We asked ChatGPT to run our agency workflow in 2023. Here’s what 2026 actually looks like.

In January 2023 we published ChatGPT’s own unedited plan for automating agency workflow. Three years later, about half of it happened, none of it the way it predicted, and the biggest change was one it never mentioned: AI stopped being a writing tool and became a coworker whose effort you have to see and manage like anyone else’s.

Yes, we really published the answer unedited

The setup was wonderfully simple.

The team was fascinated by ChatGPT. We were talking about how it might affect sales support, project work, and workflow management. So we asked ChatGPT to write a blog article about its own role in automating agency workflow.

Then we pasted the answer into the blog and told readers it was unaltered.

That was the article.

If this sounds quaint now, remember the moment. People were asking chatbots to write poems, sales emails, business plans, and explanations of quantum physics, then sharing the result because the mere fact that a machine produced coherent paragraphs felt like news.

We were not alone. We were simply honest enough to leave the fingerprints on the page.

The old post is useful now because it captured what AI thought the future would look like before any of us had lived through it. Some predictions were directionally sharp. Others mistook a confident paragraph for a working business system.

The 2023 scorecard

Sales automation: partly arrived

The post said AI could “automate tasks such as lead generation, qualification, and follow-up.”

That direction was right. AI now helps firms research prospects, summarize account context, draft outreach, prepare follow-ups, and sort information faster. But the useful version is rarely a model running sales by itself. It is a person using AI inside a process with data, review, and judgment.

The prediction underestimated the hard parts: access to trustworthy context, permission to act, knowing when not to send, and preserving the human relationship behind professional services.

Grade: useful direction, overly automatic execution.

Estimates from past data: right question, magical answer

The post claimed businesses could “quickly and accurately generate estimates for new projects” by analyzing past work.

Past delivery evidence absolutely should improve the next estimate. AI can help find patterns, compare similar Jobs, and surface Deliverables that repeatedly use more effort than planned.

But evidence does not remove uncertainty. A new estimate still needs clear scope, assumptions, input from the people doing the work, and a decision about what is truly comparable. The future arrived as evidence-based planning, not a magic estimate button.

Grade: directionally right, far too confident.

Related: How do I estimate creative and technical projects more accurately?

Resource planning: yes, with context

The post said AI could “identify areas where resources are being over or underutilized.”

That is now a practical use. An AI assistant can help review current assignments, planned effort, due dates, and capacity, then call attention to collisions or gaps.

The catch is that the assistant needs clean, current workflow data. If assignments live in one tool, effort in another, and schedule changes in chat, AI does not create truth. It creates a polished interpretation of fragments. Resource planning improved when AI gained access to the work system, not merely when the language model became smarter.

Grade: yes, when the data and permissions are real.

Profitability analysis: the framing that aged worst

The post promised AI could “better understand which projects are most profitable.” It also wandered into cash flow forecasting and robot-accountant territory.

That framing skipped the delivery truth underneath the money. An AI assistant cannot responsibly explain a Job’s economics if the firm does not know what each Deliverable was planned to take, what it actually took, how long it ran, and whether the scope changed.

The real need turned out to be seeing effort and drift. The owner can combine that evidence with financial data kept in the appropriate system. AI may help with the analysis, but it does not make weak inputs trustworthy.

Grade: wrong layer of the problem.

The change the old answer completely missed

The 2023 answer treated AI as an analyst and writing assistant. It imagined AI observing work, recommending improvements, and automating administrative steps.

It did not imagine AI doing meaningful parts of delivery.

Today, AI can help research, draft, design, code, test, analyze, summarize, transform, and prepare work across creative and technical firms. Agents can take a goal, use tools, move through several steps, and return with a result for review.

That creates a new visibility problem.

When AI produces a large part of the work, the owner’s question stops being, “Should we use AI?” The practical question becomes, “What does our work actually take now?”

A Deliverable may use less human effort but require more review. A specialist may complete work faster but spend more effort correcting weak inputs. A task may move quickly while the overall timeline still waits on client decisions. The workflow changed, so the old baseline may no longer tell the truth.

You need to see human and AI contribution in the context of the Deliverable, not celebrate the tool in isolation.

Related: How do I compare estimated vs. actual effort on agency projects?

When large language models first broke into the mainstream, plenty of teams sat down and brainstormed every way they might bolt the new capability onto their product. The more useful question turned out to be quieter: what actually helps a small firm see and manage its work? The features that lasted were the ones tied to real workflow, not the ones added because the technology was new.

What 2026 actually looks like

Hybrid human-and-AI delivery is normal in small firms. The exact mix differs by service, team, risk, and client, but the argument about whether AI belongs in professional work is largely over. The work now is learning where it helps, where it creates review burden, and how the delivery plan should change.

AI has also moved closer to the systems where work lives. Net Net has an MCP connector, so a customer’s own AI assistant can work with the customer’s workflow system on trial and paid tiers. With the right access and confirmation, the assistant can read context and help carry out supported actions instead of relying on pasted summaries.

That is more useful than a chatbot describing what an agency might do. It is also more serious. Permissions, workspace boundaries, confirmation, and auditability matter because the AI is no longer only suggesting. It can participate.

The lesson that survived every tool change

The tools changed faster than the old article expected. The management need did not.

You still need a clear promise, a usable plan, visible effort, an honest timeline, and a way to learn from what happened. AI can help at every stage. It can also hide weak process behind fast output.

The answer is not to resist the coworker or worship it. Give it good context, sensible boundaries, and work worth doing. Then keep watching the flow.

FAQ

What agency work does AI actually do well in 2026?

AI is useful for research, first drafts, variations, summaries, data cleanup, code assistance, testing support, documentation, and preparing routine follow-ups. It works best when the goal, source material, boundaries, and reviewer are clear. It is weaker when the task depends on unstated client history, sensitive judgment, original strategy, or accountability that nobody has explicitly assigned.

Did AI replace agency jobs the way people feared in 2023?

The result is mixed. AI absorbed parts of many roles and changed what clients expect, but professional work still needs judgment, relationship context, review, and ownership. Some tasks require far less human effort; other tasks appeared around directing and checking AI. The useful question is not whether a role vanished. It is how the role and delivery baseline changed.

How do I track what work takes when AI does half of it?

Keep the Deliverable and Task as the unit of context. Record the human effort used to direct, review, correct, and finish the work, and note meaningful AI contribution without pretending machine output equals human effort. Compare the new pattern with the old baseline. The goal is to understand the changed workflow well enough to plan the next Job honestly.