Most companies I meet now have an AI demo. Fewer have AI in the actual job.
A demo is easy to recognise. It lives in a slide, a sandbox, or a tool that sits beside the work. People try it, nod, and then go back to the spreadsheets, the inboxes, and the visits that still run the way they did last year. The intelligence is nearby. It is not in the process.
The useful test is simpler than a feature list. Does the system change an outcome inside the work, before a person has to notice? If the answer is no, you do not have an AI capability yet. You have a demonstration.
I have spent a lot of time on the other side of that line, building systems that have to hold up in regulated enterprises, in field teams, and in support queues. Across twenty-seven builds, the pattern that keeps proving itself is this: put the AI in the work, not next to it.
Here is what that looks like in practice. I will stay with anonymised examples from real delivery, and I will not add numbers or stories that are not in the work itself.
A large pharmaceutical manufacturer, with 1,200+ employees, was running its entire people operation across disconnected spreadsheets, manual payroll, and fragmented HR tools. Every report took hours to assemble. Attrition was going undetected. Compliance was a risk.
The useful response was not an HR chatbot parked next to the existing mess. It was a single AI-powered platform covering the complete employee lifecycle, from first application to final settlement. Seventeen module groups. 130+ features. 25+ embedded AI capabilities. Ten user roles. A 24-week roadmap in four delivery phases.
The AI sat inside the work. Resume parsing at acquisition. Predictive time-to-hire. Appraisal bias detection and succession risk scoring. Workforce management with real-time anomaly detection. A learning system with AI-personalised paths. An executive view that surfaced KPIs, attrition flags, and compliance alerts.
The moment I come back to is this one. When a department's attrition risk score climbs, the platform flags it automatically and surfaces a retention recommendation before a single resignation letter arrives. That is not a demo. That is the job, happening earlier than a person would have caught it.
The same idea shows up in support.
A support operation needed to handle high inbound message volume on WhatsApp without scaling headcount in a straight line, while keeping response quality and a clean handoff to human agents. Volume was growing faster than the team could hire against it.
The build was an AI-assisted automation layer on WhatsApp Business, delivered in twelve weeks. Common queries were recognised and resolved end to end, without an agent touching the conversation. Edge cases routed to a human with the conversation context intact, so the customer never had to repeat themselves. The resolution flow was tuned against the real query mix, so automation absorbed the repetitive volume and agents kept the complex work.
On the automated flows, first response moved from minutes to seconds. A large share of repetitive queries resolved without agent involvement, in line with mature support-automation deployments. Again, the intelligence was not a separate assistant that a tired agent might remember to open. It was in the queue, doing the repetitive part of the job, and stepping aside when the job needed a person.
Field work is a third version of the same lesson.
A major FMCG brand needed merchandising teams to execute in-store standards consistently across hundreds of retail touchpoints. The familiar problem was this: what actually happened in store looked nothing like the plan on paper.
The response was a mobile-first field execution platform. Merchandisers select the store format, and the platform loads the mandatory SKUs, shelf share targets, planograms, and scoring rules automatically. Four pillars get tracked on each visit, with photo proof: distribution, shelf share, planogram, and pricing. On visit completion, the total score calculates automatically, broken down by pillar, and syncs to management dashboards in real time. The configuration was built to deploy across food, beverage, personal care, and home care without rebuilding, and to be ready for more than one brand and more than one market from day one.
There is no large language model in that story that I need to wave at you. There is a score that used to be assembled later, if it was assembled at all, and that now exists at the shelf, while the visit is still the work in front of the merchandiser. The system is in the visit, not in a report that arrives after the window to fix the store has closed.
If I stand back from these builds, and from the rest of the twenty-seven, a few practical rules hold.
Put the intelligence where the decision actually happens. In people operations, that is inside the lifecycle, not in a side tool. In support, that is inside the conversation. In the field, that is inside the visit.
Build for the job that already exists, then let the AI change the timing of it. Attrition risk before the resignation. A first response in seconds, not minutes. A store score at the end of the visit, not at the end of the week.
Treat integration and governance as part of the product, not as a later phase you will get to. In the same body of work, a multi-site quality system and a laboratory system were taken from kickoff to go-live in 135 days, with audit trails and the controls those environments require. A regulatory information system for a multi-country pharmaceutical operation was made production-ready in twelve weeks by customising an existing platform rather than rebuilding it, with maker-checker controls so the same user cannot create and approve. Twenty-two third-party systems have been delivered with zero critical-path slippage, including SAP and SSO across regulated estates. This is unglamorous, and it is usually the reason an AI feature can actually run.
And when the date is fixed, treat the date as the brief. A national campaign stack, designed, built, and deployed for a UAE-wide scan-and-win promotion across all emirates, went live in two weeks. An animated, five-chapter product story, built inside an existing single-page site, was delivered in ten days across three phases. Launch dates are not a mood. They are a constraint the work has to meet.
None of this requires you to start with a model. It requires you to start with the job: the resignation you want to see coming, the message you want resolved, the shelf you want to be able to score before the merchandiser leaves the aisle. Then you place the AI there.
It is also the kind of work I build: enterprise systems, field tools, and automation designed so the intelligence sits inside the process. I write more about brand, growth, and this kind of building at usmanakhan.com.
If you already have an AI demo, keep it if it helps people see what is possible. Just do not stop there. Move the same intelligence into the job, and let it change an outcome before a person has to notice.
That is when AI stops being a slide, and starts being the work.
This article was written with AI assistance and reviewed by the author. It was first published on Hashnode.
Top comments (0)