AI governance.
Cancellations trace to governance, not models. Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls.
Separating interpretation from decision is the architectural rule. AI reads unstructured inputs. Deterministic rules, mapped from the organisation's own SOPs, execute every governed decision.
Legacy depth beats connector count. SAP ECC, Guidewire, Oracle and COBOL-era cores will not be replaced in the next three to five years. An agent that cannot operate inside them cannot touch the highest-value operational work.
Manufacturing quality operations on SAP are the clearest current example. Supplier document processing, 8D claim handling, PPAP review, supplier scoring and audit preparation all run end to end with SAP QM and MM write-back.
Sovereignty is architectural. Self-hosted, hybrid or on-premises, with bring-your-own-key encryption, determines whether a deployment is viable under GDPR, DORA, NIS2 and the EU AI Act.
Production benchmarks are documented. Live deployments show 3–5x ROI, 90–96% precision on real operational data, and 45–80 days from contract to production, with Santander branch operations live across five countries in under 90 days.
Test every agent on this page on whether the case closes in your system of record.
Reading the intake, querying the systems, applying policy, writing the outcome back
Book a Demo
Table of Contents
No headings found on page
Enterprise AI Agents: at a Glance
Dimension | Where Things Stand in 2026 |
Definition | Software that plans and executes multi-step operational work across enterprise systems, with AI interpreting unstructured inputs and deterministic rules governing every decision |
Unit of value | Operations completed – claims closed, invoices reconciled, documents validated and written back – rather than tokens processed or messages deflected |
Market size | $10.9bn globally in 2026, up from $7.6bn in 2025 (Grand View Research) |
Adoption | 40% of enterprise applications will embed task-specific agents by end of 2026, up from under 5% in 2025 (Gartner) |
Production reality | Around 31% of enterprises run at least one agent in production. Banking and insurance lead at 47%, healthcare sits at 18%, government at 14% (S&P Global Market Intelligence / McKinsey) |
Governance maturity | 21% of organisations report mature agentic AI governance, against 74% expecting at least moderate agentic use within two years (Deloitte, 2026) |
Failure forecast | Over 40% of agentic AI projects cancelled by end of 2027 on escalating costs, unclear business value and inadequate risk controls (Gartner) |
Where value concentrates | High-volume operations crossing three to seven systems per case, governed by written policy, with a measurable manual cost baseline |
Hardest environment | ERP-anchored operations. SAP ECC and S/4HANA, Guidewire, Oracle and COBOL-era cores, where most vendors decline to operate |
Regulatory backdrop | EU AI Act Article 11 and 12 enforcement for high-risk systems from 2 August 2026, alongside GDPR, DORA, NIS2, Solvency II and HIPAA |
No Time to Read? Your Enterprise AI Agents Deployment Checklist
What Are Enterprise AI Agents?
Enterprise AI agents are software systems that plan and carry out multi-step operational work across an organisation's systems, with limited human supervision and under defined governance controls.
What separates them from earlier categories of business software comes down to how much of a case they can carry.
An agent receives an intake, which might be an email, a scanned certificate, a portal submission or an EDI message, and works the case through to an outcome. It performs the lookups, applies the policy, executes the actions and writes results back into the system of record.
Traditional automation needs a predictable field in a predictable place, which is why it fails on most real intake. Enterprise AI agents read free text, inconsistent document layouts, mixed-language correspondence and attachments of varying quality.
They also carry the case across system boundaries. A single supplier quality claim might touch SAP QM, SAP MM, a supplier portal, an email inbox and a document store, and the agent moves through all five in one continuous run.
The category has expanded quickly. Grand View Research put the global AI agents market at $10.9 billion in 2026, up from $7.6 billion in 2025. Gartner expects 40% of enterprise applications to embed task-specific agents by the end of this year, against fewer than 5% in 2025.
Growth in supply has outpaced growth in production deployments. That gap is the subject of the rest of this guide.
What an Enterprise AI Agent Is Not
A useful test: if the output requires a person to review, copy, paste or re-key it into another system, the work was assisted rather than completed.
A chatbot answers questions in a conversation window and leaves the record untouched. A copilot produces a draft. A person still owns the action. A model with tool access can call an API. Without process mapping, codified rules and an audit trail, it cannot be trusted with a regulated operation. A sandbox pilot proves a model works on sample data and proves nothing about the production environment.
"Agent washing" – relabelling existing conversational tools as agents – is one reason Gartner attributes the coming cancellation wave to management and governance rather than model capability.
Enterprise AI Agents vs Copilots, Chatbots and RPA
Most organisations already run some form of enterprise automation.
Understanding where each category breaks explains why agents are being funded despite the failure statistics.
The distinction that matters commercially is the unit of measurement. RPA reports tasks automated and conversational AI reports deflection rate, neither of which tells a CFO what the operation now costs. Enterprise AI agents should be counted in operations completed end to end.
For a vendor-by-vendor view of the current market, see our analysis of the best enterprise AI agents.
Why Most Enterprise AI Agents Never Reach Production
Every large organisation knows where agents should help. Complaints, claims, billing, supplier documentation, account changes. The use cases are obvious and the models are capable.
Between "this should work" and "this is live" sits a set of structural problems that no model release resolves.
1. The systems were never designed to be automated
Enterprise environments were built over decades, layered on top of each other, rarely documented and never designed to communicate.
A single case moves through five or six platforms before resolution, with staff bridging every gap manually. Automation has to reproduce that bridging work, and no API contract describes it.
2. The business logic lives in people, not systems
How work actually gets done in a regulated enterprise sits in the heads of long-tenured staff: the informal rules, the exception patterns, the tolerance that is always granted for one supplier and never for another.
None of it appears in the process documentation. Until it is surfaced and codified, an agent has nothing reliable to execute against.
3. Governance arrives after the incident
Gartner's cancellation forecast points at inadequate risk controls as a primary cause.
The pattern repeats: organisations deploy at high autonomy, discover a control gap through a production incident, then bolt on logging dashboards and approval steps that do not integrate with the audit process. Retrofitted governance rarely survives a regulator.
4. RPA burned the credibility budget
Many operations leaders already funded an automation programme that broke.
Bots failed on interface updates, maintenance costs climbed past the savings, and leadership lost confidence in the category rather than the vendor. The next request faces a higher evidential bar.
5. Shadow AI is already inside the building
Staff are using consumer AI tools with operational data today, outside IT policy and without logging.
The governance question is no longer whether AI touches enterprise data. It is whether that contact is recorded.
6. Integration consumes the budget before the work starts
Every new tool needs months of IT effort to connect to existing systems. Engineering capacity goes into plumbing, and the operational problem that justified the investment stays unsolved.
Building that integration and orchestration layer internally typically costs €500k to €1M+ in engineering before a single process is automated. Most organisations never finish.
Name which of those structural problems is holding your own pilots.
No model release resolves the integration and governance layers
Book a Demo
The Architecture Behind Enterprise AI Agents That Execute
Enterprise AI agents that reach production share a common structure.
Three layers, each solving a different part of the problem:
Layer 1: Workflow design
A visual environment where operations teams define multi-step, multi-system workflows without writing code, together with knowledge base management for company policies, SOPs and reference data.
The point of this layer is ownership. When a process change needs an IT ticket and a release cycle, the workflow drifts out of date within a quarter. The people who own the process need to modify the logic.
Layer 2: Integration and orchestration
The backbone connecting the agent to the systems the business runs: SAP ECC and S/4HANA, SAP QM, MM and PP, Guidewire, ServiceNow, Oracle, Salesforce, SharePoint, Snowflake, Outlook and the proprietary platforms unique to each industry.
Noxus connects through 400+ connectors, and the deployment work concentrates on the systems other vendors decline to touch.
This layer also carries the controls that make autonomy governable:
Model independence across fourteen providers with fallback chains, on Azure OpenAI, Vertex AI, AWS Bedrock or private models, with no training on client data
Bring-your-own-key encryption
Confidence-based escalation to named humans
RBAC integrated with Azure AD and enterprise identity
Versioning, rollback and CI/CD integration
Monitoring through OpenTelemetry and Phoenix
Full lineage on every action, user, policy, model and input, replayable and exportable for regulators
The principle is integration depth over integration breadth. A marketplace of connectors matters less than the ability to write back into the one ERP where the high-value work sits.
Layer 3: Process intelligence runtime
The execution engine. It manages process lifecycle, task orchestration and system write-back in production, at a scale of millions of operations a month.
This layer determines whether an agent resolves or merely proposes. The runtime carries the case from intake to closure, posts the outcome into the system of record, and retains the trace.
Autonomy Levels and Confidence-Based Escalation
Autonomy is set by degree, and the degree is a deliberate choice made per process. Get it wrong in either direction and you end up with either a governed deployment nobody trusts to act, or an uninsurable one.
A workable model runs across four levels:
Level 1: Assisted
The agent assembles the case file, retrieves records across systems and presents a recommended action. A human executes. Useful for trust-building in the first weeks of production.
Level 2: Supervised execution
The agent executes in-policy cases and queues the outcome for approval before write-back. Common in claims and credit decisions during ramp-up.
Level 3: Autonomous within policy
The agent resolves and writes back any case falling inside codified policy. Cases outside policy, or below the confidence threshold, route to a human. Most high-volume production deployments settle here.
Level 4: Autonomous with exception ownership
The agent handles the in-policy majority and manages the exception queue itself, gathering missing evidence from suppliers, applicants or internal teams before routing.
Two controls make the higher levels defensible:
The interpretation and decision split - AI reads emails, parses scanned certificates, classifies free-text requests and extracts data from inconsistent formats. Every governed decision – approve, reject, refund, escalate, pay – runs through hard-coded business logic mapped from the organisation's own SOPs. No model invents policy. Every step replays.
The confidence floor - the client sets the threshold. Below it, the case escalates to a human who receives the complete file: every input, every lookup, every intermediate step. The escalation itself forms part of the trace.
One practical governance test during evaluation: can a non-technical administrator halt a deployed agent within seconds, without opening a ticket or hunting for credentials? If not, the kill switch is too slow for a production incident.
How to Deploy Enterprise AI Agents: From Process to Production
Production deployments follow a consistent shape.
Noxus runs it in three phases, and the six stages underneath describe what happens inside them:
Phase 1: Prioritised business case
Map how the operation runs today, quantify the value, and redesign the process for AI execution with human oversight.
Phase 2: Proof on your systems
Build the process on the client's live systems and prove it in the client's own environment within weeks, against criteria agreed upfront.
Phase 3: Production and scale
Deploy, embed the platform with the client's team, and extend into adjacent processes as the business pulls.
Stage 1 – Select the process, not the technology
Strong first candidates share four traits: high volume, several systems per case, written policy covering the majority of cases, and a measurable cost baseline. Supplier document processing, invoice reconciliation, complaint handling and document verification qualify in most organisations.
Avoid starting with the most politically visible process. Start where the baseline is cleanest, because the first deployment has to produce a defensible number.
Stage 2 – Map the process as it runs
Document every system each case touches. Record the copy-paste bridges, the spreadsheets sitting between two platforms, and the steps depending on knowledge only two people hold. The documented process and the real process diverge in almost every organisation.
Stage 3 – Codify the rules
Translate SOPs, policy documents, contractual terms and regulatory obligations into deterministic logic. Where written policy is silent, capture the informal rule and get it approved. This stage usually surfaces inconsistencies that predate any AI programme.
Stage 4 – Integrate into the systems you already run
Connect to the systems as they exist, with no API modernisation or middleware project required first. Where a modern interface exists, such as SAP OData, the agent uses it. Where none exists, the agent operates the interface the way the operations team does today.
Stage 5 – Prove it on live data, then switch on write-back
Run the agent alongside the human team on live cases with write-back disabled. Compare outcomes, measure precision, tune the confidence floor. Then enable write-back for the in-policy segment. This phase typically runs one to three weeks and produces the evidence that closes the internal governance review.
Stage 6 – Expand deliberately
Start with the cases that follow clear rules and route the rest to expert humans with the evidence pack pre-assembled. As precision improves on production data, the automated boundary moves outward.
Expansion economics are structural. Integrations and governance controls are already running, so the second and third workflows are predominantly platform spend, at 85–90% gross margin in documented deployments.
Enterprise AI Agents on SAP: Six Quality Operations Workflows
Manufacturing quality operations show the pattern at its clearest.
The work is high volume, document-heavy, contractually deadline-bound, and anchored in SAP QM, MM and PP. KIRCHHOFF Automotive runs production operations connected directly to SAP through OData, from intake to write-back. Carlsberg Group reconciles supplier invoices against purchase orders and delivery records, with only exceptions reaching finance.
Six workflows cover the majority of the supplier quality lifecycle.
1. Supplier quality document processing
Certificates, test reports, control plans and inspection records arrive by email, supplier portal, EDI or shared drive. The agent identifies the document type, matches it to the supplier, part number and PO reference against the SAP MM and QM record, extracts the required fields following type-specific protocols, and validates every value against SAP acceptance criteria, including certificate validity and accreditation.
Validated data writes into the inspection lot, goods receipt or quality notification with full traceability. Failed or incomplete documents route to the quality team with the document, the extracted data, the failure reason and the SAP context already assembled. Manual entry disappears and the inbox stops accumulating.
2. Supplier quality claims processing
When a defect is identified at incoming inspection, on the line or by a customer, the agent opens the claim as a SAP QM notification with supplier, part, batch and defect detail. Inspection records, measurement data, photos and production context are pulled from SAP QM and linked automatically.
SAP MM is queried for every location receiving the part, and containment notifications route to the affected quality teams. The supplier receives the claim, the evidence package and the required 8D format with the deadline set by contractual terms. Each 8D section is checked for completeness, gaps return to the supplier, and complete submissions route for technical review. Corrective action is tracked against the agreed timeline, with missed deadlines escalating to the Quality Director. On verification, the claim closes in SAP QM and the findings feed the supplier scorecard and incoming inspection criteria.
3. Customer quality claim processing
An OEM complaint arrives through the customer portal, email or EDI. The agent captures the details, creates the SAP QM notification, and pulls batch records, traceability and inspection results from SAP QM, PP and MM. A full traceability query identifies every batch from the same material, process or supplier lot, including parts already shipped or sitting in warehouse.
The 8D report is pre-populated in the OEM's required format with the assembled data, analysis sections left open and routed to the responsible engineer. The engineer completes root cause, corrective and preventive actions and signs off. Acknowledgement, interim and final deadlines are monitored independently, with at-risk dates escalating to the Quality Director. The completed 8D goes to the OEM, the claim closes in SAP QM, and the full case file is retained for audit.
The engineer still owns the analysis. What changes is that the four systems' worth of data behind it arrives already assembled.
4. Supplier PPAP review
The agent matches each PPAP submission to the part number, drawing revision and submission level, then maps every element against the requirement set for Level 1 to 5 and flags missing elements immediately. Measurement results are validated against drawing tolerances, with out-of-tolerance conditions flagged by characteristic, value and tolerance band. Material certifications are validated against specification, and process flow, PFMEA and control plan are checked for internal consistency. MSA studies and Cp/Cpk are reviewed against customer-specific thresholds.
The engineer receives a structured pass/fail assessment per element with the full package attached, then approves, rejects or approves with conditions. The decision records in SAP QM, and rejections return to the supplier with the specific gaps and required actions.
5. Supplier performance monitoring
Inspection results, claims, PPAP outcomes, 8D response times and closure rates are collected continuously from SAP, per supplier, across all locations. PPM, claim frequency, response time, closure rate, OTD% and audit score are weighted per quality policy and updated continuously, consolidated into one score per supplier across every receiving plant with location-level detail preserved underneath.
Green, yellow and red thresholds trigger early warning on tier changes and deteriorating trends. The Supplier Quality Director receives the scorecard, the 12-month trend, open claims, corrective action status, audit findings and affected locations as a single dossier. Red tier escalates to a decision, tracked to resolution on the supplier record. No quarterly reports, no spreadsheet consolidation.
6. Supplier audit preparation and follow-up
The audit calendar generates from supplier risk tiers, with audit type assigned per risk profile and updated when tiers change. The pre-audit briefing assembles the scorecard, twelve months of claims, corrective action history, previous findings and PPAP outcomes. The checklist pre-populates from audit type, previous findings and open risk areas, with committed corrective actions flagged for verification.
The audit itself stays human-led on site. Findings are then structured into VDA 6.3 scoring and filed against the supplier record. Each non-conformance carries an owner, deadline and required closure evidence, with missed deadlines escalating. The audit score writes back to SAP, the risk tier recalculates, and the next audit interval adjusts, feeding the scorecard that drives the following cycle.
Enterprise Automation Use Cases Beyond Manufacturing
The same pattern applies wherever operations run at volume across several systems.
The tables below summarise live use cases from the Noxus enterprise AI use case library, with the two headline metrics recorded for each.
Banking and financial services
Use case | What the agent does | Outcome |
Current account complaint handling | Retrieves records from core banking, CRM and document systems, applies complaint policy, closes in-policy cases | 60% automation, 74% first contact resolution |
Credit card complaint handling | Reconstructs transaction history and CRM context, applies complaint handling policy | 60% in-policy automation, 72% first contact resolution |
Mortgage complaint handling | Retrieves the mortgage account record, payment history and product terms from the servicing system | 80% automation, 70% reduction in resolution time |
Unexpected transaction and fee investigation | Validates charges against the fee schedule version applicable at the charge date, executes SOP-eligible refunds | 60% investigation automation, 78% first contact resolution |
Loan origination document processing | Identifies the required document pack by loan type, requests missing items from applicant or broker | 65% intake automation, 60% reduction in application cycle time |
Merchant onboarding and KYC verification | Verifies entity and UBO structure, runs PEP and sanctions screening, validates the document pack | 75% intake automation, 82% supplier data completeness |
Customer profile change and write-back | Validates identity documentation, applies standard changes across connected systems | 88% in-policy automation, 75% reduction in approval cycle time |
Pensions, investments and insurance
Use case | What the agent does | Outcome |
Pension and investment account complaint handling | Retrieves records from pension administration and custody platforms across channels | 55% automation, 65% reduction in average resolution cycle |
Pension drawdown and UFPLS processing | Validates the drawdown instruction against scheme rules and applicable tax treatment | 40% in-policy automation, 75% reduction in instruction error rate |
ISA transfer handling | Receives transfer instructions across cash and stocks and shares ISA types, collects and validates the required information | 70% transfer initiation automation, 65% transfer query automation |
Annuity payment and rate dispute resolution | Retrieves the annuity contract, original quote illustration and payment history, validates the payment calculation | 70% dispute automation, 78% first contact resolution |
Policy onboarding and renewal processing | Collects and verifies personal details and identity documents for new business and renewals | 70% straight-through completion, 65% cycle time reduction |
Property insurance complaint handling | Spans policy administration and claims, reconciles evidence from both and applies property coverage rules | 68% automation, 50% reduction in cost per complaint |
Motor and transport complaint handling | Retrieves the policy record, claims history and repair documentation, resolves in-policy complaints | 80% automation, 40% reduction in handling time |
Billing insurance claims | Extracts procedure and diagnosis codes from clinical records, matches against payer billing rules, submits clean claims | 95%+ clean claim rate, under 6% denial rate |
Healthcare
Use case | What the agent does | Outcome |
Patient onboarding | Collects patient data and consent documentation across intake channels, validates completeness | 75% automated intake, 60% reduction in intake time |
Patient complaint handling and resolution | Classifies by type and urgency, resolves administrative complaints, routes clinical cases with the full record | 40% resolution automation, 60% reduction in resolution cycle |
Patient feedback triage and resolution routing | Classifies feedback by sentiment, category and risk level, flags high-risk cases | 88% classification accuracy, 75% faster flagging of high-risk cases |
Retail, e-commerce and FMCG
Use case | What the agent does | Outcome |
Product data enrichment and catalogue management | Classifies against the defined taxonomy, applies naming conventions and description standards, writes back to PIM | 90% automated classification, 98% catalogue record accuracy |
Order, delivery and billing complaint resolution | Retrieves the order record and payment history, applies resolution policy, writes back to OMS and CRM | 65% automation, 74% first contact resolution |
Returns and refund processing | Verifies eligibility against policy, processes approval or rejection, generates the carrier return label | 85% in-policy automation, 55% reduction in return processing time |
Customer feedback triage and retention routing | Classifies inbound feedback and NPS responses, identifies detractors and high-risk accounts | 88% classification accuracy, 80% faster detractor follow-up |
Competitor and market intelligence | Retrieves competitor pricing, product launches and promotional data across monitored sources | 90% of sources monitored automatically, 80% faster market signal processing |
Utilities and cross-industry operations
Use case | What the agent does | Outcome |
Meter read exception handling | Validates reads against expected consumption ranges and estimation rules, resolves in-policy exceptions | 70% exception resolution, 75% reduction in processing time |
Billing, supply and switching complaint handling | Applies resolution policy across billing disputes, supply interruptions and switching delays | 60% automation, 72% first contact resolution |
Contract expiry and SLA monitoring | Connects to contract repositories across SAP, SharePoint and procurement systems, monitors expiry windows and SLAs | 88% of contracts monitored automatically, 80% reduction in missed renewals |
Invoice processing and reconciliation | Extracts key fields, validates against purchase orders and contract terms, posts approved invoices to SAP and Oracle | 78% automation, 70% reduction in cycle time |
Two patterns run through the whole library.
First, the metrics cluster in pairs: an automation rate and a cycle time reduction. Second, automation rates vary by how much of the process is governed by written policy. Structured document work reaches 88% and above.
Complaint handling in advice-adjacent areas such as pensions sits at 40% to 55%, because more of the caseload needs human judgement, and the return there comes from the cycle time reduction on everything else.
Enterprise AI Agents by Industry
Adoption is uneven by sector. S&P Global Market Intelligence and McKinsey data show banking and insurance leading production deployment at 47%, with healthcare at 18% and government at 14%.
1. Manufacturing
Production planning, quality control, supplier management and compliance documentation span multiple platforms with manual handoffs at every boundary. Automotive, beverage, pharmaceutical, chemical and industrial machinery manufacturers generate the highest volume of multi-system, document-heavy operational work around production, and the compliance layers on top of it add volume without adding value.
The six SAP quality workflows above describe where that concentration sits. KIRCHHOFF Automotive and Carlsberg Group are running examples.
2. Financial services and banking
Legacy-dense by definition. A single customer case touches core banking, CRM, document management, compliance systems and several communication channels, with staff navigating three to seven systems per case. GDPR, DORA and Consumer Duty make replayable lineage a structural requirement.
Santander runs branch operations as a governed process across five countries on the bank's own infrastructure, live in under 90 days, with first production reached in 45 days and 95% precision on live operational data.
3. Insurance
Claims operations span policy administration, claims platforms, document repositories and communication channels.
AI reads claim forms, parses medical reports and extracts policy details from scanned documents. Coverage and liability decisions run through rules mapped from the insurer's own product wording and regulatory obligations. Solvency II and conduct regulation are addressed through hosting architecture, lineage design and escalation thresholds.
4. Healthcare
High volume, high sensitivity.
Patient communications, appointment coordination, billing, complaint handling and referral management cross clinical and administrative systems under GDPR Article 9. No model makes a clinical or compliance-relevant decision. Production results: 3x ROI, 96% precision, 10,000+ communications per month resolved, 50 days to production.
5. Retail and FMCG
Catalogue operations, order management and complaint handling at volume.
Classification, enrichment and pricing across PIM systems and marketplace APIs support automation rates above 90%. Production results: 5x ROI, 90% precision, 80 days to production.
6. Utilities
Meter read exceptions, billing disputes, supply interruptions and switching delays all follow the same shape: retrieve the record, validate against consumption ranges or resolution policy, act, and write back to the billing system, CRM and regulatory complaints log.
Complaint handling in this sector also has to produce audit-ready records for periodic regulator review. Exception resolution runs at 70% with processing time down 75%, and first contact resolution on complaints sits at 72%.
7. Public sector
Lowest current production rate and among the highest structural need.
Administrative processing, document handling and citizen correspondence follow the same multi-system pattern, with sovereignty requirements that rule out most hosted options. On-premises deployment is usually the entry condition.
Governance, the EU AI Act and Audit Readiness
Governance maturity trails deployment by a wide margin. Deloitte's 2026 survey found only 21% of organisations report mature agentic AI governance, against 74% expecting at least moderate agentic AI use within two years.
The regulatory calendar has closed the gap between intention and obligation. Enforcement of the EU AI Act's technical documentation and event logging requirements for high-risk systems took effect on 2 August 2026. DORA adds operational resilience obligations for financial services, NIS2 does the same for critical infrastructure operators, and GDPR governs processor arrangements and special category data.
What an audit-ready enterprise AI agent deployment produces:
Full lineage per run: which inputs arrived, how they were interpreted, which policy version applied, which model was used, which systems were queried, what action followed and what was written back. Replayable and exportable for regulators.
Replayability: any historical case reconstructs step by step. When an outcome is challenged, the answer is a record.
Role-based access control: integrated with Azure AD or the existing identity provider, so agent permissions inherit from the model governing human access.
Versioning and rollback: every change to a workflow or rule set is versioned, and a faulty change reverts without a release cycle.
Named human escalation: below the confidence floor, a named role receives the case with full context, and the escalation is part of the record.
Certification: ISO 27001, SOC 2 Type II, GDPR and HIPAA in place, ISO 27701 in progress, NIS2 controls implemented.
Treated as procurement checkboxes, these produce compliance theatre: logging dashboards that do not integrate with the audit process, approval workflows layered on after go-live, access policies documented but not enforced. Built into the execution layer from the first workflow, they cost nothing extra and survive scrutiny.
Sovereignty: Running Agents on Your Own SAP Environment
For regulated organisations, the hosting architecture often determines viability before AI capability is assessed.
Three patterns cover the range:
Self-hosted: Noxus and SAP both run inside the client's own cloud, so the platform, the execution and the data all stay within the client environment.
Hybrid: Noxus hosts the platform so the client does not have to. The component that touches SAP, the agent, lives inside the client's cloud next to SAP. Only orchestration signals cross the boundary, and the data stays put.
On-premises: for SAP running on the client's own servers, Noxus places an agent alongside it, on premises, talking straight to the database. This is the pattern for estates that expose no modern APIs.
Fully managed SaaS remains available for first use cases and environments with standard data sensitivity requirements.
Across all patterns, bring-your-own-key encryption keeps data inside the client perimeter, and model independence spans fourteen providers with fallback chains on Azure OpenAI, Vertex AI, AWS Bedrock or private models, with no training on client data.
An open-core guarantee matters at the other end of the relationship: if the contract ends, the client retains all code and binaries, and the deployed workflows keep running.
US-built agent products were designed around modern SaaS stacks and North American regulatory contexts. European enterprises run on neither, and retrofitting sovereignty into an architecture that assumed a hosted model rarely works.
Measuring Enterprise AI Agents: The Metrics That Matter
The unit of value is operations completed. Token counts, message volumes and deflection rates describe activity rather than outcome.
Metric | What It Measures | Production Benchmark |
Automation rate | Share of cases resolved end to end without human intervention | 60% (complex complaint handling) to 95%+ (structured document processing) |
AI precision | Accuracy on live operational data, not test sets | 90–96% across documented deployments |
Per-case cost reduction | Reduction against the pre-deployment baseline of headcount, BPO or manual processing cost | Direct comparison against a measured baseline |
Time to production | Contract signature to live operation on the client's real systems | 45–80 days, with multi-country rollouts inside 90 |
Escalation quality | Share of escalated cases resolved by the human on first touch, with no further data gathering | Rises as the evidence pack improves |
ROI | Return attributed to a specific operational baseline | 3–5x across documented deployments; median payback across the wider market runs around 5 months |
Expansion margin | Gross margin on second and third use cases on the same infrastructure | 85–90% |
Lineage completeness | Share of runs producing a full replayable trace | 100%, or the deployment is not audit-ready |
Baseline before go-live. Cost per case, average handling time, error rate and monthly volume, captured for at least one full cycle. Without it, the post-deployment number has nothing to compare against and the business case turns into a debate about attribution.
Manual error rates in complex multi-system operations typically run 3–8% at scale, often the largest single component of the return once downstream rework is counted.
Build vs Buy: The €500k Question
Internal engineering teams can build agent infrastructure. The question is what it costs and when it delivers.
Building the integration, orchestration, governance and lineage layers internally runs €500k to €1M+ before a single process is automated, and typically 12+ months to a first production workflow. The AI portion of that effort is small. The bulk goes into legacy connectors, retry and fallback logic, audit persistence, RBAC, versioning and the operational tooling that keeps it running at 3am.
Three questions clarify the decision:
1. Is agent infrastructure a competitive asset here?
For a manufacturer, the competitive asset is the quality policy and the supplier knowledge, not the plumbing that executes them.
2. What does the delay cost?
Twelve months of internal build against 45 to 80 days to production compares when the saving starts, not just what it costs.
3. Who maintains it in year three?
Internal builds depend on the engineers who wrote them, attrition in that population is high, and documentation rarely matches intention.
Buy-then-extend is the pattern that holds: deploy on existing execution infrastructure, keep the process knowledge and rule sets as client-owned assets, and retain the code under an open-core guarantee so the exit path stays open.
Answer the third question first: who maintains it in year three.
Twelve months of internal build against 45 to 80 days to production
See Use Cases
Seven Mistakes That Kill Enterprise AI Agent Programmes
1. Starting with the model instead of the process - vendor selection before process mapping produces a shortlist optimised for demos.
2. Letting the model make governed decisions - the moment a model decides whether to approve a PPAP submission or a claim, the audit story collapses.
3. Piloting on synthetic data - a sandbox result says nothing about messy intake, timing-out legacy systems or the exception patterns that dominate real volume.
4. Counting connectors instead of testing depth - a marketplace number is a marketing figure. Ask for a live run inside your hardest system.
5. Deferring governance to phase two - retrofitted controls fail audits and cost more than building them in from the first workflow.
6. Automating the exception first - the 20% of cases needing judgement look like the interesting problem. They are the wrong place to prove value.
7. Skipping the baseline - without pre-deployment numbers, a successful deployment becomes unprovable and the second use case never gets funded.
Everything You Need to Know About Enterprise AI Agents
Decision | What It Looks Like in Practice |
Which process should go first? | High volume, three or more systems per case, written policy covering most cases, and a cost baseline you can already measure. Supplier documents, invoices and complaints qualify in most organisations |
What has to be proved before contract? | A run on your live systems and your data, against criteria agreed upfront, rather than a demo on synthetic records |
Who writes the business rules? | Your teams. The rules are mapped from your SOPs, product terms and contractual deadlines, and remain client-owned assets |
What stays with humans? | Judgement work. Engineering root cause analysis, clinical decisions, on-site audits, and anything falling below the confidence floor you set |
How is the first deployment scoped? | Prioritised business case, proof on your systems within weeks, then production and extension into adjacent processes |
What does the IT effort look like? | Deployment engineering is handled by the vendor team. No API modernisation, middleware project or re-platforming as a prerequisite |
Where does the software run? | Self-hosted in your cloud, hybrid with the agent inside your cloud next to SAP, or on-premises alongside SAP on your own servers |
How does commercial scaling work? | Monthly licence with included operations volume, scaling on usage and deployment complexity. No per-seat, per-token or outcome-based pricing |
How do you prove it worked? | Automation rate, precision on live data, per-case cost against the pre-deployment baseline, and time to production |
When does use case two start? | Once the infrastructure is operational. Later workflows are predominantly platform spend and run at 85–90% margin |
What happens if you leave? | Under an open-core guarantee, you keep all code and binaries and the deployed workflows keep running |
What would make this fail? | No baseline, no codified rules, governance deferred to phase two, or a first process chosen for visibility rather than measurability |
Take the Next Step With Noxus
Noxus runs enterprise AI agents in production inside the systems other vendors decline to touch.
No model decides whether a claim is approved or a PPAP submission passes. Business logic mapped from your own SOPs handles every governed decision, and every step replays under audit.
It runs on the stack you already operate: SAP ECC and S/4HANA, SAP QM, MM and PP, Guidewire, Oracle, ServiceNow and COBOL-era cores, through 400+ connectors, from reading the input to writing the outcome into your system of record. Deployment is self-hosted, hybrid or on-premises, with bring-your-own-key encryption, lineage exportable for regulators, and an open-core guarantee that leaves you holding the code if the relationship ends.
Noxus is built for quality directors, operations leaders, IT architects and CFOs in manufacturing, financial services, insurance, healthcare and retail who are carrying pilots that never reached production, a BPO renewal with escalation baked in, or a board mandate on operational cost. KIRCHHOFF Automotive, Santander, Carlsberg Group, CUF and Jerónimo Martins run operations on it today, with zero client churn to date.
The fastest way to know whether this works for your operation is to watch one of your own processes run end to end, on your systems and under your policy, with the result measured against criteria you set beforehand.
Watch one of your processes run under your own policy and your own criteria.
KIRCHHOFF Automotive, Santander, Carlsberg Group and CUF run operations on it today
Book a Process Assessment
FAQs About Enterprise AI Agents
What are enterprise AI agents?
Enterprise AI agents are software systems that execute multi-step operational work across enterprise systems, with AI interpreting unstructured inputs and deterministic rules governing every decision. They read intake from email, documents, portals or EDI, query three to seven systems per case, apply codified policy and write outcomes back to systems of record under full lineage. Around 31% of enterprises run at least one agent in production, led by banking and insurance at 47%. Documented deployments reach 3–5x ROI at 90–96% precision on live operational data.
How long does it take to deploy enterprise AI agents?
Deploying enterprise AI agents takes 45 to 80 days from contract to live production on the client's actual systems, with multi-country rollouts completed inside 90 days. The sequence runs from prioritised business case, to proof on live systems within weeks, to production and extension into adjacent processes. Later use cases deploy in weeks because the integrations and governance controls are already operational. Building the equivalent infrastructure internally costs €500k to €1M+ and typically takes 12 or more months.
Can enterprise AI agents work with SAP and other legacy systems?
Enterprise AI agents can work with SAP and other legacy systems, and that is where the highest-value operational work sits. Noxus writes into SAP QM, MM and PP through OData where a modern interface exists, and operates the interface directly where none does, covering SAP ECC, S/4HANA, Guidewire, Oracle and COBOL-era cores across 400+ connectors. No API modernisation, middleware project or re-platforming is required beforehand. Deployment engineering is handled by the vendor team rather than the client's IT organisation.
What is the difference between enterprise AI agents and RPA?
The difference between enterprise AI agents and RPA is that agents read unstructured inputs and orchestrate across systems, while RPA replays fixed sequences against fixed interfaces. RPA bots break when an ERP interface updates and carry maintenance costs that frequently exceed the savings they were funded to deliver. Enterprise AI agents classify messy intake, apply codified rules across several systems and write outcomes back with a replayable trace, reaching 60% automation on complex complaint handling and above 95% on structured document processing. RPA remains a reasonable fit for stable, structured, single-system tasks.
How do enterprise AI agents handle data security and compliance?
Enterprise AI agents handle data security and compliance through hosting architecture rather than contractual assurance. Noxus runs self-hosted inside the client's cloud, hybrid with the agent sitting next to SAP, or on-premises, with bring-your-own-key encryption so data never leaves the client perimeter and no training on client data. Every action produces full lineage on user, policy, model and input, replayable and exportable for regulators. Certifications cover ISO 27001, SOC 2 Type II, GDPR and HIPAA, with ISO 27701 in progress and NIS2 controls in place.
What are the best use cases for enterprise AI agents?
The best use cases for enterprise AI agents are high-volume operations crossing several systems and following written policy for most cases. In manufacturing that means supplier quality document processing, 8D claim handling, PPAP review, supplier scoring and audit preparation against SAP QM and MM. Elsewhere it means invoice matching at 78% automation, merchant onboarding and KYC at 75% intake automation, mortgage complaint handling at 80%, and catalogue enrichment at 90% automated classification. Exception-heavy work makes a poor starting point and a strong second phase.
How much do enterprise AI agents cost?
Enterprise AI agents on the Noxus model are priced as a monthly licence with included operations volume, scaling with usage rather than headcount. There is no per-seat, per-token or outcome-based pricing, which keeps total cost predictable for board approval. Pricing scales on operational volume and deployment complexity, from a single managed workflow to multi-region deployments on dedicated infrastructure. First engagements include deployment engineering; later use cases are predominantly platform spend at 85–90% margin.
We have already tried AI pilots that failed. Why would this be different?
Failed AI pilots usually stop at integration and governance rather than at model capability, which is where a process intelligence layer starts. Gartner attributes the forecast cancellation of over 40% of agentic AI projects by end of 2027 to escalating costs, unclear business value and inadequate risk controls, none of which are model problems. Noxus proves the process on your live systems and data within weeks, against criteria set upfront, before write-back is enabled. Zero clients have churned to date, and production references at KIRCHHOFF Automotive, Santander and Carlsberg Group are available for comparable evaluations.
