Blog
AI
September 15, 2026
5 min read
The agentic AI conversation in 2026 was shaped largely by vendor marketing. This piece is written from the perspective of a team that ships AI into production for mid market organizations and sees which patterns hold up after the demo audience has left the room.
Across Thessia's 2026 engagements, a narrow subset of agentic workflows has moved from pilot into production. A larger subset is still being described as production ready in marketing while sitting in pilot. The gap matters because the 2027 budget is being allocated right now.
What "agentic" actually means in production
Strip the marketing. An agent is a system that decides what to do next based on the state of its environment, calls tools or services to act on that decision, and loops until a stop condition is met. That is it.
In production, three things separate an agent that ships from a demo that does not.
The first is a tight tool surface. Production agents call three to seven tools, not thirty. Every added tool multiplies the ways the agent can fail, and the demo audience never sees the long tail of failure modes that only show up at real volume.
The second is bounded autonomy. Production agents operate inside a scope that was defined before launch. They do not expand their own scope. If a new capability is needed, a human adds it to the tool surface and the governance review happens before deployment, not after.
The third is observable state. Production agents log every decision, every tool call, and every output in a format a human can review. Agents without this fail silently and the team only discovers the failure from a customer complaint.
Agentic systems that include all three (tight tool surface, bounded autonomy, observable state) are the ones Thessia has seen reach production and stay there.
The three categories of agentic workflow that are shipping
Not every workflow is a good candidate for an agent. The ones that are tend to share a shape.
Document processing with structured output. Contract extraction, invoice processing, inspection report normalization, insurance claim intake. The common pattern: unstructured input, deterministic output schema, a human reviewer in the loop for edge cases. These workflows ship because the output can be validated against the schema and the agent fails loudly when the input does not fit the pattern.
Internal research and triage. Agents that read across a known set of internal systems (ticket trackers, CRM, knowledge base) and produce a draft response or recommendation for a human to approve. The agent is doing the first pass. The human is doing the final decision. These ship because the stakes of a wrong answer are low (a draft gets edited) and the time savings are real.
Multi step operational tasks with explicit guardrails. Reconciliation, data quality checks, status monitoring with escalation. Agents that run the same twenty step process every day, catch exceptions, and escalate to a human when the pattern breaks. These ship when the process is well documented, the exceptions are well understood, and the escalation path is staffed.
Outside these three categories, the agentic workflows Thessia has reviewed tend to fall into one of three groups: compelling demos that have not yet shipped at mid market scale, specialized research tools built for different audiences, or traditional workflows with "agent" added as a label.
The architectural pattern that holds up
A production agent that lasts in production usually has this shape:
- A planner that decomposes the task into steps, written against a stable prompt that is version controlled.
- A fixed tool registry with each tool typed, documented, and tested in isolation.
- An execution loop with a hard step limit (typically 10 to 25 steps) that halts and escalates if the limit is reached.
- A validator that checks the agent's output against a schema or acceptance criteria before anything is written back to the business system.
- A full audit log for every run, queryable by a non engineer.
- A kill switch at the queue level so one operator can disable the agent system wide.
The pattern is unglamorous. It is also the pattern Thessia has watched hold up through twelve months of production operation.
Three failure modes that appear at production scale
Three failures hit production agents in ways that rarely surface in a demo.
Tool version drift. The agent was tested against version 3 of an internal API. Version 4 ships. The agent fails in a way that looks like a model problem but is actually a schema problem. The fix is monitoring tool call contracts the same way you monitor model output.
Loop budget exhaustion in the wrong place. The agent hits its step limit on a case that was in scope but hard. Operations gets paged, the on call engineer bumps the limit, and the limit creeps up until it stops serving as a safety mechanism. The fix is treating the limit as a product decision reviewed quarterly, not an engineering default anyone can edit.
Silent drift in the planner prompt. Someone edits the planner prompt to fix one edge case. The fix works for that case and breaks three others in production. Three days later the metrics look fine on average and nobody knows anything changed. The fix is version control on prompts with the same discipline as application code.
None of these are glamorous. All of them determine whether the agent is still running in month twelve.
Three questions that clarify readiness for 2027
Three questions help clarify whether an agentic project is ready to fund in 2027.
Does the workflow fit one of the three categories Thessia has seen ship, or does the business case rest on broader agentic narratives that have not yet shipped at mid market scale?
Does the team have the operating model to maintain the audit log, the prompt versioning, and the tool contract monitoring, with named owners past launch?
Does the governance plan include a documented kill switch and a human escalation path from week one?
Across Thessia's 2026 engagements, the agentic projects that reached production shared yes answers across all three. Teams where the answers are still open typically benefit from a readiness sprint before the build budget is approved.
Where to start
For teams evaluating an agentic workflow for 2027, the Thessia Labs Enterprise AI Readiness Assessment scores the organization across the five readiness dimensions. Fifteen questions, five minutes, free, with a snapshot before any contact details are collected.
Across 2026, the agentic projects Thessia watched reach production were the ones treated as production systems from week one. The 2027 budget is the moment to decide which shape an agentic project takes from the start.
Frequently asked questions
What is agentic AI in a production setting?
An agentic AI system decides its next step based on what it sees, calls approved tools or systems to act, and repeats until a stop condition is met. In production, the version that holds up connects to a small number of systems, works inside a scope defined before launch, and logs every decision so a person can review it.
Which business workflows are good candidates for agentic AI?
The workflows Thessia has seen reach production fall into three groups: document processing with a fixed output format (contracts, invoices, inspection reports), internal research and triage where a person approves the draft, and repeatable operational tasks such as reconciliation or data quality checks with a clear escalation path.
How can a mid market team tell if an agentic project is ready to fund?
Three questions cover most of it. Does the workflow fit one of the categories that ship? Is there a named owner for the audit log, prompt versioning, and tool monitoring after launch? Is there a documented kill switch and a human escalation path from week one? Our free Enterprise AI Readiness Assessment scores the broader readiness picture in about five minutes.
