METHODOLOGY

Same methods. New help.

Your business already has ways to plan, review and approve work, and most functions have a well-understood way to do it right. We start there. We don't invent a new process: ordinary software holds the forms and checks, AI drafts, compares and helps make some of the decisions, and the people with the authority sign off.

01 · Tools

Start with tools you already use

We look up what normal looks like for the function, compare it with how you run today, and automate that. If your people came out of sales, purchasing, planning, accounting or engineering, they will recognize every row.

Table 1 · Where each tool's answer goes
Familiar toolWhat stays settledWhat AI can prepareWhat people decide
MRP and capacity planning (CRP)Bills of material, calendars, stock and unitsInitial shortage and overload analysisSchedule, overtime and capacity changes
Purchasing and supply chainOrder fields, dates, approval limits and supplier recordsLate-order comparisons and expedite draftsSupplier changes and spending
Inventory managementLocations, counts and reorder pointsStock-out and excess analysis against demandReorders, write-offs and transfers
Project trackingMilestones, owners and statusSlip analysis and the open-items listPriorities, resources and dates
Stage gatesRequired evidence and gate statusMissing-evidence list and review packetGo, hold or drop
SWOT and risk analysisFactors, scoring scale and source linksInitial analysis by AI, with evidence for each itemStrategy and which risks are accepted
Sales pipeline and CRMStages, required fields and account ownershipResearch and follow-up draftsPricing, commitments and stage changes
Budgets and forecastsAccounts, periods and calculation rulesVariance analysis with source referencesExplanations, adjustments and the close
FMEA and APQPRequired fields and source linksCandidate risks and unanswered questionsAccepted risks and release
Control plansCheck method, frequency and reaction fieldsMissing-check reviewProcess changes and response
Lean, Six Sigma, 5 whys and A3Steps, owners and evidence fieldsComparisons and possible causesWhich cause is supported and what to change
Standard work and checklistsCurrent version and completion fieldsDraft instructions and exceptionsApproval of the next version

Explore the tools in hotbotspot. The public AI agents run as open experiments at A.N. Creation Office.

02 · Owners

Give each part a clear owner

Settled in software.

Open the current record, enforce fields and allowed states, check the return, and save a version that can be read back.

What the AI does.

Read the task's instruction, compare the supplied evidence, draft the result and mark missing facts. A rejected record returns for correction.

What a person decides.

Review the substance, approve or reject the action, and decide when the work is ready to move. Anything that can't be undone waits for a yes.

Never without a yes

Anything that can't be undone waits for a yes.

Three walkthroughs

Looking for the molding example? Read Manufacturing

A late purchase order

Illustrative: Paper Lantern Distribution is a made-up company. A supplier moves a confirmed date by two weeks. The buyer supplies the PO, stock balance and customer-order dates. Software keeps the original promise date and records the new one separately. AI identifies the customer orders that may be affected, marks missing alternate-source prices, and drafts a follow-up. It offers partial delivery and a schedule change as options, without treating either as agreed. The buyer checks stock with the warehouse and chooses a partial-delivery request. The draft stays unsent until the buyer approves it. The reply and the decision return to the order record.

A month-end variance note

Illustrative: Cedar Desk Services is a made-up company. The controller supplies the ledger export and approved budget for one period. Software aligns the accounts and computes the differences. AI drafts a note separating the recorded difference from possible explanations. A travel increase has invoices; an outside-services increase has no supporting detail, so its cause stays open. The controller rejects the unsupported explanation and asks the account owner for evidence. No journal is posted. The accepted note and unresolved question are saved with their sources.

A customer complaint

Illustrative: Harbor Lane Goods is a made-up company. A customer reports the wrong item in a shipment. Software opens a case using the order number and return-policy version. AI compares the message with the shipment record and drafts a reply asking for the missing item label. It cannot infer damage or promise a replacement from the complaint alone. A service lead reviews the reply and approves sending it. When the label arrives, the lead decides whether the policy covers a replacement. Any refund or shipment requires separate approval. The case retains the evidence and the decision.

Bring AI agents on carefully

Give them the current system, the rules and the limits. Check the first work closely. Give them more responsibility as they earn it, with test results and a record. Train the people who will review it, handle exceptions and change the instructions when the business changes.

What the instructions say

Deeper in the work, four kinds of files help: principles state the shared rules; playbooks describe a job's stages and gates; skills describe one task; records hold facts, decisions and evidence. These are a flat set of working materials. Each task instruction states its purpose, inputs, boundaries, method, output, reviewer and what must be saved. Read the notes.

Where this stops

A fixed form can still contain a wrong answer. Instructions alone do not enforce permissions. Not every task justifies custom software. Test with your own data and model, including missing sources, conflicting records and denied approvals. Test again after a change. Solutions explains what has and has not been tried.