AI automation in 2026: building a workflow you can depend on

A message arrives in an inbox. AI extracts a few details, creates a CRM contact and prepares a reply. The demonstration takes two minutes. Then the same customer sends another message, an attachment is unreadable and the CRM responds late. This is where the real automation project begins.

In 2026, an important design decision is how much autonomy each step deserves. A model can interpret ambiguous language. A business rule can validate a required field. A person can approve a quotation with commercial consequences. A well-designed system gives each a clear job.

Start with a process someone owns

Choose a repetitive workflow with an observable outcome and a person who can judge whether the result is correct. “Use AI in the business” is too broad. “Prepare complete enquiries for quotation and route exceptions to an operator” is something you can verify.

Map the current process, including places where people copy information or request clarification. Check whether the problem is interpreting information or simply a missing integration. If two applications already have compatible structured fields, moving data between them may need conventional automation without a language model.

For a pilot, define the input, the required output and the owner of unsuccessful cases. If you cannot describe the exception path, the operation is not yet specified well enough.

Workflow or agent: choose according to task variability

Anthropic's guide to building agents, originally published in 2024, distinguishes predetermined workflows from agents that choose their next actions dynamically. The architectural distinction remains useful; the article is not a measure of adoption in 2026.

In our hypothetical example, receiving an enquiry, validating fields and saving a CRM record can follow stable steps. AI can interpret the free-text description and suggest a category. An agent becomes useful when it must investigate several sources and decide which information to check next.

The practical question is which uncertainty autonomy resolves. If the answer is unclear, begin with a simple workflow. Add flexibility once you have concrete examples the existing rules cannot handle.

Separate a proposal from an external action

A system can prepare a reply without having permission to send it. It can recommend a category without being allowed to change a price. This separation makes mistakes easier to notice and correct.

For the first pilot, we suggest having AI produce a structured result: extracted details, missing information, the source of each field and a proposed next step. The application validates the schema and business rules. An operator sees the original alongside the proposal and approves consequential actions.

Do not make a model's statement that it is “95% confident” the sole approval condition. Checks such as whether the customer exists, identifiers match and the necessary document is present provide more concrete evidence.

Give the integration only the access needed for its assigned role. An enquiry classifier rarely needs permission to delete customers or alter invoices. Keep that limit in the actual tool permissions rather than relying solely on instructions in a prompt.

Design for repetition and failure from the beginning

An integration may perform an action while its response gets lost. Retrying without checking what happened can create two records. Each important operation needs a stable identifier and a rule for resuming it.

Stripe's webhook documentation provides a concrete example: deliveries can repeat and their order is not guaranteed. This illustrates why integrations must account for such behaviour explicitly; check each supplier's own contract separately.

For commercial enquiries, record the processing state: received, analysing, needs review, completed or failed. Distinguish an unsuccessful attempt from an action whose outcome is unknown. In the second case, seek confirmation in the external system before repeating the effect.

Set a retry limit and an exception destination. A process that keeps spending money without finishing does not become more useful because it is autonomous. The operator needs enough information to resume the work without reconstructing the entire history.

Test the outcome, including awkward cases

Anthropic's January 2026 guide to agent evaluations distinguishes an agent's messages from the final state of the system. A claim of success needs checking against the actual outcome. It also recommends repeated trials because behaviour can vary.

For a pilot, we suggest an initial set of a few dozen anonymised examples chosen with the process owner. This is a practical starting point, not statistical certification of reliability. Include incomplete enquiries, duplicates, wrong attachments, conflicting information and an unavailable integration.

Include a document containing instructions addressed to the AI. Material received from a customer is input to analyse; it cannot grant the system new permissions by itself.

For each case, inspect the record created, the correctness of its fields, external effects and escalation behaviour. Keep separate examples that were not used to tune the prompt. Otherwise, you risk optimising the demonstration instead of the process.

Calculate the cost of an accepted outcome

The model bill is only part of the cost. Add integration work, monitoring, operator time and correction of mistakes.

A useful pilot calculation is the total cost of runs, review and correction divided by the number of correctly completed cases. Compare the same class of enquiry before and after automation. Track time to resolution too: a modest saving can matter greatly if it removes a day of waiting.

Expand autonomy when there is evidence from relevant cases. If the pilot requires approval at every step and still consumes more time, narrow the problem or improve the inputs before adding more agents.

A good automation project leaves an explainable process: what it received, what it decided, what it changed and who can intervene. That is where AI starts creating operational value.

For the wider project, read how to organise software delivery in 2026. You can also talk to us about a specific process worth automating.

PhantomTech editorial analysis. Sources checked on 10 October 2026. The commercial workflow is hypothetical; pilot recommendations should be adapted to the consequences of an error.