AI and Work in 2026: What I Take from the Trend Reports

TL;DR

The useful question is not how much AI a company uses. It is whether a specific workflow produces a better result without creating more review work than it saves. Three conclusions guide my own projects: measure the effect before making decisions based on it, treat review effort as part of the cost, and redesign the process instead of buying a tool and hoping the work changes around it.

Updated 13 September 2026: corrected the attribution of the trend list and narrowed the commentary to my own work.

What the Reports Actually Say

On 2 February, Harvard Business Review published “9 Trends Shaping Work in 2026 and Beyond” by Peter Aykens, Kaelyn Lowmaster, Emily Rose McRae, and Jonah Shepp. The article draws on Gartner's research. Gartner's own press release and open article name the actual nine trends: workforce cuts before measured AI gains, cultural dissonance, mental fitness, AI workslop, candidate fraud, corporate espionage, moves from tech into trades, process expertise, and digital doppelgangers.

I cannot assess all nine from my own work. Recruiting fraud, trade careers, and compensation for digital replicas need evidence I do not have. Three parts do connect directly to building and running AI workflows.

1. Measure the Result Before Acting on the Forecast

Gartner's first trend concerns workforce decisions made before the expected AI effect is established. The same mistake appears at a smaller scale in software projects: a team buys a tool, counts generated output, and treats activity as a result.

In my own work, the useful measurement sits at the end of a bounded process. Did the document extractor return the required fields on real documents? Did the coding agent's patch pass the repository's checks? Did the assistant save a person effort after review?

I keep the existing process and decision point in place until the replacement has been tested against representative inputs.

2. Review Work Is Part of the Cost

Gartner calls fast, poor-quality AI output “workslop”. I see the engineering version whenever several agents return plausible patches at once. Production gets cheaper; attention moves into review.

I have worked with AI agents daily since late 2024. The recurring bottleneck is deciding whether the output belongs in the system. Tests catch some failures. Source checks, browser runs, and manual review catch others. A clean-looking result can still solve the wrong problem.

Review belongs in the estimate. If an assistant saves ten minutes of drafting and adds fifteen minutes of checking, the workflow did not improve. Uncertainty markers, source links, and a separate final action can reduce that effort.

3. Start with the Process and Its Owner

Gartner's eighth trend emphasises process expertise in making AI useful. This matches what I have learned from prototypes. The model is often the quickest part. Data access, permissions, exception handling, and responsibility determine whether the result can be used.

I start with five questions:

1. What enters the process?
2. What result is needed?
3. Which difficult cases occur in real use?
4. Who decides whether the result is correct?
5. What happens when the system is uncertain or unavailable?

Only then does the choice of model or tool become useful. A document workflow may need preprocessing more than a newer model. An inbox assistant may need a clear send approval more than a longer prompt. A coding agent may need repository checks and a narrow task boundary more than another round of generated code.

What I Keep from the Reports

My takeaway is operational: measure completed work, include review in the cost, and keep responsibility with a named person or role.

That is also a practical starting point for a company. Pick one process, involve the people who run it, and define what evidence would justify changing it. The AI guide for SMEs lays out that first pass in more detail. Verification, Not Typing covers the review side of the workflow.