The first AI decision is an inventory

Most small businesses already pay for more AI than they use. Office suites, accounting systems, CRMs, help desks, and phone platforms have spent two years folding assistants and agents into plans the business was already buying. The owner's first decision is therefore not which new tool to add. It is which job to test with the capability that is already on the invoice.

The evidence is in the price sheets. Google folded Gemini into its Workspace business plans in January 2025. Microsoft now sells its business plans with Copilot included. Salesforce includes an employee agent in its small-business suites. Assistant makers now ship connectors to the accounting, payroll, payments, and commerce systems a small company already runs. A business that buys a separate subscription for drafting, notes, or summaries can easily pay twice for the same job.

Map the bundle to jobs, not features

Write down every system the business pays for and, beside each, the AI capability it now includes: drafting, summaries, meeting notes, ticket answers, transaction categorization, call answering. Then map those capabilities to jobs—the recurring work with a finish line, a frequency, and an owner. A feature that does not attach to a job is not yet worth configuring.

Two tools that both draft email replies belong to the same overlap group. The business needs one of them working well, not both half-configured, and any savings estimate must count that job once. The map usually shows three things: a capability the business pays for but never switched on, two subscriptions doing the same job, and one real gap. The gap is the only place a new purchase makes sense.

Switch it on at the lowest setting

Bundled does not mean safe by default. Before switching anything on, set the dial for one job. Draft for approval is the right starting point for replies, reminders, and marketing; a shadow test on completed cases is right for anything that operates a portal or touches money. Most of these products now stage sends, posts, and payments for approval out of the box. Keep that setting until a measured run says otherwise.

Then check what the agent can reach. A connector can see everything its account can see. Give the agent its own account where the product allows it, read access before write, and one named owner who reviews its work every week. Joint guidance from CISA and its Five Eyes partners in 2026 asks for the same controls: least privilege, human approval at decision points, and readable logs of what the agent did.

Read the pricing unit before the price

Outcome pricing is spreading, and it changes what the business should measure. Help desks now bill their AI agents per resolved conversation; HubSpot moved its customer agent to per-resolution billing in April. Intercom counts a conversation as resolved when the customer confirms the answer or leaves without asking for more help. Phone agents bill per minute or per call, and model providers bill per token.

Each unit is the seller's claim about value. Convert it to the business's own unit: cost per accepted job, with review minutes, corrections, and reopened cases included. A resolution that returns tomorrow as a complaint is not an accepted job, and a customer who gave up is not a customer who was helped.

The loop for the first month

Goal: remove measured hours from one recurring job with capability the business already pays for. Actor: the bundled assistant, set to draft for approval. Tools and context: the single system of record the job needs, read-only at first. Scoreboard: handling minutes, review minutes, and errors per case against a five-case baseline. Stop condition: two weeks, or the first error that reaches a customer. Human gate: every send, payment, and record change.

At the end of the month there are three honest outcomes: widen the dial for that job, narrow the job, or switch the feature off. Each is a better use of the next subscription dollar than a new tool bought on a vendor's resolution rate.

The cheapest AI a small business will ever deploy is the one already on its invoice.

Start there, measure it honestly, and buy only what the measurement says is missing.

sources
  1. Expanding Google AI to more of Google WorkspaceGoogle Workspace Updates, accessed September 29, 2026
  2. Microsoft 365 Copilot for business: plans and pricingMicrosoft, accessed September 29, 2026
  3. Agentforce in Salesforce SuitesSalesforce, accessed September 29, 2026
  4. HubSpot's Customer Agent and Prospecting Agent: now you pay when the task is completeHubSpot, accessed September 29, 2026
  5. Fin AI Agent outcomesIntercom, accessed September 29, 2026
  6. Careful adoption of agentic AI servicesCISA and international partners, accessed September 29, 2026