If you are planning AI workflow automation, start from one fact: the help is uneven. An agent can be excellent at one step of a process and unreliable at the next, and the gap between the two is wider than most demos let on. The job is to find which steps an agent handles well, automate those, and keep a person on the rest.
On November 18, OpenAI put numbers on this. In a report called Research acceleration: the jagged frontier, surfaced in The Rundown AI on November 20, the company measured how its own coding agents handle real research work. Agents now finish nearly half of the tasks that would take a human two to four hours, up from under a quarter in March. That is a real jump.
Then the same report breaks the work down by task, and the spread is large. Agents succeeded at 79% of jobs that monitor training runs, 35% of jobs that launch them, and 25% of writing code for inference and serving. OpenAI calls the pattern "jagged", and says humans remain a core part of the loop. This is an AI lab measuring its own best tools on its own work, and the honest read is that uplift depends heavily on the task.
Look at those three numbers again. Monitoring a training run is a well-defined job: the inputs are known, there is a clear picture of what good looks like, and a wrong answer is easy to spot. Writing new inference code is the opposite. It needs judgment, it has no single right answer, and a mistake can be subtle. The agent is strong where the step is bounded and checkable, and weak where it is open-ended.
That pattern holds outside an AI lab. In the processes founders run every day, the reliable steps tend to be the repetitive ones with clear rules: pulling data from a form, sorting an inbound request by type, reconciling two lists, drafting a first version from a template. The steps that stay human are the ones that need taste or a real decision: what to say to an unhappy customer, whether a deal is worth taking, where a number looks wrong.
So the question is not "can AI run this workflow". It is "which steps of this workflow can AI run, and which ones need me". Those are different questions, and the second one is the useful one.
Before you buy a tool or wire anything up, write the process out end to end, from the trigger that starts it to the finished result. Then mark each step: an agent can handle this reliably, a person needs to own this, or you are not sure yet. The unsure ones are where to run a small test before you trust them.
This mapping is most of the work, and it is the part people skip. It is also what our team does first when we step in as a fractional COO for an agency: we follow work from first inquiry to final invoice and find where it gets stuck, at intake, scoping, resourcing, approvals, reporting or billing. You cannot automate a process you have not written down, and once it is written down, the automation is usually obvious.
A workflow that has been mapped this way can go a long way. Hike Medical runs Step Zero, a care qualification platform that automates medical paperwork, signatures and cross-provider fax flows through a long chain of OCR, fact extraction and multi-stage compliance checks. That is a real process, automated end to end, with the checks built into the pipeline rather than bolted on after.
The OpenAI report is clear that agents are weakest where taste and novel decisions matter, and that its researchers remain central to the work. For a business, the same line falls in a predictable place: the steps that reach a customer, the ones that move money, and anything a mistake would be expensive to undo.
The way to respect that is to let the agent do the work and hold the output for a person to approve. When we build custom agents and automations into an agency's CRM, inbox and project tools, they sort inbound requests, draft follow-ups and prepare proposals, and anything going out waits for your approval. The same idea runs through Byld Ops, the system we run BYLD on: the AI assistant holds anything going out until a person signs off. The people lead and the software supports them.
A person on the final call is what makes the automation safe to leave running, and that stays true however good the models get. It is the reason you can hand an agent the tedious middle of a process and still trust what comes out the other end.
The fastest return comes from the boring work: the steps your team does many times a week, the same way each time, with clear rules. Sorting inbound, drafting first replies, moving data between two tools that will never share an API, preparing a proposal from a template. These are low-risk, high-frequency and easy to check, which is exactly where the OpenAI numbers say agents are strongest.
Leave the rare, high-stakes decisions for later, or for never. A step you run twice a year is not worth automating, and a step where a wrong call costs a client is not worth handing over. Automate the volume, keep the judgment, and you get most of the benefit with little of the risk.
The mistake I see most is treating AI workflow automation as all or nothing: either the whole process runs itself or you do not bother. The OpenAI data argues against both. Some steps are ready now and some are not, and the value is in knowing which is which and building around it.
For well-defined, repetitive steps with clear rules, yes, especially when a person approves anything that leaves the building. For open-ended steps that need judgment, not on its own. Reliability tracks how bounded and checkable the step is, so the honest answer depends on the step, not the tool.
The work your team repeats most often, with the clearest rules and the lowest cost of a mistake: sorting inbound requests, drafting first versions, moving data between tools, preparing documents from templates. Map the process first, mark which steps an agent can own, and start there.
If you want a second pair of eyes on where automation would actually help in your operations, book a call and we will walk through it.