The platform does too much, and that is the main obstacle to adopting it. When a tool fits every department, the project spreads until nobody owns it. The approach that works runs the other way: pick one department, one process with a measurable result, and finish it. Below are scenarios by business function, with what actually changes in each — and the deliberate note that these are starting points, not a rollout plan.
Why breadth is the problem, not the selling point
A demonstration that shows finance, sales, HR and operations all benefiting produces enthusiasm and no decision. Every department can imagine a use, none owns the budget, and the project becomes an initiative rather than a deployment.
The counter-move is narrowness that feels almost disappointing. One process. One team. One number that was true before and different after. That produces an owner, a result and a reference — in that order — and the second department then arrives asking rather than being persuaded.
Finance
Reconciliation and close. The monthly close involves pulling figures from several systems and chasing discrepancies. An agent that assembles the comparison and flags only the exceptions removes the assembly work while leaving judgement where it belongs.
Recurring management reporting. Reports that are structurally identical each period and rebuilt by hand each period. The measurable outcome is hours, and it is easy to establish a baseline.
Sales
Account preparation. Before a meeting, someone checks the CRM, recent tickets, usage and open invoices. An agent that assembles this into one brief saves a predictable amount of time per meeting and improves what happens in the room.
Pipeline hygiene. Identifying deals that have not moved, contacts who have gone quiet, and forecast entries missing a required field. The value is not the alert; it is that it arrives without anyone running a report.
Customer support
First-response context. Most first responses require looking up the same three things. An agent that gathers them before a human opens the ticket compresses response time without changing who answers.
Recurring-issue detection. Grouping tickets by underlying cause rather than category, which surfaces problems that individually look minor.
Operations and supply
Exception monitoring. Deliveries, stock levels or process steps that deviate from expected. The agent watches continuously; people intervene only on exceptions.
Supplier documentation. Checking that incoming documents contain what they should and match the order — high-volume, rules-based, and reliably tedious.
HR
Onboarding coordination. A new hire triggers a sequence across several systems and departments. The agent runs the sequence and reports what is stuck.
Internal enquiries. The same questions about policy, leave and process, answered from current documents rather than someone’s memory of them.
Legal and compliance
Contract review triage. Not replacing legal judgement, but sorting incoming agreements by whether they use standard terms, flagging deviations for attention. The measurable outcome is time to first review.
A caution that applies across all of these: an agent producing a recommendation is useful; an agent producing a decision that nobody checks is a liability. In regulated functions especially, design the human checkpoint before deploying, not after the first mistake.
Choosing the first one
Three criteria, applied in order. The process must be repetitive enough that a week of observation shows the pattern. It must have a measurable baseline that exists today, not one that would need to be constructed. And it must have an owner who wants it — the strongest predictor of whether a deployment survives its first month.
A process meeting all three but looking unglamorous beats an ambitious one meeting two. The purpose of the first deployment is not impact; it is evidence.
Key points
- Fitting everywhere is why these projects stall — narrowness creates ownership.
- The strongest first candidates are assembly work: gathering, comparing, checking.
- Judgement stays with people; the agent removes the preparation before judgement.
- Design the human checkpoint before deployment, particularly in regulated functions.
- Choose on repetitiveness, an existing baseline, and a willing owner — in that order.
How many agents should a first deployment include?
One. A second agent doubles the maintenance surface while the team is still learning what maintenance means, and it makes attribution harder when measuring the result.
Which function usually produces the clearest result?
Finance, because the work is already measured and the baseline exists in someone’s calendar. Support is a close second where response time is tracked. The common factor is measurement that predates the project.
What if no department volunteers?
That is a useful signal rather than an obstacle — it usually means the pain is distributed rather than acute. Deploying anyway produces an agent nobody maintains. Better to wait for a team with a problem they are already complaining about.
How long before the second department?
After the first has a measured result and has run long enough for the maintenance burden to be visible — typically a quarter. Expanding before that means repeating any structural mistake at scale.