Every company now has an AI demo. Far fewer have an AI system that has run at full volume for six months. MIT NANDA’s research found that around 95% of generative AI pilots had no measurable effect on P&L (Fortune). Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027 (Gartner). And while 88% of companies use AI in at least one function, McKinsey finds only 39% report any EBIT impact (McKinsey).
The encouraging part: production is clearly achievable. AI deals reach production in 47% of cases, against 25% for traditional SaaS (Menlo Ventures), and 70% of service teams using AI agents saw a measurable effect within 60 days (Salesforce). What kills pilots is rarely the technology. It is how the pilot is set up. This article shows the typical failure patterns and a five-stage path that gets a pilot into production.
Why AI pilots stall: eight patterns
| Pattern | What it looks like | What fixes it |
|---|---|---|
| No metric, no owner | “Let’s see what AI can do for us” | One business metric, one accountable owner, exit criteria signed before day one |
| Old process, new tool | AI bolted onto an unchanged workflow | Redesign handoffs and roles; high performers are about 3× more likely to do this (Silicon Canals on McKinsey) |
| Quality nobody can measure | Cherry-picked demo examples, anecdotes | A test set of real cases and automated evals; quality is the number one barrier for 32% of teams, latency second at about 20% (LangChain report) |
| Missing integrations | Staff copy answers from the AI into the CRM | Read and write access via APIs; integrations take 40–60% of a RAG budget (ment.tech) |
| Unclear costs | A token bill nobody forecast | Cost per outcome, budgets and hard caps |
| Data protection and legal doubts | Legal blocks the rollout in week six | DPA, hosting decision, DPIA and AI disclosure planned upfront |
| People | Staff ignore or work around the system | Early involvement, training and a clear new role for humans |
| No end date | The pilot runs “until we know more” | A fixed time box and a go/no-go meeting in the calendar |
The European numbers explain why legal and cost questions belong at the start, not at the end. Among German companies already using AI, 66% name data protection requirements as a barrier, 56% legal uncertainty, 54% unpredictable costs, 45% the risk of data leaks and 42% resistance from employees (Bitkom). Every one of these can stop a technically successful pilot.
What “production” actually means
Most disagreements between business and IT come from different ideas of “done”. This comparison makes the gap visible:
| Dimension | Pilot | Production |
|---|---|---|
| Scope | A sample, friendly users | Real volume, every customer in scope |
| Data | An export or snapshot | Live, permission-aware access |
| Integrations | Manual or read-only | Read and write via APIs, tested failure modes |
| Quality | Demo examples | Eval set and regression tests on every change |
| Monitoring | None | Dashboards, alerts, sampled reviews |
| Humans | Watching | Defined escalation and approval points |
| Cost | A credit card | Budget, caps, reported cost per outcome |
| Security | Trusted inputs | Prompt-injection defences, least-privilege access |
| Compliance | Postponed | DPA, retention rules, AI disclosure (Art. 50 EU AI Act) |
| Ownership | Innovation team or vendor | Named business and technical owners, SLA |
Since 2 August 2026, chatbots and voice agents used in the EU must tell people that they are talking to an AI (AI Act Article 50). That alone is a production requirement many pilots never considered.
A five-stage path from pilot to production
Stage 1: Audit (1–2 weeks)
Pick the process with the clearest unit and the highest volume, not the most impressive demo. Measure the baseline: volume, cost per unit, error rate, response time. Calculate a first business case with conservative assumptions — our guide on how to calculate AI automation ROI walks through the formulas.
Stage 2: Discovery and exit criteria
Before anybody builds, settle four things: data access, the integration map (which systems the AI reads and writes), a test set of real historical cases, and the exit criteria. A good exit-criteria card fits on half a page. For an order-status agent in e-commerce, an illustrative version could read:
- Primary metric: at least 45% of order-status tickets resolved without a human and not reopened within 7 days.
- Guardrails: customer satisfaction not below baseline; at least 95% correct escalations on the test set; cost per resolution below the agreed ceiling.
- Decision: the head of customer service signs go or no-go on day 28.
The numbers are yours to set. What matters is that they exist before the first line of code.
Stage 3: A paid, time-boxed pilot (2–4 weeks)
The market is used to free trials: Intercom Fin offers 14 days (Fin), HubSpot 28 days (Resolve247). A paid pilot therefore has to deliver more than a trial: a real integration with your systems, your data and a written impact report against the baseline. Start in shadow mode — the AI drafts, humans decide — so you can measure quality without customer risk. Then let it handle a small share of real traffic.
Build the evaluation before you build the agent. A test set of real cases, scored automatically, tells you whether a prompt or model change made things better or worse. Without it, every tweak is a guess. More on that in our guide to AI evals.
Stage 4: Launch with a staged rollout
Move from shadow mode to, for example, 10% of traffic, then 50%, then 100%, with a rollback switch at every step. Define human handoff explicitly: which intents always go to a person, which confidence threshold triggers escalation, how the customer reaches a human. Large companies do the same: 49% have explicitly defined scenarios in which an agent may not decide on its own, and 75% of banks do not let agents touch sensitive data without human control (KPMG).
Put a price tag on every outcome before launch. KPMG reports that 74% of companies now include a cost review in their AI approval process and 43% run explicit token budgets. Set a monthly cap and an alert well below it.
The go/no-go meeting deserves its own agenda. Keep it to 45 minutes and three questions: did the primary metric reach the threshold on real traffic; did any guardrail break (satisfaction, escalation quality, complaints, cost per outcome); and what has to change before full volume? Possible outcomes are “go”, “go with conditions” (for example, keep refunds above a limit with humans), “extend once by two weeks with a specific fix” or “stop”. A clean stop is a result, too: it frees budget for the next candidate process and you keep the baseline, the test set and the integration work.
Stage 5: Care — production is a service, not a date
After launch, the real work starts: monitoring, monthly eval runs, model updates, knowledge-base maintenance and cost reports. Knowledge bases go stale, products change, models get deprecated. A system nobody maintains degrades quietly until someone switches it off. Plan this stage as an operating budget, the way you would for any other critical system — see AI Ops after launch.
Build, buy or partner?
How you source the solution has a measurable effect on the odds. In 2025, 76% of enterprise AI use cases were bought rather than built, up from 53% a year earlier (Menlo Ventures). MIT NANDA found that purchased solutions and partnerships with specialised vendors succeeded in about 67% of cases, while internal builds succeeded far less often (Fortune).
Practical consequence: first check whether the agent built into your helpdesk or CRM already covers the use case. Build or commission custom work where it does not — custom actions, deep integrations, data residency or a genuinely differentiating process. Our build vs buy framework gives you the decision criteria.
Be careful with scope. Fully autonomous agents that operate computers and browsers have improved quickly, but prompt injection remains unsolved; even OpenAI acknowledges it may never be fully eliminated (analysis). For production in 2026, narrow agents with clear permissions and a human on risky steps are the safe default.
Warning signs that your pilot is already stalling
If you are in the middle of a pilot, these signals mean it is drifting towards the 95% rather than towards production:
- Nobody can state the success metric in one sentence. Fix: write the exit-criteria card this week and get it signed.
- The demo works, but staff still copy and paste. Fix: prioritise the one write-integration that removes the manual step.
- Quality is discussed through anecdotes. Fix: build a test set from 200–300 real cases and score every change against it.
- Legal or the data protection officer has not seen it yet. Fix: book the review now; bring the data flow, the subprocessors and the planned AI disclosure text.
- Costs are “roughly a few hundred a month”. Fix: measure cost per outcome on real traffic and set a cap.
- The end date has moved twice. Fix: schedule the go/no-go meeting and decide with the data you have.
Most stalled pilots can be rescued in two to four weeks if someone takes ownership of these points. If none of them can be fixed, stopping is cheaper than waiting.
Typical mistakes
- Choosing the most impressive use case instead of the most measurable one. A modest process with a clear unit beats a spectacular one nobody can evaluate.
- Piloting on clean demo data. Real inboxes contain scans, typos, forwarded threads and three languages. Test on those.
- Treating the pilot as free. Free pilots get free attention. A paid, time-boxed pilot creates commitment on both sides.
- No end date and no decision meeting. Put the go/no-go date in the calendar on day one.
- Measuring only at the end. Track the primary metric weekly; surprises in week four are cheaper than surprises in month six.
- Forgetting the team. People who were not involved will work around the system. Train them and give them a clear role, such as handling escalations and reviewing quality.
- Over-automating risky steps. Refunds, legal statements and medical or financial advice need approval points.
- No owner after launch. Without a named owner and a care budget, quality decays and the project is quietly cancelled.
Conclusion
Getting from AI pilot to production is less about better prompts and more about management basics: one metric, one owner, real integrations, measurable quality, controlled costs and a staged rollout. Decide what “done” means before the pilot starts, and budget for the operating phase from the beginning. Do that, and your pilot becomes a system that keeps paying back instead of a slide in last year’s strategy deck.
This is how we work in Lotus Libre Agents and Lotus Libre Systems: audit, discovery, a paid pilot with signed exit criteria, launch and care. If you have a pilot that is stuck — or one you want to set up properly — book a free 15-minute audit and we will tell you what it would take to get it into production.