AI workflow patterns: the real unit of enterprise adoption in 2026

This article is part of Turing Post's "The Org Age of AI" series, written by Will Schenk (co-founder of TheFocus.AI) and Ksenia Se. It is the fifth installment in a series that has spent four articles insisting on an uncomfortable idea: debates about AI adoption in organizations are being waged at the wrong level. People talk about models, agents, frameworks, "use cases," but the unit you can actually point to, inspect, automate and improve is smaller and more concrete: the workflow.

The authors define a workflow as "a repetitive sequence of decisions and actions that turns an input into an output, with points along the way where a human exercises judgment." The key lies in those judgment points: if they are removed, what remains is a pipeline —a script, a cron job, plumbing—; if they are kept, it is a workflow. The interesting question, they say, is not whether something can be automated, but which judgment points an agent can take on, which ones still need a human, and how the human knows when to step in.

An important organizational diagnosis they offer is that most companies have dozens of workflows running in every department —support, finance, engineering, sales— that have never been documented, because the employees who run them absorbed their complexity years ago. The first step to adopting AI seriously is not choosing a tool, but discovering these informal workflows: this is what they call the "L0-to-L2 problem," that is, making the organization legible to itself before it can improve anything.

Drawing on the operation of some thirty real production systems at TheFocus.AI —content pipelines, financial reconciliation tools, engineering automation, deal-scoring platforms, API monitors, newsletter delivery— the authors identify seven "primitives" that are the atomic actions an agent can execute in a single step, regardless of domain: Watch (wait for a trigger or condition: has something happened yet?), Validate (check against known criteria: is this correct?), Classify (assign a category or route: what kind of thing is this?), Enrich (add useful information: what can we add to make it more useful?), Generate (produce a new artifact: what should be created?), Execute (take an action with real consequences: should this happen now?) and Elicit (ask a human to reduce ambiguity: what do we still need to know?). None of these primitives alone is enough to constitute a workflow —a simple "validate" call is not a workflow—, but chained together with branching logic and one or two human checkpoints, they are.

From there, the article lays out eight workflow patterns that, according to the authors, recur again and again in real production systems (initially they believed there were five, but a deeper audit of their own code, combined with operational patterns observed at Bloomberg, Zapier, Cursor and OpenRouter, brought three more to light):

1) Triage (classify → route): an item arrives, the agent decides what type it is and sends it down the right path. No artifact is produced and no action is executed; the value lies in the routing decision. Human intervention is usually nil, because a wrong routing is cheap to correct. As an example they cite the operations of Tezlab, where DevOps and support events are combined to decide the level of reaction needed, and an in-house system ("twitter-collator") that watches lists of X accounts and classifies each tweet by relevance and engagement, routing the high-priority ones to a notification queue and archiving the rest.

2) Investigation and recommendation (validate + enrich → generate a recommendation): the agent gathers context from multiple sources, analyzes it and produces a recommendation —not a draft, but an evidence-backed judgment— for a human to act on. Examples: a system that analyzes SEC filings and press releases to decide whether an executive's departure was an orderly event or the response to a scandal; and a reconciliation system ("qbsync") for a construction company that cross-references QuickBooks data with budgets in Google Sheets, computes spending curves by line item and generates variance reports (for example, detecting that demolition is 40% over budget at the 20% progress stage). The agent does the work of gathering and cross-referencing information in minutes instead of half an hour, but the decision to act remains human.

3) Draft and review (generate artifact → human review → optional review): the agent produces a complete artifact —document, email, report, presentation— and presents it for review before it is sent anywhere. Example: a system for a private equity firm with $2 billion under management that generates draft "Go-To-Offer" presentations from meeting notes in Markdown, confidential information memoranda in PDF and financial exports from QuickBooks, producing a two-page deck plus a supporting Excel workbook. The deal lead always edits tone, emphasis and risk framing, but starts from something that used to take hours of an analyst's time and now takes the agent minutes.

4) Execution with approval (propose action → human approval → execute): the agent is ready to take an action with real consequences —send an email, post on social media, deploy code, load data into production— and waits for the human's go-ahead. The key design question is the granularity of that approval (per item, per batch or per parameter). Example: an email pipeline built on the Buttondown API that drafts newsletters from Markdown sources; the agent prepares the draft and uploads it as an unsent email, and a human reviews tone, links and content before scheduling the send, with a clear window of reversibility.

5) Monitoring and escalation (watch → validate → branch between silence if all is well or escalation to a human if there is a problem): the agent monitors conditions on a schedule; if everything is normal, nothing happens; if something breaks or goes out of range, the agent escalates the problem with enough context to act. Example: a system ("usage-monitor") that tracks API keys and credit balances for fourteen different AI providers (OpenAI, Anthropic, Google, Mistral and twenty-three others) for multiple organizations, checking every hour for validity, balance and rate limits, and firing alerts to Slack, Discord or email only when something fails —for example, an authentication failure of a production key at two in the morning.

6) Elicitation (ask → the human answers → ask again… → hand off to generate or execute): the agent and the human co-build a specification through a structured dialogue; the agent asks one question at a time and refuses to act until it has enough shared understanding. Here the human does not review an output, but provides the input that makes a good output possible. Example: a system for a private equity firm that drafts Letters of Intent (LOIs) from deal terms described conversationally by the deal lead; the agent does not generate immediately, but first asks clarifying questions based on the firm's historical corpus of executed LOIs —for example, noting that the last five LOIs with an earnout included a change-of-control acceleration clause, or that the non-compete language in healthcare has differed from the standard. The value lies in the questions, not in generating the document itself.

7) Synchronization and transformation (watch/trigger → validate → enrich → execute/load): data moves from a system A to a system B, transforming along the way, with no human judgment in the normal path because the transformation rules are known in advance. It is the pattern most likely to run unattended for weeks. Example: a media analytics platform that ingests data from five different providers (audiences in CSV, streaming metrics in Excel, celebrity recognition data in SPSS files, social followers, audience affinity scores), each with its own format and schedule; the pipeline validates the file, renormalizes column names, normalizes the data (title names, Spanish-language content) and cross-references between sources (for example, matching a celebrity's name to a talent ID, or tagging special events such as presidential debates). After each load it runs a battery of integrity tests, and the human only steps in when something fails —a new column, a change in the provider's naming convention, a sheet with an unexpected structure.

8) Curation and scheduled delivery (watch the clock → classify and enrich → generate → execute/deliver): on a schedule, the agent gathers material from multiple sources, synthesizes it and delivers a finished product to subscribers, without any human triggering the process or reviewing it before delivery. The downloaded content cuts off right at the description of this last pattern, so the complete example is not available, nor is the final section on how to chain these patterns into larger pipelines, nor the promised practical framework for deciding which workflows to automate first.

Overall, the article defends a clear thesis for those evaluating the adoption of agentic AI in the enterprise: instead of buying "agents" or adopting a trendy framework, it is best to first do the —seemingly unglamorous— work of identifying the real, undocumented workflows that already exist in the organization, describing them in terms of these seven primitives, and recognizing which of the eight patterns they fit. This makes it possible to decide with judgment where human judgment is indispensable (approvals, elicitation, draft review) and where it can disappear almost entirely (triage, synchronization, monitoring), thereby managing risk in proportion to the cost of an error in each pattern. It is a pragmatic approach grounded in real production cases —private equity funds, construction firms, media analytics platforms, AI infrastructure monitoring— rather than in abstract promises about autonomous agents, and it fits the general tone of the "The Org Age of AI" series, which had already argued in earlier installments that the ROI of AI remains elusive precisely because the conversations focus on the tools and not on the underlying organizational processes.