At nine in the morning, a small-business owner opens three different lists. The inbox contains a customer's appointment request. The calendar shows the office is open. A spreadsheet says a repair is waiting for approval.
An assistant could summarize all three beautifully and still give the wrong answer. Office hours do not prove an appointment is available. A request is not a booking. A repair awaiting approval is not permission to start work.
A useful business system keeps those distinctions clear. It helps people find the current facts, move work to the right person, and record what actually happened. AI can assist with those jobs. It does not remove the need to define them.
This reference gives you a map for connecting your work. If you are starting from scratch, use the forty-chapter AI in Business guide for the step-by-step exercises. Here, your finished product is a one-page operating map: important workflows, authoritative records, owners, handoffs, and measures.
Start with the work the business promises to complete
List the promises your organization makes. An online store promises to describe items accurately, process orders, and address problems. A service company promises to respond, schedule correctly, do the work, and bill for what was agreed. A nonprofit may promise to acknowledge donations, coordinate volunteers, and deliver services reliably.
Break each promise into a few observable steps. For a customer inquiry, those might be receive, understand, check the relevant facts, prepare an answer, obtain required approval, respond, and follow up when needed.
Put a person or team beside the handoffs. A message that leaves one queue without reaching an accountable recipient is unfinished work, even if the software labels its own step successful.
Your map should reveal the next action and the evidence needed to take it. If the process is unclear without AI, connecting an assistant may simply move the confusion faster.
Use five work areas as a map, not a shopping list
Five areas make a useful starting point: administration, content, business reporting, sales, and customer service. They are categories for finding work, not five mandatory products or a promise that one person can replace a team.
| Work area | A bounded assisted job | Authoritative record | Useful result to measure |
|---|---|---|---|
| Administration | Extract agreed tasks from meeting notes | Approved notes and task register | Correct tasks with known owners or visible gaps |
| Content | Turn an approved brief into a reviewable draft | Brief, factual sources, rights records | Accepted content after review and correction |
| Reporting | Explain a checked weekly metric table | Reconciled business records and metric definitions | Accurate explanation delivered when needed |
| Sales | Prepare a follow-up from a permitted inquiry | Customer record and actual conversation | Relevant reviewed follow-up with no invented commitment |
| Service | Draft an answer from current policy | Customer request and current policy | Correct answer within its stated scope |
Finance, fulfillment, privacy, and continuity run across these areas. Include them wherever they affect the work. The categories should fit the organization; they should not hide a missing function because the diagram has only five boxes.
For each job, distinguish assistance from authority. Summarizing a meeting does not authorize a calendar change. Preparing a follow-up does not authorize an outbound campaign. Explaining a report does not authorize a payment.
Keep one authoritative place for each kind of fact
An authoritative record is the place the business uses to establish a particular state. That might be the booking system for confirmed appointments, the accounting ledger for recorded transactions, or a controlled policy document for a service rule.
This does not mean putting every record into one enormous database or granting one assistant access to everything. Different systems can remain responsible for different facts. Record how their identifiers relate and how changes move between them.
A summary should point back to its sources and preserve uncertainty. A note saying “customer requested Friday” should not become “Friday confirmed” when another assistant reads it. A proposed policy should remain a proposal until the authorized owner adopts it.
For shared information, identify a source owner and an update route. When the source changes, determine which prompts, cached copies, published answers, and pending work need review. Treat source maintenance as part of the workflow's operating cost.
Walk one request through the map
Use fictional Mesa's inspection-fee question. The permitted current policy, SERVICE-01, says the inspection costs USD 45 and provides a credit toward a repair approved within thirty days of inspection. Appointment availability must be checked separately. The older USD 35 policy is superseded.
The customer asks about the inspection price and whether Friday is available. The workflow can prepare the supported price answer while marking the appointment question unresolved. A reviewer sees the source reference and the missing availability check.
| Stage | Record produced | State that must remain clear |
|---|---|---|
| Receive | Request identifier and original question | A request has arrived |
| Check | Current permitted policy reference | Fee rule is supported; Friday remains unknown |
| Draft | Proposed answer | Wording is awaiting review |
| Review | Decision on the exact proposal | Approval covers only its stated scope |
| Act | Receipt from the authorized destination, if an action occurs | Actual execution differs from preparation |
| Follow up | Remaining question and accountable person | Unconfirmed availability still needs work |
The guide's Part 9 example stops at staging an item in a local review queue. It uses a fixed template and simulated roles. No real customer message, appointment, or business approval occurred. Use it to understand boundaries, then evaluate any real integration on its own evidence.
Choose ordinary software where it already fits
If the next step is a fixed calculation, exact lookup, or known routing rule, ordinary software may do it directly. AI becomes useful when the task requires interpreting varied language, drafting, or comparing material that does not arrive in a neat form.
A workflow can combine both. Code can check required fields and calculate a total. A model can propose a summary. An authorized reviewer can handle the exceptions that require judgment or responsibility.
Anthropic distinguishes workflows with predefined paths from agents in which the model directs more of the process and tool use. Its guidance recommends starting with a simpler solution and adding complexity when the task warrants it. This distinction helps you avoid treating every useful process as an autonomous agent. Anthropic: Building Effective Agents
For a first project, the simplest useful arrangement might be a current source document, an assistant already approved for the task, and a manual review step. A queue or orchestration platform belongs in the design when a concrete operational need requires it.
Establish responsibility before connecting systems
Record who owns the business result, who controls the data, who maintains the integration, and who can stop it. In a small team, one person may hold several roles. Still record the responsibilities and backup coverage separately so absence does not make every exception a mystery.
NIST's voluntary governance playbook emphasizes policies, accountability, and third-party considerations throughout AI use. Apply those ideas to the actual workflow and its consequences; adopting a vocabulary does not establish that the necessary controls are working. NIST AI RMF Playbook: Govern
Give the assistant access to the records and actions needed for its defined job. Do not treat a successful connection as permission for every action the connected service offers. A person permitted to review a draft may lack authority to approve a refund or export confidential records.
For external actions, record the exact proposal, approval scope, and observed outcome. Keep an exception route for missing information, stale sources, and uncertain execution. These records make the workflow understandable to the next person who must operate it.
Pick the first improvement from your actual bottleneck
There is no universal order that requires administration first or support last. A busy service company may need better intake before it needs a content engine. A publisher may benefit more from source organization. A nonprofit may start with clearer volunteer instructions.
Compare candidate tasks on value, frequency, input quality, consequences of error, review effort, and ease of measuring the finished result. Prefer a bounded task whose sources and owner are available. Keep a manual alternative so you can compare performance and maintain service.
A workflow's attractiveness can change once you count corrections and upkeep. Measure complete effort, accepted outputs, actual cash effects, and later business outcomes where relevant. More drafts do not automatically mean more useful work or more revenue.
Also consider work shifted to someone else. A faster coordinator may create a larger review queue for the manager. The operating map should expose that bottleneck before you expand volume.
Make the system understandable on an ordinary bad day
Ask what happens when the model, connection, source file, or reviewer is unavailable. Identify the minimum service the business can maintain and the records needed to reconstruct pending work.
Check how staff find the manual process without relying entirely on the failed tool. Preserve current sources, request identifiers, and uncertain-action records through an approved route. A backup assistant cannot compensate for a missing booking record or an approval nobody can locate.
Review the map after material changes. New permissions, a different provider, additional task types, or a changed policy may invalidate an earlier evaluation. Keep the map small enough to maintain and specific enough to guide a real decision.
Finish by mapping one promise from incoming request to accepted outcome. You should be able to point to its sources, owner, next action, approval boundary, result record, and fallback. That is the foundation of an AI-assisted business system: work people can understand, verify, and improve.