Free Website Audit: Discover what's holding your digital presence back
Short answer
To move an AI pilot to production, give it one business owner, connect it to the real systems and data, set a baseline and a target before launch, add human review and logging, and fund it from an operating budget. Most pilots stop on data, risk controls, cost or unclear value, not on the model.
The pilot showed that the model can do the task. It did not show that your company can run it every day, under a named owner, inside your ERP and CRM. This guide is for a COO, CIO or head of digital at a company of 50 to 500 people. It covers what production needs, who owns what, and a 90-day plan.
Why do most AI pilots stop before production?
Most AI pilots stop for business reasons, not technical ones. In July 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. It named four causes: poor data quality, weak risk controls, rising costs and unclear business value.
The model is not on that list. A pilot avoids all four by design: chosen data, no live risk, a trial budget and no return to prove. Production removes each of those.
Leaders feel this gap. In the 2024 Work Trend Index, Microsoft and LinkedIn found that 60% of leaders worry that their leadership lacks a plan and vision to implement AI.
The value gap is real too. In 2025, BCG reported that 60% of companies got hardly any material value from AI. Only 5% were “future-built”. A year later, BCG’s Applied AI Index 2026 put future-built firms at 7.5% and scaling firms at 41%. Nearly half now get real value. BCG’s Amanda Luther put the lesson plainly: “How you redesign work and reskill your workforce matters far more than how many jobs you cut.”
What does “ready for production” mean for an AI pilot?
A pilot is ready for production when it can fail safely, be measured, and be run by people who did not build it. Use this checklist before you approve the launch.
Production-readiness checklist
- [ ] One business owner is named. That person answers for results, not the IT team.
- [ ] A baseline exists: time per case, error rate, cost per case, measured before AI.
- [ ] A target is written down, with the date you will check it.
- [ ] The system reads live data from the source systems, not exports.
- [ ] Personal data is mapped: what goes to the model, where it is stored, who can see it.
- [ ] A person reviews outputs above a set risk level before they reach a customer.
- [ ] Every input, output and human decision is logged.
- [ ] There is a way to switch the AI step off and fall back to the manual process.
- [ ] Users are trained and know how to report a bad output.
- [ ] Running costs (model fees, hosting, support hours) sit in an operating budget.
- [ ] A review date is set to keep, change or stop the system.
Each item you cannot tick is your next piece of work.
Who owns the AI system after launch?
A business owner owns the outcome, and a technical owner owns the running system. Do not give both jobs to one team.
| Role | Owns | Typical person |
|---|---|---|
| Business owner | Target, budget, decision to keep or stop | Head of the department that uses it |
| Technical owner | Uptime, integrations, model and prompt changes | IT lead or your development partner |
| Reviewers | Human checks on flagged outputs | Staff who did the task before |
| Data owner | Quality and access rules for the source data | Owner of the ERP or CRM data |
The business owner also owns adoption. BCG found that seven in ten future-built and scaling companies already retrain their staff. For the people side, see our guide to an AI enablement program.
How do you connect AI to your ERP, CRM and data?
Connect the AI step through the APIs your systems already have, and keep the systems of record as the source of truth. The model reads from your ERP or CRM, writes a draft or a proposed action, and a rule or a person decides what is saved.
Integration is where most production effort goes. Production needs authentication, error handling, retries and a record of what changed. Our AI automation work is mostly this layer: we connect a model to the systems you run, add the review steps, and log each action so you can audit it.
For larger estates with several systems and strict access rules, plan the work as an enterprise integration project, not as an add-on to the pilot. If the next step is an agent that acts inside your ERP or CRM, read our comparison of building or buying AI agents first.
How do you measure AI ROI in the P&L?
Measure AI ROI against the baseline you set before launch, in the units the P&L already uses. Hours saved only count if they turn into lower cost, more output or faster cash.
Pick two or three measures:
- Cost per case — staff time plus running costs, divided by cases handled.
- Cycle time — from request to finished result.
- Error or rework rate — how often a person must correct the output.
- Volume per person — cases one person can handle per week.
Then move the system to an operating budget that pays for model fees, hosting, support and review time every month. If nobody will put it in their operating budget, it has no business owner yet.
What do the EU AI Act and BiH data protection law require at launch?
Two rules matter at launch for most mid-size companies: AI literacy under the EU AI Act, and data protection law. This article is not legal advice.
EU AI Act reach. Article 2 applies the Act to providers outside the EU that place AI on the EU market. It also applies to providers and deployers in a third country when the AI output is used in the EU. A company in Bosnia and Herzegovina that serves EU customers should check whether this covers it.
AI literacy. Article 4 says providers and deployers “shall take measures to support the development of AI literacy” of their staff. It does not set “any specific level”. This duty has applied since 2 February 2025.
High-risk dates. The European Commission confirms that the AI Omnibus entered into force on 27 July 2026. It moved the Annex III high-risk rules to 2 December 2027 and the Annex I rules to 2 August 2028. To check if your use case is high-risk, compare it with the Annex III list.
Data protection in BiH. N1 reported that the new Law on Personal Data Protection came into force on 8 March 2025, with practical application from 4 October 2025. It aligns BiH law with the GDPR, and fines for legal entities can reach 4% of global annual revenue. Map the personal data your AI system touches before launch.
What does a 90-day pilot-to-production plan look like?
A 90-day plan moves one pilot into daily use in three stages: prepare, launch to a small group, then scale and review.
| Days | Step | Owner | Output |
|---|---|---|---|
| 1–15 | Name the owners, measure the baseline, write the target | Business owner | One-page case with baseline and target |
| 1–30 | Map data and personal data, agree review rules | Data owner, technical owner | Data map, review thresholds |
| 16–45 | Build integrations, logging and the switch-off path | Technical owner | System on live data in a test setup |
| 31–45 | Train the first users and the reviewers | Business owner | Trained group, feedback channel |
| 46–60 | Launch to one team, review every flagged output | Reviewers | First two weeks of live data |
| 61–75 | Fix the top errors, adjust thresholds, check costs | Technical owner | Updated system, cost per case |
| 76–90 | Compare results with the baseline, decide | Business owner | Keep, change or stop decision |
Keep the scope to one process until day 90.
If you want a second view on your pilot, book a call. Bring the pilot results and the systems it must connect to. We will tell you what is missing before production.
Sources
- Gartner, Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025 — 29 July 2024
- Microsoft and LinkedIn, 2024 Work Trend Index: AI at Work Is Here. Now Comes the Hard Part — 8 May 2024
- BCG, The Widening AI Value Gap — 30 September 2025
- BCG, Applied AI Index 2026 press release — 30 September 2026
- EU AI Act, Article 2 (unofficial text) — Regulation (EU) 2024/1689
- AI Act Service Desk, Article 4: AI literacy — checked 28 October 2026
- European Commission, AI Act regulatory framework — updated 19 October 2026
- N1, Bosnia’s new personal data protection law introduces stricter rules — 14 October 2025
FAQ
How long does it take to move an AI pilot to production?
For one process with working APIs, plan about 90 days from the decision to a keep-or-stop review. The time goes into owners, data mapping, integration, review rules and training, not the model. Systems with no API or poor data take longer, so set the timeline after you map the systems.
What is the difference between an AI pilot and production?
A pilot proves the model can do the task on chosen data. Production means the system runs on live data inside your ERP or CRM, every day, with a named owner, human review, logs, a fallback path and an operating budget. A pilot answers whether it can work. Production answers whether it pays for itself.
Who should own an AI system after launch?
The head of the department that uses it should own the result, the budget and the decision to keep or stop it. A technical owner, in house or a partner, runs the integrations and model changes. Reviewers check flagged outputs. This split stops the system becoming an IT project nobody uses.
Do we need to comply with the EU AI Act if we are based in Bosnia?
It depends on what you do. Article 2 covers providers outside the EU that place AI on the EU market, and third-country providers and deployers when the AI output is used in the EU. Check whether your customers or your output are in the EU. Article 4 AI literacy already applies to deployers. This is not legal advice.