Sort the job into four buckets before you approve the hire
Write down the ten things the new person would actually do, then split them into judgment, relationship, repetition and coordination. Three of those buckets are not a hiring problem.
Start with the work, not the job title. List the ten recurring tasks you expect the new hire to carry out.
Then sort that list into four categories: judgment, relationship, repetition and coordination. The shape of the list tells you whether you have a headcount problem or a workflow problem.
Judgment is work with consequences, ambiguity or accountability attached. Strategy, negotiation, creative direction, hiring, and sensitive customer decisions sit here. Software can assist, but a named person should own the outcome.
Relationship is work where trust is part of the value. Sales conversations, coaching, conflict, partnership development and customer recovery lose value when they are treated purely as an efficiency question.
Repetition is high-frequency work with stable rules: reminders, routine follow-ups, scheduling, data entry, status updates, templated messages. These are the automation candidates.
Coordination is work that exists only because systems or people are disconnected. Copying information between tools, chasing approvals, reconciling versions and asking the same status question three times all belong here.
The decision rule follows the buckets. If the bottleneck is mostly judgment or relationship, hire or develop a person. If it is mostly repetition, test automation before committing to a permanent salary. If it is mostly coordination, redesign the process first.
If a role contains all of it, split the workflow so the human job is built around the work that needs a human.
The sequence matters because automation amplifies whatever process design it finds. Clear ownership and reliable inputs make automation useful. Fuzzy ownership and wrong data make the confusion faster.
The adoption figures suggest most companies are still at the tool stage rather than the redesign stage. The Stanford AI Index 2026 reports AI use in at least one business function at 88% of surveyed organisations, with agent deployment still early. The US Census Bureau put overall business AI use at roughly 17% to 20% across late 2025 and early 2026, with higher rates among larger firms and knowledge-intensive sectors.
Census research on AI diffusion finds sales and marketing, strategy and business development, and IT among the most common adopting functions. Those are the same functions where lead routing, follow-up, scheduling and reporting quietly generate headcount pressure.
The productivity evidence argues for precision rather than blanket automation. An NBER study of more than 5,000 customer support agents found a 14% average productivity increase from AI assistance, with substantially larger gains for less experienced workers. A Harvard Business School field experiment found large speed and quality gains for consultants on tasks within the technology's capabilities, and worse performance on tasks outside them.
For a pilot, baseline the workflow before you touch it: cycle time, throughput, quality, error rate, human review time and customer impact. Then run it on one defined workflow. Heavy use with an unchanged workflow is adoption without value.
Price the workflow before you price the role. Take one recurring process, estimate the minutes it consumes and how often it happens, and you have the number to compare against a salary. That cost already exists; it is just buried inside the salaries you are paying now.