2026 Vietnam Digital Market Report: People Are Ready for AI. Company Systems Are Not.

At 4:15 every weekday morning, an AI agent on my laptop is scheduled to start work. In practice, it starts when the laptop wakes up, which on some days is after 7:00. It reads the latest messages and documents from each of our client accounts, sums up what changed since yesterday, checks off finished items on my task list, and writes today's plan. My morning starts by checking what it wrote.

2026 brought nonstop news about AI agents, along with physical AI such as humanoid robots and a steady stream of AI-powered cyberattacks. Vietnam had its own milestones. Its Law on Artificial Intelligence took effect on March 1. It was reported as the first comprehensive AI law in Southeast Asia. In August, penalties under the Personal Data Protection Law became enforceable.

In June, in a Japanese eBook on AI and jobs in Vietnam, I predicted that AI here would spread less through individuals paying out of pocket and more through companies building it into their work. Since then, we have done exactly that at our own agency. Some of it went as I expected. Some of it did not.

This report follows our 2025 Vietnam Digital Market Report. It covers three things we learned at an agency of about ten people, and what each one suggests for 2027.

The lag is in company systems, not in people

A larger share of people in Vietnam already use generative AI than in Japan. Microsoft's AI Diffusion Report puts the share of the working-age population (15 to 64) using generative AI at 27.9% in Vietnam, against 24.7% in Japan. Companies are a different story. According to an AWS study published in September, 26% of Vietnamese businesses have adopted AI, and 61% of those are still experimenting. In other words, only 39% of adopters, about one in ten of all businesses, have moved beyond the experiment stage.

I see three reasons. None of them is about skill or attitude.

• Price. A business AI seat costs around $25 a month almost everywhere. That is about 1.2% of the average monthly wage in Japan, but about 3% of the base salary of staff at Japanese companies in Ho Chi Minh City, according to JETRO's 2025 survey. So companies tend not to buy seats, and AI stays on personal accounts.
• Who decides the workflow. Engineers pick their own tools. Most office staff work inside procedures set by a manager or by headquarters in Japan, so one person's good idea rarely becomes the way the team works.
• Turnover. Staff turnover is high, so what one person figures out leaves with them unless the process keeps it.

So in July, as a decision rather than a proposal, we built AI into the work itself. Meeting minutes are now written in deliberate detail so the AI can read them. Monthly reports and slides come out of procedures that already include our brand colors and fonts. Within one or two months, the team's questions shifted from "How do I use this tool?" to "Do I understand the business well enough to hand this off?"

In 2027, AI will likely stay a personal tool in any company where the people who set the procedures do not build it in.

Stop fixing, start writing rules

In Will AI Reduce Digital Marketing Jobs in Vietnam?, I wrote that AI cuts drafting time but adds review work. That turned out to be true. When I counted one week in July and scaled it to a month, about a third of my working time went to review. Faros AI, drawing on data from 22,000 developers, found a similar pattern in software teams. When AI adoption was high, engineers completed about three times as many coding tasks, but the median time their code spent in review was more than five times longer.

So I dropped two habits I used to think made me a good manager.

The first was "It's faster if I fix it myself." Now, instead of fixing an output, I write the correction into the agent's instruction file as a rule, so I do not have to give the same correction twice. Some rules are tiny:

• Calculate totals and differences from the rounded numbers shown, so the table adds up.
• Keep a message draft to six lines.
• Write for the reader, not for the conversation we just had.

The rules get corrected too. One formatting rule was reversed after I deleted the same emoji from four messages in a row.

Some software engineers have started calling this "harness engineering": every time the agent makes a mistake, you build something so it never makes that mistake again. The harness alone can change results a lot. LangChain reported that changing only the harness lifted its agent's benchmark score from 52.8% to 66.5% on the same model. Our main instruction file is now about 12 times longer than it was in early July. When people change roles, the rules stay.

The second habit was "Read everything before it goes out." At the end of August, I stopped reviewing the work on one account. The team member now checks it with AI and sends it to the client.

Some automations did not last. We built 26 scheduled routines, and 15 were still running in October. We stopped a weekly summary that nobody read and a drafting routine whose output I never made time to review. I should have planned the reading time before building the output. Why routines that run on a timer, with no context, tend to fail is covered in Delegating to an AI Agent.

In 2027, I expect managers to spend more time adding rules than fixing outputs.

In 2027, the price moves to what AI is not trusted with

On the big ad platforms, AI now makes most of the decisions, from bidding to placements. In its Q2 2026 earnings call, Meta said its fully automated Advantage+ campaigns were on track to bring in more than $75 billion a year. Mark Zuckerberg described the model in one line: "businesses only pay us when we achieve results for them." On TikTok Shop, GMV Max has been the only campaign type since July 2025. So the price of execution work will likely keep falling.

Buyers have noticed. Some now ask for "AI discounts." KPMG pushed its own auditor to cut its fee, arguing that savings from AI should be passed on. The fee fell 14%, according to a TheStreet report citing the Financial Times. Vietnam is exposed too. In the 2025 Offshore Development White Paper, a Japanese industry survey, Vietnam was the top destination with 43% of requests, and about 65% of all contracts were still sold on engineer time. Pricing by the hour is starting to look shaky.

What we decided first was what AI does not get to finish. Sending anything to a client, setting final prices, signing contracts, and approving new Japanese copy all stay with a person, even when AI writes the draft. It is the same rule I described in Delegating to an AI Agent: drafts are autonomous, sending is human. Since late June, I have asked our AI for Slack drafts more than 1,300 times, and it has sent none of them. Clients are drawing lines too. One Japanese client asked for our generative AI policy in writing as a condition of the contract.

Our pricing is moving along the same line: less about the volume of work or the hours, more about what we check and how much responsibility we take on. To be honest, we have not found the final answer yet.

When you choose a partner in 2027, it may be worth asking not only "How many people, for how many hours?" but also "Where does a person check the work, and who is accountable?"

Tomorrow morning, the agent will start again. What it writes will still be a draft until a person checks it. Deciding where that check sits is the work I expect to matter most in 2027.

Feedforce Vietnam helps companies grow sales in Vietnam and other markets, from strategy to day-to-day operation, across digital advertising, e-commerce (Shopee, TikTok Shop, Shopify) and B2B lead generation. More about our approach is on our marketing consulting page.