Start with the conclusion: taking an enterprise AI agent from POC to production usually takes 2–4 months, not two weeks. Processes that cross systems and write back to core business often take six months or more. The deciding factor is not the model. It is data, integration, acceptance, and operations.
The 2026 market is hot, but results are uneven. Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value, or weak risk controls. In IDC’s 2026 research, 80% of organizations have started deploying agentic AI, but only 17% use it enterprise-wide. This article sets out a practical five-stage path, POC design, seven acceptance metrics, three hidden costs, and a supplier contract checklist. Timelines, budgets, and metrics are editorial recommendations, not commitments. Research cut-off: 24 October 2026.
3–5 real tasks, 200–500 evaluation cases, 2–4 months to production. Without real tasks and an evaluation set, a polished demo is not a procurement basis. Without integration and operations budget, you will rework after launch.
Why do so many agent projects exist in 2026, yet so few reach production?
Pilots are easy; scale is hard. IDC finds that 80% of organizations have started deployment and only 17% use agents enterprise-wide. The blocker is not the model. It is data, integration, cost, and governance.
| Evidence | Key number | What it means for the project |
|---|---|---|
| IDC 2026 survey | 69% of organizations see a significant gap between AI spending and measurable financial results; 73% treat excessive AI cost as a major risk | “Launched” is not “returning value”; define the financial measure at project approval |
| IDC 2026 survey | 80% have started deploying agentic AI; only 17% use it enterprise-wide | Most projects stall at pilot; scale requires process and organizational change |
| IDC 2026 survey | 34% of buyers are dissatisfied with limited integration with existing systems, opaque models and data, or vendor lock-in; 31% are dissatisfied with overpromising or unclear pricing | Integration, transparency, and contract terms are procurement priorities, not add-ons |
| Gartner forecast | By the end of 2027, more than 40% of agentic AI projects will be canceled, due to cost, unclear value, and insufficient risk controls | This is a forecast, not a statistic, but it supports “start small, then scale” |
| MIIT and seven other departments | By 2027, 1,000 high-level industrial agents, 500 typical application scenarios, and 1,000 benchmark enterprises | Policy is pushing supply and scenarios but does not guarantee any single project |
Note: IDC data comes from a public article dated 20 October 2026; Gartner data from a 25 June 2025 press release and later public articles; MIIT data from the “AI + Manufacturing” special action implementation opinion (MIIT Joint Science [2025] No. 279).
The China Academy of Information and Communications Technology (CAICT) reached a similar conclusion in its June 2026 report on enterprise agent technology and applications. Scaling enterprise agents requires model adaptation, memory, tool calling, planning, multi-agent collaboration, and private deployment. The common bottlenecks are data silos, fragmented knowledge bases, hallucination, unclear human-machine responsibility, and a lack of audit and permission controls. In other words, projects usually fail not because the model is not smart enough, but because business systems are not connected, no one owns the data, and acceptance criteria were never defined.
What stages does POC-to-production usually pass through?
Five stages: scenario initiation, POC, pilot, production, and scale. Initiation to production usually takes 2–4 months; scale proceeds quarter by quarter.
| Stage | Suggested duration | Key actions | Deliverables | Gate to the next stage |
|---|---|---|---|---|
| 0. Scenario initiation | 1–2 weeks | Business owner, 3–5 real tasks, data boundary, human baseline, budget | Project charter, data and system inventory | Business owner in place, data usable, budget and acceptance owner clear |
| 1. POC | 2–4 weeks | 200–500 evaluation cases, API or private baseline, error taxonomy, cost and latency records | POC report, evaluation set | Task success rate, cost, and latency meet the pre-agreed thresholds |
| 2. Pilot | 4–8 weeks | 1–2 teams, read-only integration, permissions and audit, training and operations | Pilot report, issue list | Adoption, human handover rate, and incident rate meet target |
| 3. Production | 4–12 weeks | Write-back and transactions, dual-node redundancy and degradation, monitoring and alerts, on-call, security review | Live system, operations manual | SLA, audit, and rollback drill passed |
| 4. Scale | Quarterly | Replicate new scenarios, update models and tools, review cost, roll out across the organization | Quarterly review, roadmap | Positive ROI, stable operations, controlled risk |
Note: durations are editorial advice. Adjust for data quality, the number of interfaces, compliance level, and team capability. They are not a delivery commitment for any project.
The most common errors are skipping the pilot and going straight to production, or buying the full platform as soon as the POC ends. A POC proves that a task is feasible under controlled conditions. A pilot proves that real users will use it and the system holds up. Production proves that incidents can be absorbed. The three stages solve different problems and cannot replace one another.
How big should a POC be to count?
3–5 real tasks, 200–500 evaluation cases, one baseline, and an error taxonomy. Replacing a POC with a demo is the root of later rework.
| Element | Minimum requirement | Why |
|---|---|---|
| Real tasks | 3–5, drawn from tickets, chat logs, reports, or production logs | Demo questions are hand-picked; real tasks expose the long tail and dirty data |
| Evaluation set | 200–500 cases, stratified by intent, difficulty, and risk, with reference answers | Without an evaluation set, “good” is a matter of opinion |
| Baseline | Human handling, the existing system, or an API approach, with time, cost, and error rate recorded | Without a baseline, you cannot show whether AI is better or more expensive |
| Error taxonomy | Retrieval errors, tool-call errors, format errors, policy or permission errors, hallucination | Different errors need different fixes; a single bucket cannot be improved |
| Cost and latency | Record tokens, tool calls, human review, and P50/P95 latency | A demo does not show the bill; production multiplies it by volume |
| Security and audit | Data boundary, permissions, logs, human confirmation points | High-risk actions cannot wait until after launch for audit |
Note: a POC is not a feature list. It is the smallest closed loop that proves one task can be completed reliably with real data. A POC without a business owner’s sign-off should not proceed to procurement.
Two formulas are worth fixing on day one of the POC: task success rate = successful tasks ÷ total tasks; cost per task = (model + tools + human review + allocated operations) ÷ successful tasks. Looking only at whether an answer sounds right hides cost and human handover.
How should acceptance metrics be defined?
Look at task success rate, closed-loop rate, human handover rate, cost per task, P95 latency, and audit coverage together. Accuracy alone leads to the wrong decision.
| Metric | Definition / formula | Suggested starting line (example) | Note |
|---|---|---|---|
| Task success rate | Successful tasks ÷ total tasks | Assisted scenarios ≥80%; high-risk ≥95% with human review retained | Define “success” before discussing percentages |
| Closed-loop rate | Tasks completed without human help ÷ total tasks | ≥60% in the first phase; ≥80% when mature | Closed loop is not “finished answering”; system actions must complete |
| Human handover rate | Tasks handed to a human ÷ total tasks | Assisted scenarios ≤20%; high-risk work can be 100% reviewed | Handover is not failure; it is part of risk control |
| Cost per task | (Model + tools + human + operations) ÷ successful tasks | Efficiency scenarios ≤50–70% of the human baseline | For revenue or risk scenarios, cost is not the only lens |
| P95 latency | End-to-end time for 95% of requests | Conversation ≤5 seconds; batch processing ≤30 seconds | Long context and tool chains raise latency significantly |
| Audit coverage | Tool calls with a complete event or trace ÷ total tool calls | 100% | Key actions must trace to person, time, input, and output |
| Adoption | Weekly active target users ÷ total target users | ≥60% within four weeks of launch | If nobody uses it, a high success rate is worthless |
Note: the figures above are editorial starting lines, not industry standards, and do not apply to every scenario. High-risk and heavily regulated businesses should set stricter targets and retain human confirmation.
Do not write acceptance as “90% accuracy” alone. Break it into task success rate, tool-call accuracy, format compliance, human handover rate, P95 latency, cost per task, and audit coverage. Leave one dimension out, and a vendor can show you an “optimization” in another.
Which three costs most often overrun?
System integration, data governance, and post-launch operations. They usually exceed the platform subscription and are the easiest to omit at project approval.
| Cost item | Typical share | Common omission | How to control it |
|---|---|---|---|
| System integration | 30–50% | Legacy systems without APIs, write-back and transactions, field-level permissions, single sign-on | Inventory interfaces first; prefer non-invasive integration; put interface changes in the contract |
| Data governance | 10–20% | Document versions, master data, knowledge-base permissions, evaluation set | Assign a data owner; build the 200–500-case evaluation set before expanding the knowledge base |
| Post-launch operations | 15–30% per year | Model upgrades, plugin changes, monitoring and alerts, security checks, on-call | Budget over three years; have at least 1–2 administrators or buy managed support |
| Platform and model | 20–40% | Subscription, tokens, GPUs, dedicated resources | Run a POC before buying; estimate from actual usage, not demo peak |
Note: the shares are experience from manufacturing agent projects and vary with the number of systems, data quality, and compliance requirements. They are not a formal quote from any vendor. For a fuller TCO breakdown, see What private AI deployment really costs in 2026.
If the budget covers only the platform and not integration and data, the project will probably stop at the demo. Conversely, connecting 1–2 high-value processes first and expanding gradually usually keeps first-year total investment more controllable.
Which projects should not start at all?
No business owner, no data owner, no evaluation set, no integration budget, demo only. Do not start.
| Red flag | Why it is dangerous | What to fix first |
|---|---|---|
| Only IT leads; the business is not involved | Requirements become a technology demo that nobody uses | Name a business owner and a process owner |
| No real tasks or logs, only demo questions | A demo cannot represent the long tail or dirty data | Pull 3–5 real tasks from tickets, chats, or reports |
| Data sits on personal computers or cloud drives with no permissions | The knowledge base cannot stay current; compliance risk is high | Classify and tier data and set access permissions first |
| Legacy systems have no API and there is no integration budget | At launch people copy data by hand; value drops sharply | Estimate the cost of the interface, middleware, or RPA work before deciding |
| Demanding “95% accuracy guaranteed” without defining the task | It cannot be verified at acceptance; vendors can only overpromise | Write task, data, risk, and human review into the acceptance criteria |
| Buying a platform with no operator | The platform becomes an asset that was bought but never used | Name an operations owner and a quarterly iteration mechanism |
Note: these red flags are unrelated to choosing Dify, Coze, Alibaba Cloud Bailian, or LingHarness. Switching platforms will not fix business, data, or organizational problems.
How should you choose a vendor and sign the contract?
Ask for a POC on real tasks, explicit acceptance metrics, data boundaries, and exit and migration terms. Do not decide from a demo and a price.
| Clause | What to ask | Why |
|---|---|---|
| POC acceptance | Which real tasks, how many evaluation cases, what counts as pass, who signs off | Avoid “great demo, poor production” |
| Data boundary | Can it run offline or on the intranet, where logs are stored, deletion mechanism, external API calls | Data compliance is a hard constraint, not an after-sales issue |
| System integration | How ERP, MES, and OA connect, write-back support, field-level permissions, legacy systems without APIs | Integration determines whether the project truly closes the loop |
| Acceptance and SLA | Task success rate, P95 latency, incident response, support hours, escalation | Turn “good results” into measurable, accountable terms |
| Exit and migration | Can data, models, prompts, tools, and the knowledge base be exported; is source code delivered; migration cost | Avoid single-vendor lock-in |
| Billing | How platform, model, integration, change requests, and renewals are priced; minimum spend | A cheap platform can recover margin on integration and change |
Note: “results” in a contract must be measurable. Do not accept promises such as “accuracy guaranteed” that cannot be verified at acceptance. Have legal counsel review legal clauses.
If you are building agents for the factory floor (scheduling, work reporting, quality inspection, traceability), start by mapping the scenario to the LingMOE AI manufacturing platform. For office collaboration, approvals, and documents, see the LingWork office agent. For a private execution engine, model independence, and audit, see the LingHarness AI foundation. The product is only the vehicle; confirm the scenario and data boundary first.
How do you prove it is useful after launch?
Use weekly and monthly metrics: adoption, task closure, cost per task, human handover, and incident rate, compared with the human baseline. Call volume is not ROI.
| Metric | Frequency | Purpose |
|---|---|---|
| Weekly active users / adoption | Weekly | Whether anyone actually uses it, or it was busy only on launch day |
| Task success / closed-loop rate | Weekly | Whether quality is stable, or it only works on easy tasks |
| Human handover / tickets | Weekly | Degree of automation and the cost of human fallback |
| Cost per task | Monthly | Combine tokens, tools, human, and operations into ROI |
| P95 latency / incident rate | Weekly | Experience and stability; avoid “it works but is unusable” |
| Audit / permissions | Monthly | Confirm sensitive actions are traceable and permissions have not drifted |
| Business results | Quarterly | Delivery cycle, labor hours, error rate, revenue and collections |
Note: do not use token consumption, generated characters, or call counts as ROI. They show usage, not value.
If adoption is still low after four weeks, do not add features. Check whether the task is wrong, the entry point too deep, permissions blocking, or answers untrustworthy. Most complaints that “the AI is not useful” come down to process and permissions.
FAQ
How long does an AI agent POC usually take?
Usually 2–4 weeks. If legacy systems have no API and data needs cleaning, 4–6 weeks is more realistic. The point of a POC is not how many features it has, but whether 3–5 real tasks and 200–500 evaluation cases prove it is usable.
Does an enterprise agent have to be privately deployed?
Not necessarily. First check whether data can leave the intranet, whether ERP/MES/OA integration is required, and whether there is an operations team. When data is not sensitive and the scenario is unproven, a cloud API or SaaS is usually more cost-effective.
Is 200 evaluation cases enough?
It is enough as a starting point for a single-scenario POC. For multiple scenarios, high risk, or coverage of long-tail questions, 500 or more is advisable, stratified by intent, difficulty, and risk.
Can a vendor guarantee 90% accuracy?
We do not recommend accepting that promise. Accuracy depends on task definition, data, tools, and human review. The contract should specify the evaluation set, the acceptance process, and what happens if the target is missed, not a number detached from the scenario.
What is a typical first-year budget for an enterprise AI agent?
A POC commonly costs RMB 50,000–150,000; pilot to production commonly costs RMB 300,000–1,000,000; post-launch operations run at about 15–30% of the build investment per year. These are editorial estimates, not quotes, and depend on integration, data, and compliance requirements.
How soon will results show after launch?
Assisted scenarios usually show changes in adoption and handling time within 4–12 weeks. Processes that cross systems and write back to core business often take 3–6 months. Look at task closure, human handover, and cost per task first, not call volume alone.
AI agent success is not decided by the model. It is decided by scenario, data, integration, acceptance, and operations: use 3–5 real tasks and 200–500 evaluation cases for the POC, prove with a pilot that people will use it, and only then discuss production and scale.
If you are building the 2027 budget, see our 2026 manufacturing AI outlook for policy targets, budget flows, and procurement timing. For operations after launch, see AI agent operations metrics on adoption, cost per task, and human handover. For decision authority, classification, and procurement compliance, see our analysis of the agent implementation opinion.
Sources and caveats
Research cut-off: 24 October 2026. Timelines, budgets, and acceptance thresholds are editorial recommendations based on public reports and project experience. They are not industry standards or delivery commitments. Institutional data is quoted on its original basis.
- IDC, “The Enterprise AI Deal Has Changed. Here’s What Vendors Need to Prove.” (20 October 2026). idc.com. Includes the 69% spending-versus-results gap, 80% deployed versus 17% enterprise-wide, and 73% cost risk.
- Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027” (25 June 2025). gartner.com.
- Ministry of Industry and Information Technology and seven other departments, “‘人工智能+制造’专项行动实施意见” (Special action implementation opinion on “AI + Manufacturing”), MIIT Joint Science [2025] No. 279. nda.gov.cn.
- China Academy of Information and Communications Technology, “企业级智能体技术与应用研究报告（2026年）” (Research report on enterprise agent technology and applications, 2026), 25 June 2026. caict.ac.cn.