Stop Hiring AI Engineers. Hire People Who Know Your Business
Generic AI talent won't fix your workflow. What a PME needs is an integrator who already understands its software and its métier.
When the Demo Works and Tuesday Doesn't
You sat through a polished pitch. A chatbot answered customer questions in seconds, a model summarized contracts, an agent drafted emails that sounded almost human. Someone on your team got excited and brought in an AI specialist, maybe even a small team of them, to build something similar. Three months later the project sits half-finished in a sandbox. It can't read your invoicing format. It doesn't know that your CRM calls a client a "dossier" and your accounting software calls the same record a "compte." Nobody on the AI team had ever opened your ERP before they started.
This is not a failure of the technology. Large language models are genuinely capable. The failure is in who got hired to deploy them. An AI engineer, however skilled, is trained to build models and pipelines. They are not trained to understand why your warehouse team enters stock corrections at the end of the day instead of in real time, or why your sales reps still fax certain orders to one supplier. Those details are not technical footnotes. They are the business logic that any automation has to respect, or it breaks something downstream that nobody notices until the month-end close doesn't balance.
The pattern repeats across PMEs that bought AI from a big vendor's menu: a generic capability, dropped into a business that has its own structure, its own exceptions, its own software stack built up over a decade. The vendor's job was to sell a model. Nobody's job was to make that model fit your invoicing workflow, your stock rules, or your client file. The gap between "the AI works" and "the AI works here" is where most of these projects quietly die.
What to Do Before You Hire Anyone for an AI Project
The fix is not avoiding AI. It's changing who you trust to bring it in. A business owner should treat an AI or automation project the same way they'd treat hiring a key employee: check that the person understands the job before checking what tools they know.
- Ask the integrator to walk through your current software stack with you before they propose anything — if they can't explain back to you how your invoicing or CRM currently works, they're not ready to automate it.
- Require a written map of the actual process the AI will touch, step by step, including the exceptions and manual workarounds your team already uses.
- Demand that the first deliverable is a small, measurable task — one invoice type, one customer segment, one report — not a platform-wide rollout.
- Get a cost breakdown that includes ongoing usage, not just setup: what each processed document, query or API call will cost once it's running daily.
- Check who owns the data and the integration once the contract ends — if the system can't run without the vendor's ongoing presence, you haven't bought a tool, you've bought a dependency.
The Difference Between a Pilot and a System
A generic AI vendor sells you a capability: summarization, classification, chat. An integrator who knows your business sells you an outcome: fewer re-entered invoices, faster quote turnaround, a support ticket that routes itself correctly on the first try. The difference shows up in the contract. One talks about models and tokens. The other talks about your accounting software, your CRM fields, and the three people on your team whose daily work actually changes.
This also changes how you evaluate risk. A model that hallucinates a product description is a nuisance. A model that miscategorizes a tax line because nobody told it how your chart of accounts is structured is a liability. The person who should catch that risk before it reaches production is someone who has actually opened your books, not someone who has only opened a model's documentation.
What to Watch in the First Ninety Days
Once a project is underway, don't wait for a glossy report to tell you it's working. Watch the operational signal directly: is the task that used to take an employee an hour now taking fifteen minutes, with the same or better accuracy? Is the error rate on the automated step lower than the manual one it replaced? Is the monthly cost of running the system — API calls, hosting, maintenance — smaller than the labor cost it displaced? If after ninety days you can't answer these three questions with a number, the project hasn't earned its place yet, regardless of how convincing the original demo was.
Talk to Someone Who Already Knows Your Software
ArkonLabs builds custom business software and automations by starting with the system you already run — your ERP, your CRM, your invoicing — not with a generic AI stack looking for a use case. If you want an integration that fits your business before it fits a sales deck, reach out through www.arkon-labs.com.