A customer asks why an order has not arrived. Your team opens the sales system, checks the warehouse, looks for a production update and reads the last email. The question took seconds to ask. Finding an answer means piecing together a small investigation. Now imagine asking an AI assistant to do that work. Writing a clear reply is one part of the job. Finding the right records, resolving contradictions and knowing whether it may take action are the parts that make the reply useful. This is where AI orchestration comes in.

THE IDEA

The answer is only as useful as the process behind it.

To apply this workflow in your business, explore our agentic automation service and implementation approach.

Follow the question through the business

AI orchestration is the coordination of models, data sources, software tools and review steps around a task. The language model can interpret a request and explain the findings. The surrounding application controls which systems are available, what happens next and when the work is complete. Some steps use AI; others are ordinary software rules or human decisions.

Consider an illustrative order enquiry. The assistant first confirms the order reference. It retrieves the promised date from the ERP, checks whether the warehouse has dispatched the goods and, if necessary, looks up the linked production job. It then prepares a response with the records supporting its explanation.

Suppose the warehouse shows a dispatch entry, but the carrier has no collection scan. The useful result is a clearly identified gap for the dispatch team to investigate. Claiming the parcel is already travelling would turn incomplete information into a promise the business cannot support.

Give routine work a clear route

For a familiar task, write down the steps before choosing an agent framework. An order-status enquiry might follow a defined route: identify the order, retrieve its records, check for missing information, prepare an answer and apply the sending rules. AI can help interpret the customer's language without deciding every part of the process.

A more open-ended request, such as investigating recurring delivery delays across several sites, may need an agent that chooses additional checks as it learns. Anthropic's guidance distinguishes predefined workflows from agents that select their own steps. Microsoft's architecture guide likewise recommends beginning with the least complex design that meets the requirements.

A useful design question is: where does this task genuinely require judgement? Keep predictable checks in software and give the model a bounded role in the parts that vary. Add freedom when testing shows why it is needed.

References: Anthropic: Building effective agents; Microsoft: AI agent orchestration patterns

Use several agents when the work can be divided

Some investigations contain independent pieces. A delivery review could examine carrier exceptions, production delays and customer complaints separately, then bring the findings together. Multiple agents may help when each has a clear question, access to the appropriate evidence and a defined result to return.

The tradeoff is coordination. Workers can repeat searches, interpret terms differently or return findings that require another round of checking. Anthropic's account of its research system describes both the value of parallel investigation and the additional token use and coordination constraints.

Reading separate records is also different from changing a shared order. Reserving stock, cancelling a shipment and issuing a refund have dependencies. Those actions need an agreed sequence and a single authoritative transaction state. A simple order lookup rarely needs a team of agents merely because several systems are involved.

References: Anthropic: How we built our multi-agent research system

Check the business facts before polishing the answer

Imagine two agents agreeing that an order is ready because both read the same outdated stock record. Agreement has added no new evidence. A stronger check looks at the source, its update time and the business rule that defines readiness.

For the order example, require a valid order identifier, matching product references and a clear distinction between packed, dispatched and delivered. Use the relevant system to calculate quantities and balances. Ask the model to explain those results in language the customer can understand.

Keep facts, interpretations and unknowns visible. 'A dispatch record exists' is a fact supported by a record. 'The carrier probably collected it' is an interpretation. A proposed delivery date needs evidence of its own. This separation gives the reviewing colleague something concrete to accept, correct or investigate.

Make interruptions part of the plan

Connections fail and people pause work. A useful workflow records enough progress to resume: which order was checked, which results were returned and which action is waiting for review. LangGraph's persistence documentation describes checkpoints for maintaining workflow state, including across interruptions when backed by persistent storage.

Saving progress does not, by itself, make every retry safe. Suppose a support ticket was created but the confirmation never arrived. Repeating the creation request could produce a second ticket. AWS's engineering guidance explains how idempotent APIs let repeated attempts represent the same intended operation without adding duplicate effects.

For a business integration, give each intended action a stable request reference where supported, check its outcome and limit retries. If the receiving system cannot recognise duplicates, build a reconciliation step or send the uncertain case to an operator. Keep 'failed' and 'outcome unknown' distinguishable.

References: LangChain: LangGraph persistence; AWS Builders' Library: Making retries safe with idempotent APIs

Separate access to information from permission to act

An assistant that can explain an order should not automatically be able to change its delivery address or release a refund. Define separate permissions for viewing records, preparing recommendations and committing changes. Enforce those permissions in the application and connected tools.

An email, document or web page can also contain instructions aimed at the assistant. Reading that material should not give it authority to change the task or send information elsewhere. OWASP recommends treating external content as untrusted, limiting tool privileges and validating actions outside the model.

Where review is required, show the person the actual proposed change: the order, the new value, the destination and the supporting evidence. If the record changes while approval is pending, check it again before applying the decision. Human review works best when the reviewer can see exactly what will happen.

References: OWASP: LLM Prompt Injection Prevention Cheat Sheet

Judge the complete job, then expand

Before a pilot, agree what a successful enquiry looks like. Did the assistant identify the correct order, use current records, explain the uncertainty and respect the user's permissions? Could a colleague verify the answer without repeating the whole investigation? These checks reveal more than whether the response sounds fluent.

Record completion time, failed attempts, human corrections and the cost of the whole run, including retries and tool calls. Keep timeouts in the results. A quick average can hide the enquiries that never finish. When comparing designs, use the same cases and record differences in models, data access and resource budgets so improvements can be attributed fairly.

Start with one recurring enquiry and a limited set of systems. Test missing records, conflicting statuses, unavailable connections and ambiguous requests alongside straightforward cases. Review the results with the people who handle that work today.

For a business connecting ERP, warehouse, production and customer support, a useful first milestone is an enquiry that reaches a checkable answer with a clear next step. Expand from that foundation as the evidence supports it.

TAKE THIS INTO YOUR NEXT DISCUSSION

A practical starting checklist

  • One recurring task with a clear completion rule
  • Named data sources, record owners and freshness checks
  • Separate permissions for reading, drafting and changing records
  • A recovery plan for interrupted or uncertain actions
  • A pilot measured on accuracy, completion time, full cost and human rework

Start with one business question. Connect the evidence, define the decisions and make the result easy to check. Let measured improvements guide how much autonomy the workflow needs.

Original practical guidance with an illustrative order-enquiry scenario, not a customer case study. The linked engineering articles and documentation support the design principles discussed. References checked on 29 October 2026.

Put the idea into practiceExplore Agentic Automation