Before you put AI into a task, check whether a plain rule or ordinary automation already does the job. Use AI only for the part rules can’t handle, keep a person approving the steps where a mistake costs money, and test it on your own real examples before it goes to production. That order keeps the build cheaper, the running costs lower and the risk visible.
This page used to summarize a 2021 list of technology trends. Most of them were about how companies would work after the pandemic. Two of them turned into decisions that founders and product owners still make on many builds today.
The first was automating whole business processes from end to end, with rules, integrations, robotic process automation and AI working together. The list called it hyperautomation. The second was the engineering work of getting AI out of a demo and into production, where it has to stay accurate, affordable and safe. The list called that AI engineering.
The questions behind those two labels are still the ones to answer: which steps of a process should a machine do, and how do you trust the ones AI does?
Most EU companies haven’t reached the second one yet. In 2025, about 20% of EU enterprises with 10 or more employees used AI, 6.5 percentage points more than in 2024, and 55% of large ones did. Among firms that had considered AI and didn’t use it, the most common reason was a lack of relevant expertise, named by 71% (Eurostat).
Climb only as high as the task needs. Each step up adds cost and something new to check.
| Option | Use it when | Running cost | Example |
|---|---|---|---|
| A rule in code | The input is structured and the decision fits an if-then | Close to zero per run | IPOPBOOKS: once a bookkeeper approves a merchant, the next charge from it is sorted from memory, with no AI call |
| Automation | Data moves between systems or documents change format, with clear steps | Low | Hays: formatting one CV went from up to 2 hours to 2 minutes for 200+ recruiters |
| AI with a person approving | The input is messy and a mistake costs money | AI fees per call plus review time | IPOPBOOKS: AI suggests a category for each bank transaction, with a reason, and a bookkeeper approves it |
| AI on its own | Mistakes are cheap, visible and easy to correct | AI fees per call | Subflare: lead scores go out on their own, and each shows the visitor’s timeline behind it |
Most real processes use several rungs at once. IPOPBOOKS, a bookkeeping platform we built for US small businesses, uses three: AI for the first suggestion, a person for approval and a rule that remembers the decision. The IPOPBOOKS case study shows how the pieces fit. Rungs can also move down over time: once a kind of item keeps getting the same answer, a rule can take it over and the AI call stops costing money.
Plenty of slow office work can be automated without AI: copying data between systems, reformatting documents, writing the same report every week.
At Hays, a recruitment company, recruiters spent up to 2 hours formatting each CV by hand. The tool we built does it in 2 minutes for 200+ recruiters, and it later grew into a mini-ERP linked to recruiter KPIs. The Hays case study has the details. SmartBuilding replaced paper site journals with QR-code access control and automatic reports on safety violations, and it went from start to launch in 2 months. Zuzan saves teams up to 60 hours a month on booking B2B meetings.
Go through your process one step at a time and answer these:
Structured input with a cheap mistake points to a rule. Messy input with an expensive mistake points to AI with a person approving. Messy input with a cheap mistake, such as sorting incoming leads, can go to AI on its own, as long as people can see why each answer was given.
A demo that works on ten hand-picked examples proves little. Before AI goes to production, check four things.
Test on real examples from your business and compare the AI’s answers with what your people decided. Subflare, a lead scoring product, tested its risky AI part in 2 weeks before the full build started.
Each answer has to cost less than it earns or saves. If the product depends on AI, the first release has to prove both that the answers are useful and that each one costs less than a user pays. Rules that remember past decisions, as in IPOPBOOKS, cut the number of AI calls over time.
In IPOPBOOKS, AI never writes to the books on its own. It suggests, and a bookkeeper approves. None of our AI agent projects is fully autonomous: where a mistake costs money, a person approves. Keep that pattern wherever a wrong answer moves money or reaches a customer. Reported failures are rising, too: Stanford’s 2026 AI Index reports that documented AI incidents rose to 362, up from 233 in 2024.
NIST’s AI Risk Management Framework (version 1.0, released in January 2023) organizes the work into four functions: govern, map, measure and manage. NIST added a Generative AI Profile in July 2024 and says version 1.0 is now being revised under the White House AI Action Plan. For a small team, a written owner, a test set and a monthly review of errors and costs cover the basics.
If you sell in the EU, the AI Act sets dates as well. Obligations for general-purpose AI models have applied since 2 August 2025, and the Act as a whole has applied since 2 August 2026, with some exceptions. Rules for high-risk uses in sensitive areas, such as CV-sorting software for recruitment or credit scoring, now start on 2 December 2027 (European Commission). This isn’t legal advice; ask a lawyer which rules cover your product.
You can estimate your AI pilot’s cost and timeline before you pick the first step to test.
We start this kind of work with a pilot: $5,000-15,000, ready in 2-4 weeks. If the pilot works, we price the production version then. See our process automation and AI development pages for what we have built.
It is the idea of automating a whole business process from end to end, combining rules, integrations, robotic process automation and AI, instead of automating single tasks one at a time.
The work of taking AI from a demo to production: testing it on real data, keeping costs per answer under control, adding human approval where mistakes are expensive and monitoring it after launch.
Use rules and plain automation wherever the input is structured and the steps are clear. Add AI only for steps with messy input, and keep a person approving where a mistake costs money.
At Rocketech, a pilot costs $5,000-15,000 and takes 2-4 weeks. You pick one task, we build it and test it on real examples with your team, and then you decide whether to take it to production.
Carefully curated content for resourceful Devs, CTOs, and PMs. No spam.