6 min read
Only 22% of Belgian companies feel ready for agentic AI, and barely 4% have actually rolled it out (AWS/Strand Partners, 2026). That gap is rarely down to the technology. It is down to AI not knowing the company: the knowledge that keeps an organisation running sits in mailboxes, Excel files and people's heads — and no model can reach it there. What changes when you do capture that knowledge can be put in figures: in a documented client case, handling a single absence file fell from an hour to a quarter of an hour.
Most AI pilots in SMEs do not fail on the quality of the model. They fail because the model has no access to the context in which the answer lies. Gartner expects 40% of agentic AI projects to be cancelled before the end of 2027 — a figure that matches what we see in practice.
Take an ordinary question: "is the inspection certificate for that machine still valid?" The answer exists. It is in a PDF, somewhere. Someone forwarded it once. A generic AI assistant can phrase that question beautifully but cannot answer it, because the document sits nowhere in a form it can reach.
That is the heart of it: the missing layer is not better AI, but organisational memory.
Organisational memory is everything your company knows, captured somewhere both people and agents can work with it: documents, decisions, processes and exceptions. Not as a folder of files, but with metadata — who, what, which type, which date, which status.
You notice the difference at three moments:
With that foundation, automation becomes something other than a script. A concrete example from practice, start to finish:
At 08:41 a site worker uploads his sick note through the app. An agent recognises the employee, the period and the type, tags the document and queues the absence in the file. HR approves with one click. Only then are scheduling and payroll notified. From an hour of paperwork to fifteen minutes — 75% less admin time, and every step is in the audit log.
Note the order: the agent prepares, the human decides, and only after that decision does anything irreversible happen. That is not a limitation but the design. You approve. They do the rest.
A second reason pilots stall is costs that are not predictable. So the platform routes every task to the cheapest engine that can handle the job: classic automation (RPA) for routine work such as watching deadlines, generative AI for recognising and tagging, and agents for the complex work such as assembling a dossier or resolving an exception.
We call that AI-first, not AI-native. It sounds like a detail, but it is the difference between an invoice you can explain and an invoice that surprises you every month.
| What | Figure | Source |
|---|---|---|
| Less admin time per file | 75% (1 hour → 15 min) | documented client case |
| Organisations live | 14, from 5 to 3,000 users | own figures, Aug 2026 |
| Client churn since the start | 0% | own figures |
| Time to live at a reference client | 2 weeks (Cibus Real Estate) | client case |
"Our data used to be scattered everywhere. Now HR and Prevention finally have the information they need." — Steffy De Meulemeester, HR Manager at Spuntini
This article appears on the day Ishtar365 continues as Offiks.ai. The reason is described above: what began as document management in Microsoft 365 has grown into the brain of a company — the memory and the agents that work on it. That promise deserved a name that says what it is. Offiks: the office, the real work. .ai: the agents that help carry it.
For existing clients nothing changes about the platform, the team, the contracts or the prices. Exactly what does and does not change is on the page about the name change.
Not with a transformation programme. With one workflow — the one that costs the most time today. Day 1 the document foundation stands, within the month the first workflow runs with agents under human approval, and the first ROI is visible within 30 days.
Further reading: Data & Documents (how the memory is built) and Process automation (what agents do with it).