The first agent sets the tone for everything after it. Pick the workflow that can prove value quickly, safely and in numbers your finance team accepts.
Gartner predicts that 40% of enterprise applications will feature task-specific AI agents by the end of 2026. The pressure to start somewhere is real. Starting in the wrong place is how organisations end up with a stalled pilot and a sceptical board.
Six tests for a first workflow
1. Volume
The work happens often enough that time saved adds up. A task done thousands of times a month makes the case in weeks of data. A task done ten times a quarter never will.
2. Rules you can write down
Most of the work follows rules the team can explain: what to check, where to look, what counts as an exception. Agents handle judgement support well; they should not be asked to invent the rules on day one.
3. A baseline you can measure
You can say today how many items flow through, how long each takes and how often they are reworked. Without a baseline there is no proof of value, only an opinion.
4. Contained risk
A mistake is caught before it reaches a customer and can be corrected. For the first build, avoid workflows where the agent’s output is itself a decision with legal consequences for a person; those need the POPIA section 71 design work and are better as a second step.
5. Accessible data and systems
The documents and systems the agent needs can be reached with proper access controls, without a six-month integration project before anything runs.
6. A named owner
One person owns the workflow, wants it fixed, and can decide whether the pilot goes to production.
Good first candidates
Workflows that tend to pass these tests include document intake and extraction, email and case triage, drafting standard replies for approval, reconciliation and recurring reporting, and internal knowledge questions answered from approved sources. Each has volume, rules, a measurable baseline and a natural human checkpoint.
Poor first candidates
- Workflows nobody owns, or that three departments each partly own
- Processes that are broken and need redesign before anything is automated
- Rare, high-stakes decisions where the agent would decide rather than prepare
- Anything whose data cannot be accessed without a separate integration programme
The Automation Savings Estimator takes your volumes, time per item and cost, and shows the hours each workflow could release and which agent to build first.
When you have a candidate, the next step is a paid proof-of-value pilot on that one workflow. See how the pilot works.
Sources
- Gartner press release, 26 August 2025: 40% of enterprise apps will feature task-specific AI agents by 2026, up from less than 5% in 2025.
- Protection of Personal Information Act 4 of 2013, section 71.