Enterprise agents are moving into production, but the highest-value deployments are narrower and more controlled than the market narrative suggests. Production evidence in 2026 points toward specialised agents embedded in specific workflows, with limited action horizons, explicit tools and permissions, human escalation, strong observability and outcome-based measurement. The central scaling constraint is not model intelligence alone; it is the redesign of work, data, architecture, evaluation and governance around an agent that can act.
Research question and scope
Core question: Where are AI agents genuinely operating in production, what patterns distinguish scaled value from stalled pilots, and what should enterprises build or govern differently?
Scope: Global enterprise use across software engineering, service operations, finance, sales, supply chain, healthcare/insurance and knowledge work. The report distinguishes vendor-labelled “agents” from systems that plan or act across multiple steps and tools.
Method: Triangulation of global surveys, production-practitioner research, public deployment studies, proprietary enterprise usage datasets and current case evidence. Claims about “production” are separated from future expectations and marketing language.
1. The 2026 reality: agents are scaling, but unevenly
McKinsey’s August 2026 global survey finds that 40% of respondents from organisations with more than $1 billion in annual revenue report scaling AI agents, up from 27% a year earlier; smaller organisations remain at 22%. About one in five organisations are scaling coding agents, rising to 31% among large enterprises. S1 Deloitte’s August 2026 research finds only 15% of organisations have scaled orchestrated, cross-functional multi-agent adoption, despite very high expectations for the next four years. S2
The gap between these figures is useful rather than contradictory. “Scaling an agent in at least one function” is a lower threshold than cross-functional multi-agent orchestration. The evidence therefore describes a market that has crossed from pilots into real deployment, but not yet into pervasive autonomous enterprise operations.
| Evidence | Latest result | Interpretation |
|---|---|---|
| Large-enterprise scaling | 40% scaling agents in at least one function; 31% scaling coding agents. S1 | Production has become material in large firms. |
| Cross-functional orchestration | 15% scaled orchestrated multi-agent adoption. S2 | Enterprise-wide agent systems remain early. |
| Governance maturity | Only 21% report mature agentic governance. S3 | Guardrails lag deployment. |
| Production design | 68% of surveyed production agents execute at most 10 steps before human intervention; 74% rely primarily on human evaluation. S4 | Real agents are generally bounded and human-supervised. |
| Public deployment study | Gartner analysed 107 intermediate/advanced agent deployments; highest ROI concentrated in specialised domain agents. S5 | Specialisation beats generic autonomy for near-term ROI. |
| Enterprise usage depth | Frontier firms generate 8.3× output tokens per active user versus typical firms; agentic work is spreading beyond engineering. S6 | Depth of workflow integration, not access, is creating a frontier gap. |
2. What a production agent actually looks like
The strongest research challenge to the “autonomous digital employee” narrative comes from IBM Research’s Measuring Agents in Production study. Based on 20 case studies and a survey of 306 practitioners across 26 domains, it finds that successful production agents are built with comparatively simple, controllable techniques: 68% run no more than ten steps before human intervention; 70% primarily prompt off-the-shelf models rather than fine-tuning weights; and 74% rely primarily on human evaluation. Reliability is the dominant development challenge. S4
- Short action horizons reduce compounding error and make evaluation tractable.
- Off-the-shelf frontier models are often sufficient when workflow context, tools and controls are strong.
- Human evaluation remains central because enterprise success is contextual: correctness is necessary but not sufficient.
- Systems-level design—state, permissions, retries, validation, escalation and monitoring—matters as much as prompt or model choice.
3. The use cases moving fastest
Production adoption is clustering where work is digital, repetitive enough to structure, valuable enough to justify integration, and bounded enough to control. McKinsey’s 2026 survey shows software engineering as the clearest scaled use case, but agent deployments are spreading by industry: technology firms toward engineering, consumer/retail toward marketing and sales, and advanced manufacturing toward supply chain and production. S1 OpenAI’s enterprise usage data show the fastest growth in enterprise Codex users outside engineering in legal, sales, recruiting and marketing, albeit within OpenAI’s own customer base and therefore not a market-wide benchmark. S6
| Domain | Production pattern | Why it fits agents | Primary control problem |
|---|---|---|---|
| Software engineering | Issue-to-code, testing, review preparation, maintenance | Digital tools, measurable artifacts, reversible changes | Code quality, security, technical debt, reviewer load |
| Customer/service operations | Case triage, resolution steps, knowledge retrieval, follow-up | High volume, repeatable workflows, clear escalation | Incorrect action, customer harm, policy compliance |
| Finance/insurance | Claims, document review, reconciliations, exception handling | Rules + documents + systems + audit need | Regulation, traceability, false decisions |
| Sales/marketing | Account research, campaign operations, CRM actions | Many tools and context-rich repetitive tasks | Brand, privacy, revenue attribution |
| Supply chain/manufacturing | Inventory/parts workflows, diagnostics, scheduling support | Operational rules and structured data | Safety, real-world side effects, data latency |
| Legal/knowledge work | Research, drafting, document workflows | High information load and tool use | Privilege, source accuracy, accountability |
4. ROI comes from execution inside a workflow
Gartner’s September 2026 synthesis of 107 public deployments argues that tangible ROI is most likely to come from specialised domain agents. Examples include healthcare claims review, workers’ compensation review, prior authorisation, manufacturing analysis and parts replenishment; one industrial-services deployment cited by Gartner returned 90,000 technician hours and approximately $3 million in annual ROI. S5 These case figures are not population averages, but they illustrate a repeatable principle: value appears when the agent completes measurable work inside an operating process rather than simply answering questions.
This is consistent with McKinsey’s September 14 blueprint: CFOs and boards are moving past experimentation and asking for explicit returns; scaling requires redesigning the underlying work, not merely adding agents to existing steps. S7
Sources and notes
- McKinsey, The state of AI in 2026: On the road to ROI (25 Aug 2026). [Source](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai?os=av.)) — 1,719 participants across 97 nations; fielded May–June 2026.
- Deloitte, AI readiness gap and agentic success survey (12 Aug 2026). [Source](https://www.deloitte.com/us/en/about/press-room/deloitte-survey-examines-ai-readiness-agentic-ai-success.html)) — Executive survey on process readiness, autonomy expectations and cross-functional scaling.
- Deloitte, Agentic AI is scaling faster than guardrails (24 Apr 2026). [Source](https://www.deloitte.com/us/en/insights/topics/emerging-technologies/ai-agents-scaling-faster.html)) — 3,235 IT/business leaders in 24 countries.
- IBM Research, Measuring Agents in Production (ICLR 2026). [Source](https://research.ibm.com/publications/measuring-agents-in-production)) — 20 case studies; 306 practitioners across 26 domains.
- Gartner, Mastering Agentic AI: Multimillion-Dollar ROI Lessons (9 Sep 2026). [Source](https://www.gartner.com/en/articles/agentic-ai-roi)) — Analysis of 107 public agentic deployments.
- OpenAI, Enterprise Signals: What frontier firms are doing differently (12 Aug 2026). [Source](https://openai.com/signals/enterprise-data/)) — Proprietary enterprise-customer usage data; treated as vendor evidence.
- McKinsey, Stacking the odds: A blueprint for successfully scaling agentic AI (14 Sep 2026). [Source](https://www.mckinsey.com/capabilities/operations/our-insights/stacking-the-odds-a-blueprint-for-successfully-scaling-agentic-ai)) — Latest operating-model blueprint.
8i TECHNOLOGIES
Agents in Production
Where enterprise AI agents are actually scaling, what works, what fails, and the operating model required for measurable value
Reference date: 15 September 2026
See what matters next.
| Executive answer Enterprise agents are moving into production, but the highest-value deployments are narrower and more controlled than the market narrative suggests. Production evidence in 2026 points toward specialised agents embedded in specific workflows, with limited action horizons, explicit tools and permissions, human escalation, strong observability and outcome-based measurement. The central scaling constraint is not model intelligence alone; it is the redesign of work, data, architecture, evaluation and governance around an agent that can act. |
|---|
Research question and scope
Core question: Where are AI agents genuinely operating in production, what patterns distinguish scaled value from stalled pilots, and what should enterprises build or govern differently?
Scope: Global enterprise use across software engineering, service operations, finance, sales, supply chain, healthcare/insurance and knowledge work. The report distinguishes vendor-labelled “agents” from systems that plan or act across multiple steps and tools.
Method: Triangulation of global surveys, production-practitioner research, public deployment studies, proprietary enterprise usage datasets and current case evidence. Claims about “production” are separated from future expectations and marketing language.
1. The 2026 reality: agents are scaling, but unevenly
McKinsey’s August 2026 global survey finds that 40% of respondents from organisations with more than $1 billion in annual revenue report scaling AI agents, up from 27% a year earlier; smaller organisations remain at 22%. About one in five organisations are scaling coding agents, rising to 31% among large enterprises. [S1] Deloitte’s August 2026 research finds only 15% of organisations have scaled orchestrated, cross-functional multi-agent adoption, despite very high expectations for the next four years. [S2]
The gap between these figures is useful rather than contradictory. “Scaling an agent in at least one function” is a lower threshold than cross-functional multi-agent orchestration. The evidence therefore describes a market that has crossed from pilots into real deployment, but not yet into pervasive autonomous enterprise operations.
| Evidence | Latest result | Interpretation |
|---|---|---|
| Large-enterprise scaling | 40% scaling agents in at least one function; 31% scaling coding agents. [S1] | Production has become material in large firms. |
| Cross-functional orchestration | 15% scaled orchestrated multi-agent adoption. [S2] | Enterprise-wide agent systems remain early. |
| Governance maturity | Only 21% report mature agentic governance. [S3] | Guardrails lag deployment. |
| Production design | 68% of surveyed production agents execute at most 10 steps before human intervention; 74% rely primarily on human evaluation. [S4] | Real agents are generally bounded and human-supervised. |
| Public deployment study | Gartner analysed 107 intermediate/advanced agent deployments; highest ROI concentrated in specialised domain agents. [S5] | Specialisation beats generic autonomy for near-term ROI. |
| Enterprise usage depth | Frontier firms generate 8.3× output tokens per active user versus typical firms; agentic work is spreading beyond engineering. [S6] | Depth of workflow integration, not access, is creating a frontier gap. |
2. What a production agent actually looks like
The strongest research challenge to the “autonomous digital employee” narrative comes from IBM Research’s Measuring Agents in Production study. Based on 20 case studies and a survey of 306 practitioners across 26 domains, it finds that successful production agents are built with comparatively simple, controllable techniques: 68% run no more than ten steps before human intervention; 70% primarily prompt off-the-shelf models rather than fine-tuning weights; and 74% rely primarily on human evaluation. Reliability is the dominant development challenge. [S4]
- Short action horizons reduce compounding error and make evaluation tractable.
- Off-the-shelf frontier models are often sufficient when workflow context, tools and controls are strong.
- Human evaluation remains central because enterprise success is contextual: correctness is necessary but not sufficient.
- Systems-level design—state, permissions, retries, validation, escalation and monitoring—matters as much as prompt or model choice.
3. The use cases moving fastest
Production adoption is clustering where work is digital, repetitive enough to structure, valuable enough to justify integration, and bounded enough to control. McKinsey’s 2026 survey shows software engineering as the clearest scaled use case, but agent deployments are spreading by industry: technology firms toward engineering, consumer/retail toward marketing and sales, and advanced manufacturing toward supply chain and production. [S1] OpenAI’s enterprise usage data show the fastest growth in enterprise Codex users outside engineering in legal, sales, recruiting and marketing, albeit within OpenAI’s own customer base and therefore not a market-wide benchmark. [S6]
| Domain | Production pattern | Why it fits agents | Primary control problem |
|---|---|---|---|
| Software engineering | Issue-to-code, testing, review preparation, maintenance | Digital tools, measurable artifacts, reversible changes | Code quality, security, technical debt, reviewer load |
| Customer/service operations | Case triage, resolution steps, knowledge retrieval, follow-up | High volume, repeatable workflows, clear escalation | Incorrect action, customer harm, policy compliance |
| Finance/insurance | Claims, document review, reconciliations, exception handling | Rules + documents + systems + audit need | Regulation, traceability, false decisions |
| Sales/marketing | Account research, campaign operations, CRM actions | Many tools and context-rich repetitive tasks | Brand, privacy, revenue attribution |
| Supply chain/manufacturing | Inventory/parts workflows, diagnostics, scheduling support | Operational rules and structured data | Safety, real-world side effects, data latency |
| Legal/knowledge work | Research, drafting, document workflows | High information load and tool use | Privilege, source accuracy, accountability |
4. ROI comes from execution inside a workflow
Gartner’s September 2026 synthesis of 107 public deployments argues that tangible ROI is most likely to come from specialised domain agents. Examples include healthcare claims review, workers’ compensation review, prior authorisation, manufacturing analysis and parts replenishment; one industrial-services deployment cited by Gartner returned 90,000 technician hours and approximately $3 million in annual ROI. [S5] These case figures are not population averages, but they illustrate a repeatable principle: value appears when the agent completes measurable work inside an operating process rather than simply answering questions.
This is consistent with McKinsey’s September 14 blueprint: CFOs and boards are moving past experimentation and asking for explicit returns; scaling requires redesigning the underlying work, not merely adding agents to existing steps. [S7]
5. Why pilots stall
- Agent washing: relabelling a copilot or deterministic automation as an “agent” creates unrealistic expectations and weak measurement.
- Poor process definition: an agent cannot reliably own a workflow whose rules, exceptions and decision rights are unclear.
- Weak context and data: missing enterprise knowledge, stale records and fragmented identity create unreliable actions.
- Insufficient evaluation: offline benchmark scores do not establish business reliability under real permissions and exceptions.
- Permission sprawl: tool access is granted faster than least-privilege policies, monitoring and revocation are designed.
- No owner for the outcome: technology teams deploy an agent while the process owner continues to own neither metrics nor redesign.
- Cost variability: long loops, retries and high-context tool calls can make seemingly successful agents economically unattractive.
6. Governance is an operating capability
Deloitte’s multicountry survey of 3,235 technology and business leaders found only 21% reporting mature governance for agentic AI, while expected usage rises quickly through 2027. [S3] McKinsey’s AI-trust research reports security and risk as a leading barrier to scaling agents, with inaccuracy and cybersecurity among the most salient risks. [S8]
- Authenticate the agent, the sponsoring user/service and the target system.
- Define allowed tools, data, spend, actions and temporal scope before runtime.
- Preserve the decision trail: request, retrieved evidence, tool calls, state changes and approvals.
- Use deterministic policy checks around nondeterministic reasoning.
- Design explicit stopping points and human escalation for uncertainty, policy exceptions and high-impact actions.
- Measure error severity and recoverability, not only average task success.
- Version prompts, models, tools and policies so changed behaviour can be explained.
7. Build, buy or orchestrate?
The architecture choice is shifting away from “one platform or custom build” toward a layered decision. Packaged agents accelerate common workflows; custom orchestration is justified where proprietary process logic, data, integration or differentiation matter; and most enterprises will operate a portfolio. Gartner’s current guidance emphasises bounded agents for exception-heavy workflows, while McKinsey and Deloitte both stress process redesign and orchestration. [S9][S2]
| Choice | Best fit | Economic advantage | Primary risk |
|---|---|---|---|
| Packaged agent | Common horizontal process | Fast deployment, vendor-maintained integrations | Generic logic, lock-in, limited differentiation |
| Custom agent | Proprietary workflow / strategic capability | Process-specific value, control, differentiation | Engineering/evaluation burden |
| Agent + workflow/orchestration layer | Cross-system process with explicit controls | Balances reasoning with deterministic state and audit | Architecture complexity |
| Human + agent team | Ambiguous/high-stakes knowledge work | Leverages speed without surrendering accountability | Benefits depend on job redesign and adoption |
8. The production maturity model
| Level | Agent role | Control model | Success measure |
|---|---|---|---|
| 1 — Assist | Suggests/drafts; human executes | User review of every output | Quality and time saved |
| 2 — Bounded execute | Acts within one workflow with narrow tools | Rules + approval checkpoints | Accepted tasks / cycle time |
| 3 — Managed autonomy | Completes end-to-end cases within policy | Monitoring, exception routing, audit | Outcome rate / intervention / cost |
| 4 — Orchestrated agents | Multiple specialised agents coordinate | Shared state, identity, policy and observability | End-to-end process performance |
| 5 — Adaptive operating model | Work is designed around human-agent allocation | Continuous evaluation + governance | Business value, resilience, learning speed |
9. 8i assessment
| 8i view — TEST / SCALE SELECTIVELY Enterprise agents are ready for serious production in bounded, measurable workflows. General-purpose autonomy across consequential processes remains a WATCH proposition. The evidence supports selective scaling where the task boundary, data, tools, decision rights, evaluation and economics are explicit. |
|---|
10. What leaders should do now
- Create an agent portfolio around business workflows, not vendor products.
- Require an outcome metric and a human baseline before production funding.
- Define autonomy by action class: prepare, recommend, execute reversible, execute consequential.
- Instrument every agent for cost, success, intervention, error and rollback.
- Centralise identity, policy and observability even if agents are decentralised by business function.
- Treat process redesign as part of the implementation budget.
- Scale proven specialised agents before pursuing broad multi-agent autonomy.
Source register
Sources are prioritised toward primary institutions, transparent methodology, observable behavioural data, large-sample surveys and current production evidence. Vendor evidence is identified as such and used cautiously where it provides proprietary usage data or live implementation detail.
[S1] McKinsey, The state of AI in 2026: On the road to ROI (25 Aug 2026). Source — 1,719 participants across 97 nations; fielded May–June 2026.
[S2] Deloitte, AI readiness gap and agentic success survey (12 Aug 2026). Source — Executive survey on process readiness, autonomy expectations and cross-functional scaling.
[S3] Deloitte, Agentic AI is scaling faster than guardrails (24 Apr 2026). Source — 3,235 IT/business leaders in 24 countries.
[S4] IBM Research, Measuring Agents in Production (ICLR 2026). Source — 20 case studies; 306 practitioners across 26 domains.
[S5] Gartner, Mastering Agentic AI: Multimillion-Dollar ROI Lessons (9 Sep 2026). Source — Analysis of 107 public agentic deployments.
[S6] OpenAI, Enterprise Signals: What frontier firms are doing differently (12 Aug 2026). Source — Proprietary enterprise-customer usage data; treated as vendor evidence.
[S7] McKinsey, Stacking the odds: A blueprint for successfully scaling agentic AI (14 Sep 2026). Source — Latest operating-model blueprint.
[S8] McKinsey, State of AI trust in 2026 (25 Mar 2026). Source — Risk and trust barriers to agentic scaling.
[S9] Gartner, Bounded Agentic AI Products Scale Revenue the Fastest (1 Sep 2026). Source — Current guidance on bounded, human-in-the-loop agents.
[S10] Salesforce, Agentic Enterprise Index — second edition. Source — Proprietary platform usage data; useful for scale trends, not market share.
[S11] Deloitte + Google Cloud, Scaling Agentic AI to realize business value (29 May 2026). Source — Architecture, governance and organisational scaling framework.
[S12] OpenAI, From assistance to execution (12 Aug 2026). Source — Companion enterprise adoption synthesis; vendor evidence.