AI Adoption SME Wiki

AI agents are moving from assistive chat into workflows where they autonomously execute tasks: reviewing documents, extracting and moving data between systems, or triggering decision steps. For SMEs, this shift means access to specialist functions once reserved for larger competitors.

What counts as an agent use case

Agents that matter to SMEs share a pattern: they operate inside existing workflows rather than as standalone chat interfaces. An agent runs in the background or is triggered by an event—a contract arrives, an invoice appears, a status changes—and executes a complete task without waiting for a human to prompt each step. This is a shift from assistive AI (which retrieves or drafts information) into operational AI (which changes systems, creates records, and makes decisions).

Aaron Levie, CEO of Box.com, argues that this shift is from "read-only operation to fundamentally a read/write operation." The distinction matters: read-only AI helps a person retrieve or analyse information; read-write AI executes complete workflow steps. For an SME, read-write agents mean less manual work moves into software that runs unattended.

To be operationally viable, agents need three things: access to the systems where work lives (Box, Linear, CRM, accounting software); reliable context about what they are working on (company data, process definitions, rules); and infrastructure to run safely in the background (long-running execution, code sandboxes, audit trails). Most SMEs have the first and may lack the second and third—which is where implementation work comes in.

The competitive angle for smaller firms

Levie identifies one outcome worth naming: agents could narrow the specialist-talent gap that historically favours large companies. A 20-person firm cannot afford a dedicated legal analyst, senior accountant, or customer-service manager; firms with 500 people can. But an agent doing routine legal review, invoice processing, or customer escalation triage gives the small firm equivalent capacity. Levie argues that access to expertise once meant hiring specialist staff; now it can come as an agent running a workflow.

This does not eliminate the need for human judgement, but it changes where judgement goes. Instead of spending half their week on mechanical work, staff redirect time to decisions an agent cannot make: whether a contract's terms are acceptable given the relationship, whether an exceptional invoice error reflects a real problem, whether a customer escalation needs an exception. (This is inference: Levie names the shift in staff time; the specific decisions are how that plays out in business functions.) Manual work can shift to agents, freeing people to move toward customer focus, innovation, and judgement-intensive tasks.

It also shifts how IT works inside an organisation. Levie notes that rather than buying software for human workers, IT becomes a supplier of operational capacity: agents that run alongside people. IT teams will shift from software procurement to agent management and maintenance.

What deployment actually entails

Model capability alone does not produce a stable business process. Getting from "the AI can do this task" to "the AI does this task reliably in our workflows" requires substantial work: upgrading IT systems to support agent integration, giving agents context about your business (rules, data, definitions), redesigning workflows around how humans and agents will actually work together, and managing adoption so staff see the benefit rather than threat.

Levie notes that the major AI labs have launched new initiatives to help enterprises deploy agents inside organisations, signalling implementation is now core to the market. The gap is real and is creating opportunities for new jobs and firms (which is why forward-deployed engineers now help SMEs deploy agents).

This is worth tracking as you consider which use cases to prioritise. A use case that seems simple ("let an agent screen customer emails") may depend on context work that isn't: does the agent know your SLA targets? Your escalation rules? Which competitor's email you treat as urgent? That context usually lives in people's heads or scattered across tools. Extracting and formalising it is the unglamorous groundwork that makes an agent reliable.

Where agents work

Agents operate in workflows where work lives inside systems the SME already uses. Levie's examples span contract review feeding into Linear task tracking, client onboarding, invoice processing, M&A due diligence, and data extraction pipelines. The pattern is consistent: documents or records arrive, an agent reviews and extracts what matters, and writes the outcome into the next system humans need it in. This saves the manual transcription step.

Because agent value depends on reliable context about your business (rules, definitions, decision thresholds), and connection to systems where your work lives, use cases rarely transfer unchanged between organisations. What works for one firm's invoice approval may not work for another's, and implementation is where FDEs and internal teams do real work.

The sub-categories below cover functions where SMEs are deploying agents. Each sub-category page details the tasks, approaches, and tools specific to that area:

Customer service: agent-handled triage, escalation routing, known-answer responses
Sales: lead qualification, prospect research, pipeline updates
Operations: data movement, document review, system-to-system synchronisation
Admin and HR: onboarding workflows, scheduling, routine queries
Finance: invoice processing, expense categorisation, reconciliation
Compliance: policy application, audit trail generation, flagged-item routing
Knowledge management: extracting and indexing organisational knowledge, retrieval for agent context

What makes a use case defensible

Not all agent use cases survive as models mature and cheaper competitors emerge. Sara Normous, a venture capitalist investing in AI-native companies, distinguishes between the kind of work that gets commoditised and the kind that stays valuable. Measurable correctness gets eaten from both directions: below by task saturation and margin compression, above by frontier labs absorbing scaffolding (tool integration, retrieval, routing) into model weights.

What persists is work where correctness is singular—where whether the agent is right is established by judgement inside your system, not by a public benchmark. A token spent answering a generic question is worth almost nothing, since anyone's model can answer it, while a token spent reasoning over your company's data is worth much more. This is another reason context work matters: an agent that reasons over your rules, your customer history, your outcome definitions, is harder to commoditise than an agent that retrieves and summarises.

For an SME, agent value lives in specificity rather than generality. The agent that categorises your firm's invoices using your chart of accounts and approval rules is more defensible than one that does generic expense categorisation. Specificity is what survives margin erosion.